mirror of
https://github.com/wassname/ml_debug.git
synced 2026-09-09 11:27:07 +08:00
cite arxiv pdf, never abs
An /abs/ link is a stub: it costs a second lookup before anyone can check the quote, and it is what let abstract-only caches look sourced. 77 links across the skill files and 4 cache headers. Verbatim source bodies untouched, their reference lists are the authors' text.
This commit is contained in:
wassname
+1
-1
@@ -217,7 +217,7 @@ Folklore sources (the quotes above trace to these):
|
||||
[^jones]: Andy Jones, "Debugging RL, Without the Agonizing Pain" — https://andyljones.com/posts/rl-debugging.html ([cache](docs/evidence/andyljones_rl_debugging.md): anomalies L103-109, write-from-scratch L155, assume-bug L176-180, raise-threshold L182, loss-curve L186-188)
|
||||
[^rahtz]: Matthew Rahtz (Amid Fish), "Lessons Learned Reproducing a Deep RL Paper" — http://amid.fish/reproducing-deep-rl ([cache](docs/evidence/amid_fish_reproducing_deep_rl.md): frame-diff confusion L85-87, investigate-confusion L100-102, think-more L145-153, don't-implement-RL-yourself L497-501)
|
||||
[^schulman]: John Schulman, "Nuts and Bolts of Deep RL Research" slides — http://joschu.net/docs/nuts-and-bolts.pdf ([cache](docs/evidence/joschu_nuts_and_bolts.md): Always-Be-Ablating L71-75, standardize-observations L84-88; clean slide transcript)
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) — https://arxiv.org/abs/1709.06560 ([cache](docs/evidence/henderson_2018_deep_rl_matters.md): seeds-create-different-distributions L235, implementation-differences L251)
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) — https://arxiv.org/pdf/1709.06560 ([cache](docs/evidence/henderson_2018_deep_rl_matters.md): seeds-create-different-distributions L235, implementation-differences L251)
|
||||
[^cs231n]: Stanford CS231n, "Neural Networks Part 3" — https://cs231n.github.io/neural-networks-3/ ([cache](docs/evidence/cs231n_neural_networks_3.md): overfit-tiny-subset L89)
|
||||
[^slavv]: Slav Ivanov, "37 Reasons why your Neural Network is not working" (2017) — https://blog.slavv.com/37-reasons-why-your-neural-network-is-not-working-4020854bd607 ([cache](docs/evidence/slavv_37_reasons_nn.md): opening anecdote L19, emergency checklist L45-51)
|
||||
[^fsdl]: Josh Tobin, Full Stack Deep Learning Spring 2021, Lecture 7 "Troubleshooting DNNs" — https://fullstackdeeplearning.com/spring2021/lecture-7/ ([cache](docs/evidence/fsdl_spring2021_lecture7.md))
|
||||
|
||||
@@ -347,7 +347,7 @@ Open the relevant one when the task calls for it. These are synthesized checklis
|
||||
|
||||
Start here rather than treating the bibliography as flat:
|
||||
|
||||
- **Beginner / broad checklist:** Lones, ["How to avoid machine learning pitfalls"](https://arxiv.org/abs/2108.02497), with its full do/don't list extracted in [refs/checklist.md](refs/checklist.md).
|
||||
- **Beginner / broad checklist:** Lones, ["How to avoid machine learning pitfalls"](https://arxiv.org/pdf/2108.02497), with its full do/don't list extracted in [refs/checklist.md](refs/checklist.md).
|
||||
- **Debugging a neural net:** Karpathy, ["A Recipe for Training Neural Networks"](https://karpathy.github.io/2019/04/25/recipe/).
|
||||
- **Designing tuning experiments:** Google, [Deep Learning Tuning Playbook](https://developers.google.com/machine-learning/guides/deep-learning-tuning-playbook).
|
||||
- **Transformer and LLM runs:** [refs/transformers.md](refs/transformers.md), then the HF, Axolotl, Unsloth, nanochat, and Bekman sources below.
|
||||
@@ -360,13 +360,13 @@ Folklore sources (the quotes above trace to these):
|
||||
[^karpathy-mistakes]: Andrej Karpathy, "most common neural net mistakes" tweet thread, 1 Jul 2018 — https://x.com/karpathy/status/1013244313327681536 ([cache](docs/evidence/karpathy_common_mistakes_tweet_2018.md): tweets 1-3 verbatim, cross-checked against threadreaderapp; x.com itself blocks fetching)
|
||||
[^sculley]: Sculley et al., "Hidden Technical Debt in Machine Learning Systems" (NIPS 2015) — https://papers.nips.cc/paper_files/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf ([cache](docs/evidence/sculley_2015_hidden_technical_debt.md): abstract, CACE/entanglement, ensemble caveat)
|
||||
[^schulman]: John Schulman, "Nuts and Bolts of Deep RL Research" slides — http://joschu.net/docs/nuts-and-bolts.pdf ([cache](docs/evidence/joschu_nuts_and_bolts.md): Always-Be-Ablating L71-75, standardize-observations L84-88; clean slide transcript)
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) — https://arxiv.org/abs/1709.06560 ([cache](docs/evidence/henderson_2018_deep_rl_matters.md): seeds-create-different-distributions L235, implementation-differences L251)
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) — https://arxiv.org/pdf/1709.06560 ([cache](docs/evidence/henderson_2018_deep_rl_matters.md): seeds-create-different-distributions L235, implementation-differences L251)
|
||||
[^irpan]: Alex Irpan, "Deep Reinforcement Learning Doesn't Work Yet" (2018) — https://www.alexirpan.com/2018/02/14/rl-hard.html ([cache](docs/evidence/alexirpan_rl_hard.md): variance-bug-or-unlucky L674-678, seed-canary L705-707)
|
||||
[^cs231n]: Stanford CS231n, "Neural Networks Part 3" — https://cs231n.github.io/neural-networks-3/ ([cache](docs/evidence/cs231n_neural_networks_3.md): overfit-tiny-subset L89)
|
||||
[^slavv]: Slav Ivanov, "37 Reasons why your Neural Network is not working" (2017) — https://blog.slavv.com/37-reasons-why-your-neural-network-is-not-working-4020854bd607 ([cache](docs/evidence/slavv_37_reasons_nn.md): opening anecdote L19, emergency checklist L45-51)
|
||||
[^goodfellow]: Goodfellow, Bengio, Courville, *Deep Learning*, ch. 11 "Practical Methodology" — https://www.deeplearningbook.org/ ([cache](docs/evidence/goodfellow_ch11_practical_methodology.md): one-part-broken-others-adapt L198, weights-adapt-to-compensate L204)
|
||||
[^mccandlish]: McCandlish, Kaplan et al., "An Empirical Model of Large-Batch Training" (2018) — https://arxiv.org/abs/1812.06162 ([cache](docs/evidence/mccandlish_2018_large_batch.md))
|
||||
[^goyal]: Goyal et al., "Accurate, Large Minibatch SGD" (2017) — https://arxiv.org/abs/1706.02677
|
||||
[^mccandlish]: McCandlish, Kaplan et al., "An Empirical Model of Large-Batch Training" (2018) — https://arxiv.org/pdf/1812.06162 ([cache](docs/evidence/mccandlish_2018_large_batch.md))
|
||||
[^goyal]: Goyal et al., "Accurate, Large Minibatch SGD" (2017) — https://arxiv.org/pdf/1706.02677
|
||||
[^lucidrains]: Phil Wang (lucidrains), x-transformers README — https://github.com/lucidrains/x-transformers ([cache](docs/evidence/lucidrains_x_transformers_readme.md): post-embedding LayerNorm / BLOOM+YaLM L366, attention-overflow / cosine-sim norm L1230, autoregressive validation L1234, "wiping out a source of instability" / QK RMSNorm L1292)
|
||||
[^koaning]: Vincent D. Warmerdam (koaning), "Bad Labels" (2021) — https://koaning.io/posts/labels/ ([cache](docs/evidence/koaning_bad_labels.md): bad-labels-huge-problem L13, confidence-sort trick L21, spend-less-time-tuning L33)
|
||||
[^nanochat]: nanochat (Karpathy), documented via DeepWiki — https://deepwiki.com/karpathy/nanochat ([cache](docs/evidence/nanochat_deepwiki_llm_pretraining_2026.md): BOS fake-improvement L97, all-ranks-clip-on-inf L131)
|
||||
@@ -386,7 +386,7 @@ Folklore sources (the quotes above trace to these):
|
||||
[^tuning-playbook]: Godbole, Dahl, Gilmer, Shallue, Nado, "Deep Learning Tuning Playbook" (Google Research / Google Developers, 2023; Google Developers page last updated 2025-08-25) — https://developers.google.com/machine-learning/guides/deep-learning-tuning-playbook ([cache](docs/evidence/google_tuning_playbook.md): exploration-over-exploitation L24, scientific/nuisance/fixed L34-38, incremental-tuning L14-18)
|
||||
[^domingos]: Pedro Domingos, "A Few Useful Things to Know About Machine Learning" (CACM, Oct 2012) — https://homes.cs.washington.edu/~pedrod/papers/cacm12.pdf ([cache](docs/evidence/domingos_2012_few_useful_things.md): test-on-train illusion L20, insidious-contamination L22, overfitting-bugbear L26, features-are-key L32)
|
||||
[^bekman-book]: Stas Bekman, *Machine Learning Engineering Open Book*, "Understanding Training Loss Patterns" + "Instabilities" — https://github.com/stas00/ml-engineering ([cache](docs/evidence/bekman_ml_engineering_instabilities.md): heartbeat L10, 104B post-mortem L18, spike types + bad-data-pocket L22-24, init-std L28-32, PaLM batch-skipping L36, logbooks L40)
|
||||
[^lones]: Michael A. Lones, "How to avoid machine learning pitfalls" (2021, updated annually) — https://arxiv.org/abs/2108.02497 ([cache](docs/evidence/lones_2021_ml_pitfalls.md): full do/don't TOC L18-22, leakage L26, look-ahead bias L30). Aimed at beginners but the most exhaustive checklist here: 36 do/don'ts across data prep, training, evaluation, comparison, and reporting.
|
||||
[^lones]: Michael A. Lones, "How to avoid machine learning pitfalls" (2021, updated annually) — https://arxiv.org/pdf/2108.02497 ([cache](docs/evidence/lones_2021_ml_pitfalls.md): full do/don't TOC L18-22, leakage L26, look-ahead bias L30). Aimed at beginners but the most exhaustive checklist here: 36 do/don'ts across data prep, training, evaluation, comparison, and reporting.
|
||||
|
||||
For modern transformer pretraining specifically (most sources above predate it), see [Karpathy's recipe](https://karpathy.github.io/2019/04/25/recipe/) and the [nanochat deepwiki](https://deepwiki.com/karpathy/nanochat) (320+ empirical HP sweeps for a GPT-2-scale run). For LLM-as-judge eval debugging workflow more broadly, Hamel Husain's ["Your AI Product Needs Evals"](https://hamel.dev/blog/posts/evals/) covers the error-analysis-first approach for LLM products. Most multi-source claims trace to quotes in [docs/ml_debug_folklore.argdown](docs/ml_debug_folklore.argdown) (vargdown); the full evidence set is in [docs/evidence/](docs/evidence/).
|
||||
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
Source: https://arxiv.org/abs/1709.06560
|
||||
Source: https://arxiv.org/pdf/1709.06560
|
||||
Title: Deep Reinforcement Learning that Matters - Henderson et al. (2018)
|
||||
Fetched-via: curl https://r.jina.ai/https://arxiv.org/pdf/1709.06560
|
||||
Fetch-status: verbatim
|
||||
|
||||
@@ -5,19 +5,19 @@ Fetch-status: verbatim (abstracts in full or near-full; verdict pages are short
|
||||
|
||||
# LLM judge biases (excerpts)
|
||||
|
||||
## "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" — Zheng et al. (LMSYS), NeurIPS 2023 — https://arxiv.org/abs/2306.05685
|
||||
## "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" — Zheng et al. (LMSYS), NeurIPS 2023 — https://arxiv.org/pdf/2306.05685
|
||||
|
||||
The canonical paper naming the bias taxonomy:
|
||||
|
||||
> We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. [...] Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans.
|
||||
|
||||
## "Large Language Models are not Fair Evaluators" — Wang et al., ACL 2024 — https://arxiv.org/abs/2305.17926
|
||||
## "Large Language Models are not Fair Evaluators" — Wang et al., ACL 2024 — https://arxiv.org/pdf/2305.17926
|
||||
|
||||
Positional bias is large enough to flip rankings outright:
|
||||
|
||||
> We find that the quality ranking of candidate responses can be easily hacked by simply altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other, e.g., Vicuna-13B could beat ChatGPT on 66 over 80 tested queries with ChatGPT as an evaluator. To address this issue, we propose a calibration framework with three simple yet effective strategies: 1) Multiple Evidence Calibration, which requires the evaluator model to generate multiple evaluation evidence before assigning ratings; 2) Balanced Position Calibration, which aggregates results across various orders to determine the final score; 3) Human-in-the-Loop Calibration [...]
|
||||
|
||||
## "LLM Evaluators Recognize and Favor Their Own Generations" — Panickssery, Bowman, Feng (NYU/MATS), 2024 — https://arxiv.org/abs/2404.13076
|
||||
## "LLM Evaluators Recognize and Favor Their Own Generations" — Panickssery, Bowman, Feng (NYU/MATS), 2024 — https://arxiv.org/pdf/2404.13076
|
||||
|
||||
Self-preference is causally linked to self-recognition:
|
||||
|
||||
@@ -95,13 +95,13 @@ Title: evaluator blind spots (Doddapaneni), criteria drift (Shankar), Judgemark
|
||||
Fetched-via: WebFetch of arXiv abstract pages, 2026-07-22; Judgemark scores read from the checked-in judgemark-v4.js in EQ-bench/EQ-bench-site
|
||||
Fetch-status: paper quotes verbatim from abstracts; Judgemark numbers copied from the repo's data rows (not the rendered site table)
|
||||
|
||||
## Doddapaneni et al., "Finding Blind Spots in Evaluator LLMs with Interpretable Checklists" (2024) — https://arxiv.org/abs/2406.13439
|
||||
## Doddapaneni et al., "Finding Blind Spots in Evaluator LLMs with Interpretable Checklists" (2024) — https://arxiv.org/pdf/2406.13439
|
||||
|
||||
Evaluator LLMs miss most injected quality drops:
|
||||
|
||||
> Our findings reveal significant shortcomings in current Evaluator LLMs, which failed to identify quality drops in over 50% of cases on average.
|
||||
|
||||
## Shankar et al., "Who Validates the Validators?" (2024) — https://arxiv.org/abs/2404.12272
|
||||
## Shankar et al., "Who Validates the Validators?" (2024) — https://arxiv.org/pdf/2404.12272
|
||||
|
||||
Criteria drift, and the validator-needs-validation problem:
|
||||
|
||||
@@ -216,7 +216,7 @@ Independent, outsider-run, continuously-updated public harness (strong trust). M
|
||||
|
||||
Rule of thumb: even in 2026, judges flip ~43% of decisive verdicts on order swap alone; the worst flip >70%. Direction is not universal (Mistral Large 3 goes second-position). Always judge both orders. (Note: an earlier subagent draft misattributed 27.4% to Claude Opus; the raw table puts 27.4% first-shown pick on Mistral Large 3, so I dropped the per-model attributions except the verified worst-case.)
|
||||
|
||||
## "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" — Shi et al. (Dartmouth), IJCNLP-AACL 2025 — https://arxiv.org/abs/2406.07791 — [FT]
|
||||
## "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" — Shi et al. (Dartmouth), IJCNLP-AACL 2025 — https://arxiv.org/pdf/2406.07791 — [FT]
|
||||
Most-cited dedicated position-bias study, largest scale (150k+ instances, 15 judges, 22 tasks).
|
||||
|
||||
> Our experiments, involving 15 LLM judges across MTBench and DevBench with 22 tasks and approximately 40 solution-generating models, result in over 150,000 evaluation instances.
|
||||
@@ -225,7 +225,7 @@ Most-cited dedicated position-bias study, largest scale (150k+ instances, 15 jud
|
||||
|
||||
Rule of thumb: position bias is systematic, not random noise, and it gets WORSE as the two answers converge in quality (exactly when you most need the judge). Metrics introduced: repetition stability, position consistency, preference fairness.
|
||||
|
||||
## "Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models" — Norman, Rivera, Hughes (UC Berkeley), 2026 — https://arxiv.org/abs/2606.19544 — [ID]
|
||||
## "Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models" — Norman, Rivera, Hughes (UC Berkeley), 2026 — https://arxiv.org/pdf/2606.19544 — [ID]
|
||||
2026 audit of 21 judges. Title confirmed via arxiv abs this turn; in-body numbers via subagent WebFetch (re-fetched HTML mirror after PDF parse failed).
|
||||
|
||||
> High test-retest reliability (>0.95) coexists with severe position bias (>0.10) in two production-deployed judges (instantiating a consistency-bias paradox).
|
||||
@@ -238,7 +238,7 @@ Two rules of thumb: (1) a judge being REPRODUCIBLE (same verdict on re-run) does
|
||||
|
||||
## Self-preference / self-enhancement scales inversely with size
|
||||
|
||||
## "Beyond the Surface: Measuring Self-Preference in LLM Judgments" — Chen et al., EMNLP 2025 main — https://arxiv.org/abs/2506.02592 — [FT]
|
||||
## "Beyond the Surface: Measuring Self-Preference in LLM Judgments" — Chen et al., EMNLP 2025 main — https://arxiv.org/pdf/2506.02592 — [FT]
|
||||
Machine-accessible: https://github.com/zhiyuanc2001/self-preference . Its DBG score uses gold judgments to separate bias from genuine quality. Quotes copied from arxiv HTML full-text this turn:
|
||||
|
||||
> the DBG score of Qwen2.5-0.5B-Instruct is 41.7%. In contrast, the DBG score of Qwen2.5-14B-Instruct is only 2.1%.
|
||||
@@ -251,7 +251,7 @@ Machine-accessible: https://github.com/zhiyuanc2001/self-preference . Its DBG sc
|
||||
|
||||
Rule of thumb: self-preference is inversely proportional to size. Tiny judges (<1B) can inflate their own scores ~40%; strong large judges drop to low single digits. Reasoning does NOT reliably remove it. Do not use a small model to judge its own family's outputs.
|
||||
|
||||
## "Do LLM Evaluators Prefer Themselves for a Reason?" — Chen et al., 2025 — https://arxiv.org/abs/2504.03846 — [ID]
|
||||
## "Do LLM Evaluators Prefer Themselves for a Reason?" — Chen et al., 2025 — https://arxiv.org/pdf/2504.03846 — [ID]
|
||||
The counter-intuitive one: CoT REDUCES self-preference (contradicts naive "more thinking = more bias"). Title confirmed this turn; HSPP numbers via subagent WebFetch.
|
||||
|
||||
> generating reasoning traces substantially reduces harmful self-preference across all models
|
||||
@@ -262,14 +262,14 @@ Tension to flag: this paper says CoT roughly halves self-preference; 2506.02592
|
||||
|
||||
## Reasoning judges: accuracy up, superficial bias not fixed
|
||||
|
||||
## "JudgeLRM: Large Reasoning Models as a Judge" — Chen et al., 2025 — https://arxiv.org/abs/2504.00050 — [FT]
|
||||
## "JudgeLRM: Large Reasoning Models as a Judge" — Chen et al., 2025 — https://arxiv.org/pdf/2504.00050 — [FT]
|
||||
RL-trained reasoning judge. Full text re-pulled from raw PDF 2026-08-14; F1 numbers confirmed.
|
||||
|
||||
> JudgeLRM achieves an average improvement of 8.14% in F1 score [vs same-size SFT judges]
|
||||
|
||||
> On the human-annotated PandaLM benchmark, JudgeLRM-3B surpasses GPT-4 [F1 72.12% vs 61.80%]
|
||||
|
||||
## "Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases" — Huang et al., Jan 2026 — https://arxiv.org/abs/2601.03630 — [ID]
|
||||
## "Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases" — Huang et al., Jan 2026 — https://arxiv.org/pdf/2601.03630 — [ID]
|
||||
|
||||
> LRMs outperform non-reasoning LLMs in terms of judgment accuracy, particularly on reasoning-intensive tasks
|
||||
|
||||
@@ -277,7 +277,7 @@ RL-trained reasoning judge. Full text re-pulled from raw PDF 2026-08-14; F1 numb
|
||||
|
||||
Rule of thumb across these two: prefer a reasoning judge for reasoning-heavy grading (~5-8 F1 gain), but it does not remove length/position/style bias. "Use a reasoning judge" is defensible; "reasoning fixes bias" is not.
|
||||
|
||||
## "Explicit Reasoning Makes Better Judges" — 2025 — https://arxiv.org/abs/2509.13332 — [ID]
|
||||
## "Explicit Reasoning Makes Better Judges" — 2025 — https://arxiv.org/pdf/2509.13332 — [ID]
|
||||
Directly tests small judges (Qwen3 0.6B/1.7B/4B). (Subagent first mislabeled the title as "Thinking Small Models..."; corrected to the real arxiv title this turn.)
|
||||
|
||||
> thinking models achieve approximately 10% points higher accuracy with little overhead (under 2x), in contrast to augmentation strategies like few-shot learning, which deliver modest gains at a higher cost (>8x).
|
||||
@@ -288,7 +288,7 @@ Directly tests small judges (Qwen3 0.6B/1.7B/4B). (Subagent first mislabeled the
|
||||
|
||||
Rule of thumb: sub-1B judges fall to random on hard/safety pairs; turning on reasoning buys ~+10 accuracy and higher bias-consistency far cheaper than few-shot ICL (<2x cost vs >8x).
|
||||
|
||||
## "RLAIF vs. RLHF" — Lee et al. (Google), ICML 2024 — https://arxiv.org/abs/2309.00267 — [FT]
|
||||
## "RLAIF vs. RLHF" — Lee et al. (Google), ICML 2024 — https://arxiv.org/pdf/2309.00267 — [FT]
|
||||
The canonical "smaller = more position-biased" source. Main-text quotes copied from arxiv HTML this turn; the 18/21/56% per-size figures are in its Appendix B (table, not captured by my main-text grep).
|
||||
|
||||
> We find evidence of position bias, which is especially prevalent in smaller LLM labelers
|
||||
@@ -301,14 +301,14 @@ Reported (Appendix B, via subagent): PaLM 2 L/S/XS keep the same position after
|
||||
|
||||
## Overthinking: token budget vs task difficulty is non-monotonic
|
||||
|
||||
## "Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models" — Ghosal et al., 2025 — https://arxiv.org/abs/2506.04210 — [ID]
|
||||
## "Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models" — Ghosal et al., 2025 — https://arxiv.org/pdf/2506.04210 — [ID]
|
||||
Cleanest non-monotonic curve. Title confirmed this turn; numbers via subagent WebFetch.
|
||||
|
||||
> accuracy increases from 82.2% to 87.3% as the average number of thinking tokens increases from 385 to 1100. However... pushing the average thinking token count from 1100 to 15980 reduces accuracy from 87.3% to 70.3%
|
||||
|
||||
Rule of thumb: return on thinking tokens peaks then declines. In their setup peak was ~1.1k tokens; 14x more tokens (16k) cost ~17 accuracy points. Past the peak, extra tokens add variance, not reasoning.
|
||||
|
||||
## "OptimalThinkingBench: Evaluating Over and Underthinking in LLMs" — Aggarwal et al., 2025 — https://arxiv.org/abs/2508.13141 — [ID]
|
||||
## "OptimalThinkingBench: Evaluating Over and Underthinking in LLMs" — Aggarwal et al., 2025 — https://arxiv.org/pdf/2508.13141 — [ID]
|
||||
|
||||
> Thinking models often overthink for hundreds of tokens on the simplest user queries without improving performance. In contrast, large non-thinking models underthink, often falling short of much smaller thinking models.
|
||||
|
||||
@@ -316,14 +316,14 @@ Rule of thumb: easy items hit negative marginal utility of thinking earlier than
|
||||
|
||||
## Self-consistency convergence: how many samples N
|
||||
|
||||
## "Self-Consistency Improves Chain of Thought Reasoning" — Wang et al., 2022 — https://arxiv.org/abs/2203.11171 — [ID]
|
||||
## "Self-Consistency Improves Chain of Thought Reasoning" — Wang et al., 2022 — https://arxiv.org/pdf/2203.11171 — [ID]
|
||||
Foundational (several-thousand citations), PaLM-540B era.
|
||||
|
||||
> GSM8K (+17.9%) [self-consistency over CoT; 56.5% -> 74.4% at N=40]
|
||||
|
||||
Widely-reproduced pattern: gain is monotonic in N with diminishing returns; bulk arrives by N=5-10, saturates ~N=40 for that era's models.
|
||||
|
||||
## "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" — Loo, 2025 — https://arxiv.org/abs/2511.00751 — [FT]
|
||||
## "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" — Loo, 2025 — https://arxiv.org/pdf/2511.00751 — [FT]
|
||||
Single-author preprint (low citation signal, flagged; author states AI tools assisted drafting), but directly answers "how has N moved." Full text re-pulled from raw PDF 2026-08-14.
|
||||
|
||||
> [MATH-500, Gemini-2.5-Flash-Lite] accuracy improved through approximately 10 sampled paths before plateauing... declining slightly beyond 15
|
||||
@@ -332,7 +332,7 @@ Single-author preprint (low citation signal, flagged; author states AI tools ass
|
||||
|
||||
Rule of thumb for N: on strong 2026 models the self-consistency plateau moved in to N~10-15 (from ~40), total gain shrank to <2 points, and accuracy can DECLINE past the plateau. Reserve repeats for genuinely hard items where the base model is well below ceiling. This also sets the sane N for your N=4 repeat-variance check: 4-10 is plenty to see instability; going past ~15 buys nothing.
|
||||
|
||||
## "Inference-Time Scaling for Generalist Reward Modeling" (DeepSeek-GRM) — Liu et al., 2025 — https://arxiv.org/abs/2504.02495 — [ID]
|
||||
## "Inference-Time Scaling for Generalist Reward Modeling" (DeepSeek-GRM) — Liu et al., 2025 — https://arxiv.org/pdf/2504.02495 — [ID]
|
||||
Vendor paper (mild caution). Numbers via subagent WebFetch.
|
||||
|
||||
> Direct voting with 32 samples of DeepSeek-GRM-27B could achieve comparable performance to the 671B MoE model
|
||||
@@ -352,7 +352,7 @@ Industry report (not peer-reviewed), 18-model controlled study. Numbers via suba
|
||||
|
||||
Rule of thumb: degradation is continuous and starts well before the window fills; a 1M-token window does not reliably reason over 1M tokens. Accuracy is highest when the key info sits near the START of the sequence.
|
||||
|
||||
## "NoLiMa: Long-Context Evaluation Beyond Literal Matching" — Modarressi et al., ICML 2025 — https://arxiv.org/abs/2502.05167 — [FT]
|
||||
## "NoLiMa: Long-Context Evaluation Beyond Literal Matching" — Modarressi et al., ICML 2025 — https://arxiv.org/pdf/2502.05167 — [FT]
|
||||
Repo: https://github.com/adobe-research/NoLiMa . Removes literal lexical overlap, so it measures latent-association retrieval (closest analog to a judge matching a rubric to a semantically-distant answer). Full text re-pulled from raw PDF 2026-08-14.
|
||||
|
||||
> The effective length is defined as the longest context where a model maintains at least 85% of its base score.
|
||||
@@ -363,7 +363,7 @@ Repo: https://github.com/adobe-research/NoLiMa . Removes literal lexical overlap
|
||||
|
||||
Rule of thumb: once literal cues are gone, most models fall below their 85% effective length by 1-4K tokens (median 2K, Table 3); only GPT-4o reaches 8K and GPT-4.1 16K. By 32K, 11 of 13 are below half their short-context score.
|
||||
|
||||
## "Lost in the Middle" — Liu et al., TACL 2024 — https://arxiv.org/abs/2307.03172 — [FT]
|
||||
## "Lost in the Middle" — Liu et al., TACL 2024 — https://arxiv.org/pdf/2307.03172 — [FT]
|
||||
Origin of the U-shaped/middle-penalty result, replicated across 6 model families.
|
||||
|
||||
> performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models.
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
Source: https://arxiv.org/abs/2108.02497 (v5)
|
||||
Source: https://arxiv.org/pdf/2108.02497 (v5)
|
||||
Title: "How to avoid machine learning pitfalls: a guide for academic researchers" -- Michael A. Lones (Heriot-Watt University)
|
||||
Fetched-via: curl https://r.jina.ai/https://arxiv.org/pdf/2108.02497, 2026-08-14
|
||||
Fetch-status: verbatim, full paper (33 pages)
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
Source: https://arxiv.org/abs/1812.06162
|
||||
Source: https://arxiv.org/pdf/1812.06162
|
||||
Title: An Empirical Model of Large-Batch Training - McCandlish & Kaplan (2018)
|
||||
Fetched-via: curl https://r.jina.ai/https://arxiv.org/pdf/1812.06162
|
||||
Fetch-status: verbatim
|
||||
|
||||
+27
-27
@@ -49,7 +49,7 @@ This also makes the input domain ~[0,1] naturally, which is what NeuralPDE.jl ex
|
||||
If you can't nondimensionalize cleanly (unknown material properties, mixed units), at minimum z-score each input/output channel so the network sees zero-mean unit-variance data.
|
||||
|
||||
> Rathore et al. 2024: "the estimate of the κ grows polynomially with nres" -- but this is in raw units. Nondimensionalization reduces the effective condition number by making all PDE coefficients O(1).
|
||||
> Source: https://arxiv.org/abs/2402.01868, Section 8.2 (Theorem 8.4), empirical check in Appendix F.5
|
||||
> Source: https://arxiv.org/pdf/2402.01868, Section 8.2 (Theorem 8.4), empirical check in Appendix F.5
|
||||
|
||||
---
|
||||
|
||||
@@ -67,12 +67,12 @@ From NeuralPDE.jl tests/docs + Wang et al. 2021:
|
||||
|
||||
**Modified MLP** (Wang et al. 2021, credence ~70%):
|
||||
> Wang et al. propose a modified MLP with multiplicative interactions. With `U = φ(XW1 + b1)`, `V = φ(XW2 + b2)` two nonlinear encodings of the input (φ = tanh) and a per-layer gate `Z(k) = φ(H(k)Wz,k + bz,k)` computed from the hidden state, the update is `H(k+1) = (1 - Z(k)) * U + Z(k) * V`. Authors claim a ~3x decrease in the leading Hessian eigenvalue.
|
||||
> Source: https://arxiv.org/abs/2001.04536, Section 2.6, equations 43-47
|
||||
> Source: https://arxiv.org/pdf/2001.04536, Section 2.6, equations 43-47
|
||||
> Evidence: on Helmholtz the architecture alone (M3) improves relative L2 error 3.4-8.7x over vanilla PINN (M1) across 9 width/depth settings (Table 2); with LR annealing (M4) it is 23-97x. On Klein-Gordon M3 is 9.1x, M4 is 64x (Table 3). The paper's headline "50-100x" is the combination, not the architecture. Only tested by the proposing authors; no independent replication found.
|
||||
|
||||
**Random Weight Factorization (RWF)** (arXiv 2210.01274, credence ~60%):
|
||||
> Factorize each neuron's weight vector as w = s * w_unit, where s is a trainable scalar and w_unit is the unit-normalized direction. This changes the optimization geometry so the loss surface has better-conditioned local minima. "Predictions obtained by RWF are in excellent agreement with ground truth, while other weight parameterizations result in poor or non-physical approximations."
|
||||
> Source: https://arxiv.org/abs/2210.01274
|
||||
> Source: https://arxiv.org/pdf/2210.01274
|
||||
> Used in the PirateNet architecture alongside causal training, sequence-to-sequence, and Fourier features. Simple to implement as a custom parameterization on Linear layers.
|
||||
> Credence: plausible mechanism, but proposing-author result; check jaxpi repo for independent adoption.
|
||||
|
||||
@@ -107,7 +107,7 @@ The loss landscape of PINNs is ill-conditioned. First-order methods (Adam) conve
|
||||
```
|
||||
|
||||
> Rathore et al. 2024 (ICML, credence ~80%): "Adam+L-BFGS attains 14.2x smaller L2RE than Adam on convection and 6.07x smaller than L-BFGS on wave." Tested on 3 PDEs (convection, reaction, wave), 5 seeds, widths 50-400.
|
||||
> Source: https://arxiv.org/abs/2402.01868, Table 1
|
||||
> Source: https://arxiv.org/pdf/2402.01868, Table 1
|
||||
> Code: https://github.com/pratikrathore8/opt_for_pinns
|
||||
|
||||
**Alternative from NeuralPDE.jl**: Stepwise LR decrease with warm-start:
|
||||
@@ -120,12 +120,12 @@ Adam(0.1) -> Adam(0.01) -> Adam(0.001)
|
||||
|
||||
**Near-zero loss required** (Rathore et al., credence ~85%):
|
||||
> "on the convection PDE, a loss of 10^-3 yields an L2RE around 10^-1, but decreasing the loss by a factor of 100 to 10^-5 yields an L2RE around 10^-2, a 10x improvement."
|
||||
> Source: https://arxiv.org/abs/2402.01868, Section 4, Figure 2
|
||||
> Source: https://arxiv.org/pdf/2402.01868, Section 4, Figure 2
|
||||
> Implication: you need to drive the loss very low for useful accuracy. Don't stop at "loss looks flat."
|
||||
|
||||
**L-BFGS stalls but gradient is still useful** (Rathore et al., credence ~80%):
|
||||
> "L-BFGS stops in these cases without reaching a critical point: the gradient norm is around 10^-2 or 10^-3. The gradient still contains useful information for improving the loss."
|
||||
> Source: https://arxiv.org/abs/2402.01868, Section 7.1, Figure 4 (line-search failure in Appendix E, Figure 9)
|
||||
> Source: https://arxiv.org/pdf/2402.01868, Section 7.1, Figure 4 (line-search failure in Appendix E, Figure 9)
|
||||
> Cause: strong Wolfe line search fails, step size goes to zero.
|
||||
> Fix: switch to NNCG (Armijo only) or restart with different LR.
|
||||
|
||||
@@ -147,7 +147,7 @@ The PINN loss has multiple terms (PDE residual, BCs, ICs, data) with different g
|
||||
> Source: Brunton, S. "AI/ML+Physics Part 4 - Crafting a Loss Function." https://www.youtube.com/watch?v=3SNkQ8jhKXc
|
||||
|
||||
> Wang et al. 2021 (credence ~80%): "the gradients corresponding to the boundary loss term Lub(θ) in each layer are sharply concentrated around zero and overall attain significantly smaller values than the gradients corresponding to the PDE residual loss Lr(θ)." Shown via per-layer histograms of back-propagated gradients; the paper does not quantify the gap in orders of magnitude.
|
||||
> Source: https://arxiv.org/abs/2001.04536, Section 2.2, Figures 2-3
|
||||
> Source: https://arxiv.org/pdf/2001.04536, Section 2.2, Figures 2-3
|
||||
|
||||
**Consequences:**
|
||||
- BC/IC losses are undertrained (gradient signal drowned out)
|
||||
@@ -157,7 +157,7 @@ The PINN loss has multiple terms (PDE residual, BCs, ICs, data) with different g
|
||||
### Hessian stiffness
|
||||
|
||||
> Wang et al. 2021: "many eigenvalues of the residual-loss Hessian are extremely large up to 1e5" while the boundary-loss Hessian eigenvalues stay small, so the gradient-flow stiffness is dominated by the residual term. This is an absolute magnitude, not a condition number; Wang never reports one.
|
||||
> Source: https://arxiv.org/abs/2001.04536, Section 2.4, Figures 4-5
|
||||
> Source: https://arxiv.org/pdf/2001.04536, Section 2.4, Figures 4-5
|
||||
> For a condition number, use Rathore Figure 3: outlier eigenvalues > 1e4 (convection), > 1e3 (reaction), > 1e5 (wave).
|
||||
|
||||
### Solutions (in order of preference)
|
||||
@@ -166,7 +166,7 @@ The PINN loss has multiple terms (PDE residual, BCs, ICs, data) with different g
|
||||
|
||||
**2. Learning rate annealing** (Wang et al. 2021, credence ~75%):
|
||||
> Adaptively weight each loss term inversely proportional to its gradient magnitude. EMA of gradient statistics for stability.
|
||||
> Source: https://arxiv.org/abs/2001.04536, Algorithm 1
|
||||
> Source: https://arxiv.org/pdf/2001.04536, Algorithm 1
|
||||
> NeuralPDE.jl implements this as `GradientScaleAdaptiveLoss`.
|
||||
|
||||
**3. Gradient aggregation methods** (ConFIG or UPGrad):
|
||||
@@ -184,18 +184,18 @@ ConFIG and UPGrad are both reasonable candidates when the losses cannot be repla
|
||||
> Standard PINNs use penalized (soft) constraints: add physics as a loss term. The alternative is constrained optimization: minimize data error while exactly satisfying the physics constraints. "With a loss function you're not exactly satisfying your constraints. With constrained optimization you are."
|
||||
> Source: Brunton, S. "AI/ML+Physics Part 5 - Employing an Optimization Algorithm." https://www.youtube.com/watch?v=T4iJ10TAIMg
|
||||
> Physics-informed DMD (Baddoo et al. 2021) is the cleanest example: restrict the DMD matrix to a symmetry-preserving manifold (Hermitian, symplectic, etc.) via the Procrustes problem. KKT closed-form solutions exist because DMD is linear in its parameters -- the constraint is linear in both the output and the parameters simultaneously.
|
||||
> Baddoo et al. 2021. "Physics-informed dynamic mode decomposition." Proc. R. Soc. A. https://arxiv.org/abs/2112.04307
|
||||
> Baddoo et al. 2021. "Physics-informed dynamic mode decomposition." Proc. R. Soc. A. https://arxiv.org/pdf/2112.04307
|
||||
> **Critical caveat for PINNs**: A BC like u(0)=0 is affine in the output u, but it is nonlinear in the NN weights theta. Closed-form KKT does NOT apply to neural network parameters. For NN-based PINNs, the two options for hard constraints are: (a) architectural -- multiply output by a distance function that satisfies the BC (Section 4 item 8), or (b) Augmented Lagrangian Methods (ALM), which are iterative and substantially more complex than Adam. Constrained optimization is most practical for linear models (DMD, SINDy, linear state-space) where the parameters enter linearly.
|
||||
|
||||
**5. Curriculum regularization** (Krishnapriyan et al. 2021 NeurIPS, credence ~80%):
|
||||
> When the PINN fails on hard PDE regimes (high convection coefficient, strong reaction), don't start there. Start with easy parameters (small coefficient), train to convergence, then warm-start and increase to the target regime. 1-2 orders of magnitude improvement over naive training.
|
||||
> "The curriculum training approach achieves significantly better errors, as well as lower variance in the error." (From Figure E.2 showing 10 seeds)
|
||||
> Source: https://arxiv.org/abs/2109.01050, Sections 5.1 and Figure 4
|
||||
> Source: https://arxiv.org/pdf/2109.01050, Sections 5.1 and Figure 4
|
||||
> Evidence: evidence/krishnapriyan2021_failure_modes.md
|
||||
|
||||
**6. Sequence-to-sequence (time-marching)** (Krishnapriyan et al. 2021, credence ~75%):
|
||||
> For time-dependent PDEs: train on a short time window, predict next state, step forward. Don't train on full space-time at once. "Posing the problem as seq2seq learning results in significantly lower error. The difference is particularly striking for reaction and reaction-diffusion cases, where seq2seq decreases error by almost two orders of magnitude."
|
||||
> Source: https://arxiv.org/abs/2109.01050, Section 5.2
|
||||
> Source: https://arxiv.org/pdf/2109.01050, Section 5.2
|
||||
> NeuralPDE.jl calls this time-marching; see `WeightedIntervalTraining`.
|
||||
> Note: these failures are not due to limited NN expressivity -- the architecture has enough capacity. The problem is optimization difficulty from the soft PDE constraint.
|
||||
|
||||
@@ -203,7 +203,7 @@ ConFIG and UPGrad are both reasonable candidates when the losses cannot be repla
|
||||
> Standard PINNs trained by gradient descent are implicitly biased toward minimizing residuals at *later* times before even fitting the initial conditions -- violating physical causality. The NTK analysis shows the residual at time t is influenced more by residuals at later t' > t than earlier ones. This makes PINNs fail on chaotic/turbulent systems.
|
||||
> Fix: weight each temporal residual point by wi = exp(-epsilon * sum_j<i R_j(theta)), where R_j is the accumulated residual before time i. This forces earlier times to converge first before the loss "turns on" at later times.
|
||||
> "10-100x improvements in accuracy compared to competing approaches. First time PINNs succeeded on chaotic Lorenz, Kuramoto-Sivashinsky, and 2D Navier-Stokes in turbulent regime."
|
||||
> Source: https://arxiv.org/abs/2203.07404, Abstract and Section 3
|
||||
> Source: https://arxiv.org/pdf/2203.07404, Abstract and Section 3
|
||||
> Evidence: evidence/wang2022_causal_training.md
|
||||
> Key difference from seq2seq/curriculum: causal weighting works within a single continuous training, without requiring separate time windows or changing the PDE coefficients. Can be combined with seq2seq for further gains.
|
||||
> Sensitivity: epsilon controls the steepness of the causal weights. Too small = residuals at later times turn on too early. Too large = training stalls on early time steps. Anneal epsilon during training.
|
||||
@@ -212,7 +212,7 @@ ConFIG and UPGrad are both reasonable candidates when the losses cannot be repla
|
||||
> Instead of penalizing BC violations (soft), multiply the PINN output by a distance function phi(x) that is zero on the boundary. Then u(x) = phi(x) * NN(x) satisfies BCs exactly by construction.
|
||||
> "We eliminate modeling error associated with the satisfaction of boundary conditions. The sole contribution to the loss function is from the residual error at interior collocation points."
|
||||
> "The proposed approach consistently outperforms a standard PINN-based collocation method."
|
||||
> Source: https://arxiv.org/abs/2104.08426, Abstract and Section 1
|
||||
> Source: https://arxiv.org/pdf/2104.08426, Abstract and Section 1
|
||||
> Evidence: evidence/sukumar2022_exact_bc_distance.md
|
||||
> Domain-specific failure modes and hard BC examples: see [refs/heat_exchanger.md](refs/heat_exchanger.md).
|
||||
|
||||
@@ -343,7 +343,7 @@ optimizer.step()
|
||||
|
||||
For large or geometrically complex domains, split into subdomains each with a local PINN. Interface conditions enforce continuity between subdomains.
|
||||
|
||||
> Jagtap et al. 2020. "Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations." Commun. Comput. Phys. https://arxiv.org/abs/2005.11025
|
||||
> Jagtap et al. 2020. "Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations." Commun. Comput. Phys. https://arxiv.org/pdf/2005.11025
|
||||
> Credence ~70%: Multiple citations, implemented in DeepXDE. Enables parallelization; each subdomain network is smaller and easier to optimize.
|
||||
> Key: interface residuals must be added as additional loss terms. Continuity of u and its normal derivative across interfaces.
|
||||
> Useful when the solution has different character in different regions (e.g., different phases, boundary layers).
|
||||
@@ -353,7 +353,7 @@ For large or geometrically complex domains, split into subdomains each with a lo
|
||||
## 10. PIKANs (Kolmogorov-Arnold Networks for PINNs)
|
||||
|
||||
> Toscano et al. 2024: PIKANs "lead to smaller models and may also contribute to lowering computational cost while maintaining good accuracy."
|
||||
> Source: https://arxiv.org/abs/2410.13228
|
||||
> Source: https://arxiv.org/pdf/2410.13228
|
||||
> Credence ~40%: New, no independent replication. Other authors focus on improving PINNs within the MLP framework, not validating PIKANs as an alternative. Interesting but unproven.
|
||||
|
||||
---
|
||||
@@ -361,39 +361,39 @@ For large or geometrically complex domains, split into subdomains each with a lo
|
||||
## References
|
||||
|
||||
### Comprehensive guides (start here)
|
||||
- Wang et al. 2023. "An Expert's Guide to Training Physics-Informed Neural Networks." arXiv:2308.08468. https://arxiv.org/abs/2308.08468
|
||||
- Wang et al. 2023. "An Expert's Guide to Training Physics-Informed Neural Networks." arXiv:2308.08468. https://arxiv.org/pdf/2308.08468
|
||||
- Key: most thorough practical guide. Covers architecture (modified MLP), sampling, loss weighting, Fourier features, causal training, code. By the same group as 2021 paper.
|
||||
- Wang et al. 2021. "Understanding and mitigating gradient pathologies in physics-informed neural networks." SIAM J. Sci. Comput. https://arxiv.org/abs/2001.04536
|
||||
- Wang et al. 2021. "Understanding and mitigating gradient pathologies in physics-informed neural networks." SIAM J. Sci. Comput. https://arxiv.org/pdf/2001.04536
|
||||
- Key: gradient imbalance diagnosis, learning rate annealing, modified MLP architecture
|
||||
|
||||
### Loss landscape and optimization
|
||||
- Rathore et al. 2024. "Challenges in Training PINNs: A Loss Landscape Perspective." ICML. https://arxiv.org/abs/2402.01868
|
||||
- Rathore et al. 2024. "Challenges in Training PINNs: A Loss Landscape Perspective." ICML. https://arxiv.org/pdf/2402.01868
|
||||
- Key: Adam+L-BFGS, ill-conditioning from differential operators, near-zero loss required
|
||||
- Krishnapriyan et al. 2021. "Characterizing possible failure modes in physics-informed neural networks." NeurIPS. https://arxiv.org/abs/2109.01050
|
||||
- Krishnapriyan et al. 2021. "Characterizing possible failure modes in physics-informed neural networks." NeurIPS. https://arxiv.org/pdf/2109.01050
|
||||
- Key: curriculum regularization, seq2seq, failure is optimization not expressivity
|
||||
|
||||
### Training strategies
|
||||
- Wang et al. 2022. "Respecting causality for training physics-informed neural networks." J. Comput. Phys. https://arxiv.org/abs/2203.07404
|
||||
- Wang et al. 2022. "Respecting causality for training physics-informed neural networks." J. Comput. Phys. https://arxiv.org/pdf/2203.07404
|
||||
- Key: causal weighting of temporal residuals, 10-100x improvement on chaotic systems, first PINN success on turbulence
|
||||
- Jagtap et al. 2020. "Extended Physics-Informed Neural Networks (XPINNs)." Commun. Comput. Phys. https://arxiv.org/abs/2005.11025
|
||||
- Jagtap et al. 2020. "Extended Physics-Informed Neural Networks (XPINNs)." Commun. Comput. Phys. https://arxiv.org/pdf/2005.11025
|
||||
- Key: domain decomposition, parallelizable, implemented in DeepXDE
|
||||
|
||||
### Hard constraints and boundary conditions
|
||||
- Sukumar & Srivastava 2022. "Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks." CMAME. https://arxiv.org/abs/2104.08426
|
||||
- Sukumar & Srivastava 2022. "Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks." CMAME. https://arxiv.org/pdf/2104.08426
|
||||
- Key: distance-function trial functions, eliminates BC loss, consistently outperforms soft BCs
|
||||
- Lagaris et al. 1998. "Artificial neural networks for solving ordinary and partial differential equations." IEEE Trans. Neural Netw. doi:10.1109/72.712178
|
||||
- Key: original paper on hard BCs via trial functions. Sukumar 2022 is the modern extension.
|
||||
|
||||
### Architecture and parameterization
|
||||
- Wang et al. 2022. "Random Weight Factorization Improves the Training of Continuous Neural Representations." https://arxiv.org/abs/2210.01274
|
||||
- Wang et al. 2022. "Random Weight Factorization Improves the Training of Continuous Neural Representations." https://arxiv.org/pdf/2210.01274
|
||||
- Key: factorize w = s * w_unit, better local minima, used in PirateNet
|
||||
- Toscano et al. 2024. "From PINNs to PIKANs." https://arxiv.org/abs/2410.13228
|
||||
- Toscano et al. 2024. "From PINNs to PIKANs." https://arxiv.org/pdf/2410.13228
|
||||
- PirateNet / jaxpi: https://github.com/PredictiveIntelligenceLab/jaxpi (bundles RWF + causal + seq2seq + Fourier)
|
||||
- Ling et al. 2016. "Machine learning strategies for systems with invariance properties." J. Comput. Phys. https://doi.org/10.1016/j.jcp.2016.05.003
|
||||
- Key: tensor-layer architecture enforcing Galilean invariance by construction for turbulence closure; symmetry-via-architecture beats symmetry-via-augmentation
|
||||
|
||||
### Constrained optimization and physics-informed DMD
|
||||
- Baddoo et al. 2021. "Physics-informed dynamic mode decomposition." Proc. R. Soc. A. https://arxiv.org/abs/2112.04307
|
||||
- Baddoo et al. 2021. "Physics-informed dynamic mode decomposition." Proc. R. Soc. A. https://arxiv.org/pdf/2112.04307
|
||||
- Key: restrict DMD to symmetry-preserving matrix manifolds (Hermitian, symplectic) via Procrustes problem; exactly satisfies conservation laws without penalty terms
|
||||
|
||||
### Lecture series
|
||||
@@ -405,7 +405,7 @@ For large or geometrically complex domains, split into subdomains each with a lo
|
||||
- Part 5 (optimization): https://www.youtube.com/watch?v=T4iJ10TAIMg
|
||||
|
||||
### Alternative formulations
|
||||
- Weinan E & Bing Yu 2018. "The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems." Commun. Math. Stat. https://arxiv.org/abs/1710.00211
|
||||
- Weinan E & Bing Yu 2018. "The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems." Commun. Math. Stat. https://arxiv.org/pdf/1710.00211
|
||||
- Key: energy minimization formulation instead of strong-form residuals. Better-conditioned for elliptic PDEs, can be easier to optimize since the loss is an energy (always positive, no cancellation).
|
||||
|
||||
### Multi-loss training
|
||||
|
||||
+1
-1
@@ -2,7 +2,7 @@
|
||||
|
||||
Appendix to the [ML Debugging skill](../SKILL.md).
|
||||
|
||||
This is the full do/don't list from Michael A. Lones, ["How to avoid machine learning pitfalls: a guide for academic researchers"](https://arxiv.org/abs/2108.02497) (v5, updated annually). Read the paper for the reasoning and examples behind each item; the local evidence excerpt is [here](../docs/evidence/lones_2021_ml_pitfalls.md).
|
||||
This is the full do/don't list from Michael A. Lones, ["How to avoid machine learning pitfalls: a guide for academic researchers"](https://arxiv.org/pdf/2108.02497) (v5, updated annually). Read the paper for the reasoning and examples behind each item; the local evidence excerpt is [here](../docs/evidence/lones_2021_ml_pitfalls.md).
|
||||
|
||||
> Mistakes in machine learning practice are commonplace, and can result in a loss of confidence in the findings and products of machine learning.
|
||||
|
||||
|
||||
@@ -13,14 +13,14 @@ Verify: **current LLM judges carry large, size-dependent biases (order, self-pre
|
||||
|
||||
epistemic context: outsider-run public benchmark with a reproducible swapped-order harness (193 pairs, 36 models); no arXiv paper, the numbers are the raw output of the author's own runs. The headline order-flip figure elsewhere in the same README is "the model-average order-flip rate is 43.0%".
|
||||
|
||||
## "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" — Shi et al. (Dartmouth), IJCNLP-AACL 2025 — [arXiv:2406.07791](https://arxiv.org/abs/2406.07791)
|
||||
## "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" — Shi et al. (Dartmouth), IJCNLP-AACL 2025 — [arXiv:2406.07791](https://arxiv.org/pdf/2406.07791)
|
||||
- page date: arXiv June 2024; IJCNLP-AACL 2025
|
||||
|
||||
> The findings confirm that position bias is not due to random chance and varies significantly across judges and tasks. **While position bias is weakly influenced by the length of prompt components, it is strongly affected by the quality gap between solutions.** Our agreement and disagreement analysis among judges further provides insights into the distribution of judging difficulty across the dataset, and highlights the potential for dataset modifications.
|
||||
|
||||
epistemic context: peer-reviewed; largest-scale dedicated position-bias study (over 150,000 evaluation instances, 15 judges, 22 tasks); "quality gap" here means the closer the two answers in quality, the more the judge flips on order. The abstract's counts moved across versions (v1-v3: 9 judges / 80,000 instances), and Section 3.1 of the current version still says "more than 100,000", contradicting its own abstract. Best judges in Table 2 reach position consistency 0.82, so ~18% of pairs flip on order even at the top.
|
||||
|
||||
## "RLAIF vs. RLHF" — Lee et al. (Google), ICML 2024 — [arXiv:2309.00267](https://arxiv.org/abs/2309.00267)
|
||||
## "RLAIF vs. RLHF" — Lee et al. (Google), ICML 2024 — [arXiv:2309.00267](https://arxiv.org/pdf/2309.00267)
|
||||
- page date: arXiv Sept 2023; ICML 2024
|
||||
|
||||
> We find evidence of position bias, which is especially prevalent in smaller LLM labelers (see Appendix B). **To mitigate the effect of position bias, two inferences are made for every pair of candidates, where the order in which candidates are presented to the LLM is reversed for the second inference.** The results from both inferences are then averaged to obtain the final preference distribution.
|
||||
@@ -29,7 +29,7 @@ epistemic context: peer-reviewed; the standard citation for both the "smaller =
|
||||
|
||||
## Self-preference scales inversely with judge size
|
||||
|
||||
## "Beyond the Surface: Measuring Self-Preference in LLM Judgments" — Chen et al., EMNLP 2025 main — [arXiv:2506.02592](https://arxiv.org/abs/2506.02592)
|
||||
## "Beyond the Surface: Measuring Self-Preference in LLM Judgments" — Chen et al., EMNLP 2025 main — [arXiv:2506.02592](https://arxiv.org/pdf/2506.02592)
|
||||
- page date: arXiv June 2025; EMNLP 2025 main conference. Data + code: [github.com/zhiyuanc2001/self-preference](https://github.com/zhiyuanc2001/self-preference)
|
||||
|
||||
> As observed in the figure, models larger than 7B exhibit significantly less self-preference bias compared to those of 7B or smaller. **For example, the DBG score of Qwen2.5-0.5B-Instruct is 41.7%. In contrast, the DBG score of Qwen2.5-14B-Instruct is only 2.1%.** This suggests that LLM judging tasks should utilize larger models to obtain more accurate and unbiased judgment results.
|
||||
@@ -38,7 +38,7 @@ epistemic context: peer-reviewed; the DBG (Difference-based Bias Gauge) score ne
|
||||
|
||||
## Reasoning judges: accuracy up, superficial bias not fixed
|
||||
|
||||
## "JudgeLRM: Large Reasoning Models as a Judge" — Chen et al., 2025 — [arXiv:2504.00050](https://arxiv.org/abs/2504.00050)
|
||||
## "JudgeLRM: Large Reasoning Models as a Judge" — Chen et al., 2025 — [arXiv:2504.00050](https://arxiv.org/pdf/2504.00050)
|
||||
- page date: arXiv April 2025
|
||||
|
||||
> Empirical results demonstrate that JudgeLRM not only surpasses proprietary models like GPT-4 and DeepSeek-R1 but also outperforms SFT and RL baselines of comparable sizes, **with an average improvement of 8.14% in F1 score over SFT counterparts.**
|
||||
@@ -47,7 +47,7 @@ epistemic context: quoted from Section 1 (Introduction), v3; the 8.14% figure is
|
||||
|
||||
## Overthinking: the reasoning-token budget is non-monotonic
|
||||
|
||||
## "Does Thinking More always Help? ... Mirage of Test-Time Scaling in Reasoning Models" — Ghosal et al., 2025 — [arXiv:2506.04210](https://arxiv.org/abs/2506.04210)
|
||||
## "Does Thinking More always Help? ... Mirage of Test-Time Scaling in Reasoning Models" — Ghosal et al., 2025 — [arXiv:2506.04210](https://arxiv.org/pdf/2506.04210)
|
||||
- page date: arXiv June 2025
|
||||
|
||||
> We observe an initial increase (similar to (Muennighoff et al., 2025; Aggarwal & Welleck, 2025)) in accuracy as the average thinking budget increases. **For example, in Figure 2(a), accuracy increases from 82.2% to 87.3% as the average number of thinking tokens increases from 385 to 1100.** However, this trend does not continue indefinitely.
|
||||
@@ -68,7 +68,7 @@ epistemic context: config lines quoted verbatim (not prose, so no surrounding se
|
||||
|
||||
## Self-consistency: how many samples N
|
||||
|
||||
## "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" — Loo, 2025 — [arXiv:2511.00751](https://arxiv.org/abs/2511.00751)
|
||||
## "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" — Loo, 2025 — [arXiv:2511.00751](https://arxiv.org/pdf/2511.00751)
|
||||
- page date: arXiv Oct 2025 (v2 May 2026)
|
||||
|
||||
> **Results confirm that accuracy gains plateau early and, in some configurations, decline at high sample counts** — a pattern inconsistent with diminishing returns alone and more consistent with noise introduction on problems that were already solved. This suggests self-consistency should be reserved for genuinely difficult problems rather than applied as a default scaling strategy.
|
||||
@@ -79,14 +79,14 @@ epistemic context: single-author preprint (low citation signal, flagged), and th
|
||||
|
||||
## Context rot: long inputs and rubrics degrade judging
|
||||
|
||||
## "NoLiMa: Long-Context Evaluation Beyond Literal Matching" — Modarressi et al., ICML 2025 — [arXiv:2502.05167](https://arxiv.org/abs/2502.05167)
|
||||
## "NoLiMa: Long-Context Evaluation Beyond Literal Matching" — Modarressi et al., ICML 2025 — [arXiv:2502.05167](https://arxiv.org/pdf/2502.05167)
|
||||
- page date: arXiv Feb 2025; ICML 2025. Repo: [github.com/adobe-research/NoLiMa](https://github.com/adobe-research/NoLiMa)
|
||||
|
||||
> While they perform well in short contexts (<1K), performance degrades significantly as context length increases. **At 32K, for instance, 11 models drop below 50% of their strong short-length baselines.** Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%.
|
||||
|
||||
epistemic context: peer-reviewed; removes literal lexical overlap so the test measures latent-association retrieval, the closest analog to a judge matching a rubric to a semantically-distant answer. The paper defines "effective length as the maximum length at which the score remains above a threshold, set at 85% of the model's base score" (Table 3). Effective lengths are shorter than they sound: 1-4K tokens for most of the 13 models (median 2K), 8K for GPT-4o, and 16K for GPT-4.1 in the extended Table 10.
|
||||
|
||||
## "Lost in the Middle: How Language Models Use Long Contexts" — Liu et al., TACL 2024 — [arXiv:2307.03172](https://arxiv.org/abs/2307.03172)
|
||||
## "Lost in the Middle: How Language Models Use Long Contexts" — Liu et al., TACL 2024 — [arXiv:2307.03172](https://arxiv.org/pdf/2307.03172)
|
||||
- page date: arXiv July 2023; TACL 2024
|
||||
|
||||
> We find that performance can degrade significantly when changing the position of relevant information, indicating that current language models do not robustly make use of information in long input contexts. **In particular, we observe that performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models.** Our analysis provides a better understanding of how language models use their input context and provides new evaluation protocols for future long-context language models.
|
||||
|
||||
+13
-13
@@ -157,9 +157,9 @@ The sections above are about pairwise preference judging. This one is about the
|
||||
|
||||
For a worked example, wassname has a ~300-line async OpenRouter judge (WIP) that implements many of these: bounded thinking, pinned quantisation, a versioned eval, JSON-schema output, JSONL of everything, OpenRouter error handling, and position-bias swapping: [gist](https://gist.github.com/wassname/b7f76e42de131887c02d9e9835be80ef). The same gist has `judge_inspect.py`, the inspect-ai port (`.eval` logs instead of JSONL, epochs for the repeat passes, a provider subclass that retries OpenRouter's transient-status-in-HTTP-400), and `audit.py`, a scout scanner for the identical-arms / refusal / truncation / saturation checks that runs both inline and over saved logs.
|
||||
|
||||
[^zheng]: Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (NeurIPS 2023) — https://arxiv.org/abs/2306.05685
|
||||
[^wang]: Wang et al., "Large Language Models are not Fair Evaluators" (ACL 2024) — https://arxiv.org/abs/2305.17926
|
||||
[^panickssery]: Panickssery, Bowman, Feng, "LLM Evaluators Recognize and Favor Their Own Generations" (2024) — https://arxiv.org/abs/2404.13076
|
||||
[^zheng]: Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (NeurIPS 2023) — https://arxiv.org/pdf/2306.05685
|
||||
[^wang]: Wang et al., "Large Language Models are not Fair Evaluators" (ACL 2024) — https://arxiv.org/pdf/2305.17926
|
||||
[^panickssery]: Panickssery, Bowman, Feng, "LLM Evaluators Recognize and Favor Their Own Generations" (2024) — https://arxiv.org/pdf/2404.13076
|
||||
[^verdict]: Haize Labs, verdict docs: [best practices](https://verdict.haizelabs.com/docs/best-practices/), [distributional bias cookbook](https://verdict.haizelabs.com/docs/cookbook/distributional-bias/)
|
||||
[^hamel]: Hamel Husain, "Creating a LLM-as-a-Judge That Drives Business Results" (2024) — https://hamel.dev/blog/posts/llm-judge/ (critique-shadowing workflow: look at the data first, iterate the prompt with a domain expert, prefer binary pass/fail) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
[^databricks]: Databricks, "Best Practices for LLM Evaluation of RAG Applications" (2023) — https://www.databricks.com/blog/LLM-auto-eval-best-practices-RAG (use a low-precision 0-3 / 1-5 scale; few-shot examples help weak judges but shift the score distribution) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
@@ -168,18 +168,18 @@ For a worked example, wassname has a ~300-line async OpenRouter judge (WIP) that
|
||||
[^judgemark]: EQ-Bench, "Judgemark v4" — https://eqbench.com/judgemark-v4.html (meta-eval of a model's judging ability, scored by how well its ratings separate stronger from weaker writing; leaderboard shows cost per model)
|
||||
[^speechmap]: SpeechMap.ai — https://speechmap.ai/ (refusal / completion rates across providers on contentious prompts; useful for spotting a judge that will refuse ambiguous or red-teaming scenarios)
|
||||
[^epoch]: Epoch AI, "Output length" data insight — https://epoch.ai/data-insights/output-length (reasoning models emit ~8x more tokens than non-reasoning; medium->high effort = 1.6x; reasoning-model response length growing ~5x/year)
|
||||
[^r1]: DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning" (2025) — https://arxiv.org/abs/2501.12948 (pure-RL reasoning; response length / test-time compute grows over training and with problem difficulty)
|
||||
[^survey]: Gu et al., "A Survey on LLM-as-a-Judge" (2024) — https://arxiv.org/abs/2411.15594 (broad survey of methods, biases, and reliability; complements Yan's practitioner review)
|
||||
[^doddapaneni]: Doddapaneni, Khan, Verma, Khapra, "Finding Blind Spots in Evaluator LLMs with Interpretable Checklists" (2024) — https://arxiv.org/abs/2406.13439 (evaluator LLMs missed injected quality drops in >50% of cases) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
[^shankar]: Shankar, Zamfirescu-Pereira, Hartmann, Parameswaran, Arawjo, "Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences" (2024) — https://arxiv.org/abs/2404.12272 (criteria drift; LLM evaluators need human validation) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
[^r1]: DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning" (2025) — https://arxiv.org/pdf/2501.12948 (pure-RL reasoning; response length / test-time compute grows over training and with problem difficulty)
|
||||
[^survey]: Gu et al., "A Survey on LLM-as-a-Judge" (2024) — https://arxiv.org/pdf/2411.15594 (broad survey of methods, biases, and reliability; complements Yan's practitioner review)
|
||||
[^doddapaneni]: Doddapaneni, Khan, Verma, Khapra, "Finding Blind Spots in Evaluator LLMs with Interpretable Checklists" (2024) — https://arxiv.org/pdf/2406.13439 (evaluator LLMs missed injected quality drops in >50% of cases) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
[^shankar]: Shankar, Zamfirescu-Pereira, Hartmann, Parameswaran, Arawjo, "Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences" (2024) — https://arxiv.org/pdf/2404.12272 (criteria drift; LLM evaluators need human validation) ([cache](../docs/evidence/llm_judge_biases.md))
|
||||
[^lechmazur]: Lech Mazur, position_bias benchmark — https://github.com/lechmazur/position_bias (independent, outsider-run swapped-order harness; 193 pairs, 36 models, 2026-era; strong trust signal, but a solo-run leaderboard not a paper) ([litreview](llm_judge_litreview.md))
|
||||
[^shi]: Shi et al., "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" (IJCNLP-AACL 2025) — https://arxiv.org/abs/2406.07791 (peer-reviewed; 150k+ instances, 15 judges; bias worsens as the answer quality gap shrinks) ([litreview](llm_judge_litreview.md))
|
||||
[^selfpref]: Chen et al., "Beyond the Surface: Measuring Self-Preference in LLM Judgments" (EMNLP 2025 main) — https://arxiv.org/abs/2506.02592 (peer-reviewed; DBG nets out quality; larger judges less self-biased, though authors credit capability; reasoning models still biased) ([litreview](llm_judge_litreview.md))
|
||||
[^judgelrm]: Chen et al., "JudgeLRM: Large Reasoning Models as a Judge" (2025) — https://arxiv.org/abs/2504.00050 (single-group preprint, not independently replicated; RL-trained reasoning judges beat same-size SFT, ~+8 F1 headline in body) ([litreview](llm_judge_litreview.md))
|
||||
[^overthink]: Ghosal et al., "Does Thinking More always Help? ... Mirage of Test-Time Scaling in Reasoning Models" (2025) — https://arxiv.org/abs/2506.04210 (preprint; one controlled study, accuracy-vs-thinking-token curve is non-monotonic, peak setup-specific) ([litreview](llm_judge_litreview.md))
|
||||
[^shi]: Shi et al., "Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge" (IJCNLP-AACL 2025) — https://arxiv.org/pdf/2406.07791 (peer-reviewed; 150k+ instances, 15 judges; bias worsens as the answer quality gap shrinks) ([litreview](llm_judge_litreview.md))
|
||||
[^selfpref]: Chen et al., "Beyond the Surface: Measuring Self-Preference in LLM Judgments" (EMNLP 2025 main) — https://arxiv.org/pdf/2506.02592 (peer-reviewed; DBG nets out quality; larger judges less self-biased, though authors credit capability; reasoning models still biased) ([litreview](llm_judge_litreview.md))
|
||||
[^judgelrm]: Chen et al., "JudgeLRM: Large Reasoning Models as a Judge" (2025) — https://arxiv.org/pdf/2504.00050 (single-group preprint, not independently replicated; RL-trained reasoning judges beat same-size SFT, ~+8 F1 headline in body) ([litreview](llm_judge_litreview.md))
|
||||
[^overthink]: Ghosal et al., "Does Thinking More always Help? ... Mirage of Test-Time Scaling in Reasoning Models" (2025) — https://arxiv.org/pdf/2506.04210 (preprint; one controlled study, accuracy-vs-thinking-token curve is non-monotonic, peak setup-specific) ([litreview](llm_judge_litreview.md))
|
||||
[^cais]: CAIS simple-evals .env.example vs litellm constants.py (fetched 2026-07, directly verifiable config) — https://github.com/centerforaisafety/simple-evals/blob/main/.env.example (effort high=24576/med=8192/low=1024, overriding litellm stock 4096/2048/1024)
|
||||
[^loo]: Loo, "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" (2025) — https://arxiv.org/abs/2511.00751 (single-author preprint, low citation signal; plateau ~N=10-15 on modern models, can decline past it) ([litreview](llm_judge_litreview.md))
|
||||
[^nolima]: Modarressi et al., "NoLiMa: Long-Context Evaluation Beyond Literal Matching" (ICML 2025) — https://arxiv.org/abs/2502.05167 (peer-reviewed; effective length = length holding 85% of base score; most models below half by 32K once literal cues removed) ([litreview](llm_judge_litreview.md))
|
||||
[^loo]: Loo, "Self-Consistency Is Losing Its Edge: Diminishing Returns and Rising Costs in Modern LLMs" (2025) — https://arxiv.org/pdf/2511.00751 (single-author preprint, low citation signal; plateau ~N=10-15 on modern models, can decline past it) ([litreview](llm_judge_litreview.md))
|
||||
[^nolima]: Modarressi et al., "NoLiMa: Long-Context Evaluation Beyond Literal Matching" (ICML 2025) — https://arxiv.org/pdf/2502.05167 (peer-reviewed; effective length = length holding 85% of base score; most models below half by 32K once literal cues removed) ([litreview](llm_judge_litreview.md))
|
||||
[^petri]: Petri 3.0 judge — https://github.com/meridianlabs-ai/inspect_petri/blob/main/src/inspect_petri/_judge/judge.py (dynamic `create_model` answer schema; `JUDGE_PROMPT` with the prefill-attribution caps and the empty-transcript rule; refusal returns `Result(value=None, metadata={"refusal": True})`). Maintained by Meridian Labs, used in Anthropic's alignment audits; strong trust signal as engineering, but it is one team's design, not a measured result.
|
||||
[^petri-dim]: Petri judge dimensions — https://github.com/meridianlabs-ai/inspect_petri/tree/main/src/inspect_petri/_judge/dimensions (one markdown file per dimension with YAML front matter; `concerning.md` quoted above)
|
||||
[^petri-docs]: Petri docs, results — https://github.com/meridianlabs-ai/inspect_petri/blob/main/docs/using/results.qmd (`inspect score ... --action append` to re-judge a saved log; read the audit-quality dimensions first)
|
||||
|
||||
@@ -218,6 +218,6 @@ Less central but useful:
|
||||
[^steinhardt]: Jacob Steinhardt, "Research as a Stochastic Decision Process" - https://cs.stanford.edu/~jsteinhardt/ResearchasaStochasticDecisionProcess.html ([cache](../docs/evidence/steinhardt_research_stochastic_decision_process.md)).
|
||||
[^spinningup]: Joshua Achiam, "Spinning Up as a Deep RL Researcher" (OpenAI, 2018-10-13) - https://spinningup.openai.com/en/latest/spinningup/spinningup.html ([research cache](../docs/evidence/spinningup_research_source_graph.md), [debugging cache](../docs/evidence/spinningup_researcher.md)).
|
||||
[^rahtz]: Matthew Rahtz, "Lessons Learned Reproducing a Deep Reinforcement Learning Paper" (2018) - http://amid.fish/reproducing-deep-rl ([cache](../docs/evidence/amid_fish_reproducing_deep_rl.md)).
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) - https://arxiv.org/abs/1709.06560 ([cache](../docs/evidence/henderson_2018_deep_rl_matters.md)).
|
||||
[^henderson]: Henderson et al., "Deep Reinforcement Learning that Matters" (AAAI 2018) - https://arxiv.org/pdf/1709.06560 ([cache](../docs/evidence/henderson_2018_deep_rl_matters.md)).
|
||||
[^schulman]: John Schulman, "Nuts and Bolts of Deep RL Research" (2016) - http://joschu.net/docs/nuts-and-bolts.pdf ([cache](../docs/evidence/joschu_nuts_and_bolts.md)).
|
||||
[^irpan]: Alex Irpan, "Deep Reinforcement Learning Doesn't Work Yet" (2018) - https://www.alexirpan.com/2018/02/14/rl-hard.html ([cache](../docs/evidence/alexirpan_rl_hard.md)).
|
||||
|
||||
+1
-1
@@ -149,7 +149,7 @@ usually more precise than saying the market personally moves against the model.
|
||||
[^fpp3-predictability]: Rob J. Hyndman and George Athanasopoulos, *Forecasting: Principles and Practice*, 3rd ed., ["What can be forecast?"](https://otexts.com/fpp3/what-can-be-forecast.html) ([local book](../docs/evidence/fpp3/01-getting-started.md#11-what-can-be-forecast)). This is the authors' own organizing framework, with residential electricity demand and currency exchange rates as contrasting examples.
|
||||
[^fpp3-accuracy]: Hyndman and Athanasopoulos, ["Evaluating point forecast accuracy"](https://otexts.com/fpp3/accuracy.html) ([local book](../docs/evidence/fpp3/05-forecasters-toolbox.md#58-evaluating-point-forecast-accuracy)).
|
||||
[^fpp3-tscv]: Hyndman and Athanasopoulos, ["Time series cross-validation"](https://otexts.com/fpp3/tscv.html) ([local book](../docs/evidence/fpp3/05-forecasters-toolbox.md#510-time-series-cross-validation)).
|
||||
[^cerqueira]: Vitor Cerqueira, Luis Torgo, and Igor Mozetic, ["Evaluating time series forecasting models: an empirical study on performance estimation methods"](https://arxiv.org/abs/1905.11744), *Machine Learning* 109 (2020), 1997-2028. This is empirical evidence across real and synthetic series, not a universal proof that one split is always best.
|
||||
[^cerqueira]: Vitor Cerqueira, Luis Torgo, and Igor Mozetic, ["Evaluating time series forecasting models: an empirical study on performance estimation methods"](https://arxiv.org/pdf/1905.11744), *Machine Learning* 109 (2020), 1997-2028. This is empirical evidence across real and synthetic series, not a universal proof that one split is always best.
|
||||
[^fpp3-missing]: Hyndman and Athanasopoulos, ["Dealing with outliers and missing values"](https://otexts.com/fpp3/missing-outliers.html) ([local book](../docs/evidence/fpp3/13-practical-issues.md#139-dealing-with-outliers-and-missing-values)).
|
||||
[^fpp3-stationarity]: Hyndman and Athanasopoulos, ["Stationarity and differencing"](https://otexts.com/fpp3/stationarity.html) ([local book](../docs/evidence/fpp3/09-arima-models.md#91-stationarity-and-differencing)).
|
||||
[^fpp3-long]: Hyndman and Athanasopoulos, ["Very long and very short time series"](https://otexts.com/fpp3/long-short-ts.html) ([local book](../docs/evidence/fpp3/13-practical-issues.md#137-very-long-and-very-short-time-series)).
|
||||
|
||||
+17
-17
@@ -103,12 +103,12 @@ Hermes 4 and Qwen3 both lean on filtered synthetic/verifiable data, but with gua
|
||||
|
||||
When debugging or designing a modern transformer run, read reports that disclose the model-building process rather than only final benchmark scores:
|
||||
|
||||
- [Olmo 3](https://arxiv.org/abs/2512.13961) releases the "entire model flow," including stages, checkpoints, data, and dependencies; code lives in [OLMo-core](https://github.com/allenai/OLMo-core).
|
||||
- [Olmo 3](https://arxiv.org/pdf/2512.13961) releases the "entire model flow," including stages, checkpoints, data, and dependencies; code lives in [OLMo-core](https://github.com/allenai/OLMo-core).
|
||||
- Microsoft's [MAI-Thinking-1](https://microsoft.ai/pdf/mai-thinking-1.pdf) treats model development as a system-level optimization problem and gives a long-form account of scaling and RL decisions.
|
||||
- Nous Research's [Hermes 4](https://arxiv.org/abs/2508.18255) describes failures and solutions across data curation, synthesis, training, and evaluation; Nous also releases open training/evaluation tooling such as [Atropos](https://github.com/NousResearch/atropos).
|
||||
- [DeepSeek-V3](https://arxiv.org/abs/2412.19437) reports architecture, infrastructure, training, and a run with no irrecoverable loss spikes or rollbacks.
|
||||
- [Qwen3](https://arxiv.org/abs/2505.09388) documents a dense/MoE family from `0.6B` to `235B`, including pretraining and post-training details.
|
||||
- Secondary postmortems: [The Llama 3 Herd](https://arxiv.org/abs/2407.21783) for large-scale pretraining operations, and [OPT-175B](https://arxiv.org/abs/2205.01068) for training interruptions, instability, and mid-flight recovery.
|
||||
- Nous Research's [Hermes 4](https://arxiv.org/pdf/2508.18255) describes failures and solutions across data curation, synthesis, training, and evaluation; Nous also releases open training/evaluation tooling such as [Atropos](https://github.com/NousResearch/atropos).
|
||||
- [DeepSeek-V3](https://arxiv.org/pdf/2412.19437) reports architecture, infrastructure, training, and a run with no irrecoverable loss spikes or rollbacks.
|
||||
- [Qwen3](https://arxiv.org/pdf/2505.09388) documents a dense/MoE family from `0.6B` to `235B`, including pretraining and post-training details.
|
||||
- Secondary postmortems: [The Llama 3 Herd](https://arxiv.org/pdf/2407.21783) for large-scale pretraining operations, and [OPT-175B](https://arxiv.org/pdf/2205.01068) for training interruptions, instability, and mid-flight recovery.
|
||||
|
||||
These are useful as working implementations and experiment logs: copy proven priors, compare the exact computation graph and recipe, and look for engineering details absent from method papers.
|
||||
|
||||
@@ -119,23 +119,23 @@ For experiment design, keep the [Google Deep Learning Tuning Playbook](https://d
|
||||
[^hfcourse]: Hugging Face LLM Course, ["Debugging the training pipeline"](https://huggingface.co/learn/llm-course/chapter8/4) ([cache](../docs/evidence/hf_llm_course_ch8_4_debugging_pipeline.md))
|
||||
[^axolotl-stability]: Axolotl, ["Training Stability"](https://docs.axolotl.ai/docs/training_stability.html) ([cache](../docs/evidence/axolotl_training_stability.md))
|
||||
[^unsloth]: Unsloth, ["Troubleshooting & FAQs"](https://docs.unsloth.ai/basics/troubleshooting-and-faqs) ([cache](../docs/evidence/unsloth_troubleshooting_faqs.md))
|
||||
[^goyal]: Goyal et al., ["Accurate, Large Minibatch SGD"](https://arxiv.org/abs/1706.02677)
|
||||
[^super-convergence]: Smith and Topin, ["Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates"](https://arxiv.org/abs/1708.07120)
|
||||
[^wsd]: Wen et al., ["Understanding Warmup-Stable-Decay Learning Rates: A River Valley Loss Landscape Perspective"](https://arxiv.org/abs/2410.05192)
|
||||
[^goyal]: Goyal et al., ["Accurate, Large Minibatch SGD"](https://arxiv.org/pdf/1706.02677)
|
||||
[^super-convergence]: Smith and Topin, ["Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates"](https://arxiv.org/pdf/1708.07120)
|
||||
[^wsd]: Wen et al., ["Understanding Warmup-Stable-Decay Learning Rates: A River Valley Loss Landscape Perspective"](https://arxiv.org/pdf/2410.05192)
|
||||
[^nanochat]: Karpathy, [nanochat experiment log](https://github.com/karpathy/nanochat/blob/main/dev/LOG.md) ([cache](../docs/evidence/nanochat_deepwiki_llm_pretraining_2026.md))
|
||||
[^karpathy-recipe]: Karpathy, ["A Recipe for Training Neural Networks"](https://karpathy.github.io/2019/04/25/recipe/) ([cache](../docs/evidence/karpathy_recipe_training_nn_2019.md))
|
||||
[^nanochat-optimizer]: Karpathy, [`nanochat`](https://github.com/karpathy/nanochat) (`optim.py`: AdamW + Muon)
|
||||
[^optimizer-benchmark]: Wen et al., ["Fantastic Pretraining Optimizers and Where to Find Them"](https://arxiv.org/abs/2509.02046) (ICLR 2026)
|
||||
[^optimizer-benchmark]: Wen et al., ["Fantastic Pretraining Optimizers and Where to Find Them"](https://arxiv.org/pdf/2509.02046) (ICLR 2026)
|
||||
[^tuning-playbook]: Google Developers, ["Deep Learning Tuning Playbook"](https://developers.google.com/machine-learning/guides/deep-learning-tuning-playbook)
|
||||
[^bekman]: Stas Bekman, [`DebugUnderflowOverflow`](https://github.com/huggingface/transformers/blob/main/src/transformers/debug_utils.py) ([cache](../docs/evidence/bekman_debug_utils_transformers.md))
|
||||
[^eval-awareness]: Chaudhary et al., ["Evaluation Awareness Scales Predictably in Open-Weights Large Language Models"](https://arxiv.org/abs/2509.13333)
|
||||
[^steering-reliability]: Braun et al., ["Understanding (Un)Reliability of Steering Vectors in Language Models"](https://arxiv.org/abs/2505.22637)
|
||||
[^olmo3-report]: OLMo Team, ["Olmo 3"](https://arxiv.org/abs/2512.13961) ([cache](../docs/evidence/reports/olmo3_technical_report.md); [OLMo-core](https://github.com/allenai/OLMo-core), [cache](../docs/evidence/reports/code/olmo_core_readme.md))
|
||||
[^eval-awareness]: Chaudhary et al., ["Evaluation Awareness Scales Predictably in Open-Weights Large Language Models"](https://arxiv.org/pdf/2509.13333)
|
||||
[^steering-reliability]: Braun et al., ["Understanding (Un)Reliability of Steering Vectors in Language Models"](https://arxiv.org/pdf/2505.22637)
|
||||
[^olmo3-report]: OLMo Team, ["Olmo 3"](https://arxiv.org/pdf/2512.13961) ([cache](../docs/evidence/reports/olmo3_technical_report.md); [OLMo-core](https://github.com/allenai/OLMo-core), [cache](../docs/evidence/reports/code/olmo_core_readme.md))
|
||||
[^mai-thinking-report]: Microsoft AI Team, ["MAI-Thinking-1: Building a Hill-Climbing Machine"](https://microsoft.ai/pdf/mai-thinking-1.pdf) ([cache](../docs/evidence/reports/mai_thinking_1_technical_report.md))
|
||||
[^hermes4-report]: Nous Research, ["Hermes 4 Technical Report"](https://arxiv.org/abs/2508.18255) ([cache](../docs/evidence/reports/hermes4_technical_report.md); [Atropos](https://github.com/NousResearch/atropos), [cache](../docs/evidence/reports/code/nous_atropos_readme.md))
|
||||
[^deepseek-v3-report]: DeepSeek-AI, ["DeepSeek-V3 Technical Report"](https://arxiv.org/abs/2412.19437) ([cache](../docs/evidence/reports/deepseek_v3_technical_report.md))
|
||||
[^qwen3-report]: Qwen Team, ["Qwen3 Technical Report"](https://arxiv.org/abs/2505.09388) ([cache](../docs/evidence/reports/qwen3_technical_report.md))
|
||||
[^llama3-report]: Meta AI, ["The Llama 3 Herd of Models"](https://arxiv.org/abs/2407.21783) ([cache](../docs/evidence/reports/llama3_herd_technical_report.md))
|
||||
[^opt175b-report]: Zhang et al., ["OPT: Open Pre-trained Transformer Language Models"](https://arxiv.org/abs/2205.01068) ([cache](../docs/evidence/reports/opt175b_technical_report.md))
|
||||
[^hermes4-report]: Nous Research, ["Hermes 4 Technical Report"](https://arxiv.org/pdf/2508.18255) ([cache](../docs/evidence/reports/hermes4_technical_report.md); [Atropos](https://github.com/NousResearch/atropos), [cache](../docs/evidence/reports/code/nous_atropos_readme.md))
|
||||
[^deepseek-v3-report]: DeepSeek-AI, ["DeepSeek-V3 Technical Report"](https://arxiv.org/pdf/2412.19437) ([cache](../docs/evidence/reports/deepseek_v3_technical_report.md))
|
||||
[^qwen3-report]: Qwen Team, ["Qwen3 Technical Report"](https://arxiv.org/pdf/2505.09388) ([cache](../docs/evidence/reports/qwen3_technical_report.md))
|
||||
[^llama3-report]: Meta AI, ["The Llama 3 Herd of Models"](https://arxiv.org/pdf/2407.21783) ([cache](../docs/evidence/reports/llama3_herd_technical_report.md))
|
||||
[^opt175b-report]: Zhang et al., ["OPT: Open Pre-trained Transformer Language Models"](https://arxiv.org/pdf/2205.01068) ([cache](../docs/evidence/reports/opt175b_technical_report.md))
|
||||
[^tinker-dpo]: Thinking Machines, [tinker-cookbook recipes/preference/dpo README](https://github.com/thinking-machines-lab/tinker-cookbook/tree/main/tinker_cookbook/recipes/preference/dpo)
|
||||
[^tinker-rlhf]: Thinking Machines, [tinker-cookbook recipes/preference/rlhf README](https://github.com/thinking-machines-lab/tinker-cookbook/tree/main/tinker_cookbook/recipes/preference/rlhf)
|
||||
|
||||
+2
-2
@@ -149,9 +149,9 @@ Sometimes (rarely) you don't. Schulman:
|
||||
### Articles
|
||||
- Andy Jones, "Debugging RL, Without the Agonizing Pain" (2021): https://andyljones.com/posts/rl-debugging.html
|
||||
- Matthew Rahtz, "Lessons Learned Reproducing a Deep RL Paper" (2018): http://amid.fish/reproducing-deep-rl
|
||||
- Henderson et al., "Deep Reinforcement Learning that Matters" (2018): https://arxiv.org/abs/1709.06560
|
||||
- Henderson et al., "Deep Reinforcement Learning that Matters" (2018): https://arxiv.org/pdf/1709.06560
|
||||
- Alex Irpan, "Deep Reinforcement Learning Doesn't Work Yet" (2018): https://www.alexirpan.com/2018/02/14/rl-hard.html
|
||||
- McCandlish & Kaplan, "An Empirical Model of Large-Batch Training" (2018): https://arxiv.org/abs/1812.06162
|
||||
- McCandlish & Kaplan, "An Empirical Model of Large-Batch Training" (2018): https://arxiv.org/pdf/1812.06162
|
||||
- Slav Ivanov, "37 Reasons why your Neural Network is not working" (2017): https://blog.slavv.com/37-reasons-why-your-neural-network-is-not-working-4020854bd607
|
||||
- Huang et al., "The 37 Implementation Details of PPO" (ICLR Blog Track 2022): https://iclr-blog-track.github.io/2022/03/25/ppo-implementation-details/ ([cache](../docs/evidence/cleanrl_37_ppo_details.md))
|
||||
- Joshua Achiam, "Spinning Up as a Deep RL Researcher" (OpenAI, 2018): https://spinningup.openai.com/en/latest/spinningup/spinningup.html ([cache](../docs/evidence/spinningup_researcher.md))
|
||||
|
||||
Reference in New Issue
Block a user