weight-steering/docs/fork_plan.md

Fork plan: weight steering benchmark + analysis

Updated: 2026-04-27

Goal

Test whether weight steering is a useful method, and if it is, understand what part of the learned weight delta carries the behavior.

Two questions are intentionally separated:

1. Benchmark question: Does weight steering beat simple alternatives such as prompting and activation steering on sycophancy and daily-dilemmas honesty transfer?
2. Analysis question: If weight steering works, can the learned delta dW = \theta^+ - \theta^- be factorized into a simpler causal intervention: a cross-adapter shared subspace, module, low-rank component, or adapter parameterization?

Context

This is a fork of Anthropic's weight-steering method. Original recipe: train one positive adapter and one negative adapter, merge each adapter into base-weight deltas, then steer with:

dW = \Delta W_{pos} - \Delta W_{neg}.

This repo removes Axolotl/vLLM/API orchestration and rebuilds the method in HF + PEFT + uv for cheap iteration on small models.

Current main model: Qwen/Qwen3-0.6B.

Current behavior: honesty training (positive = honest persona, negative = dishonest persona), evaluated on wassname/daily_dilemmas-self-honesty (OOD) and held-out sycophancy Yes/No claims (IID).

Links

• Paper / blog:
  • docs/weight_steering_paper.md
  • docs/weight_steer_blog.md
• Adapter-as-hypothesis notes:
  • docs/blog_adapter_as_hypothesis/README.md
• Steering/subspace concepts:
  • docs/AntiPaSTO_concepts/README.md
• Current user-facing summaries:
  • README.md
  • RESEARCH_JOURNAL.md
• Key code:
  • src/ws/data.py
  • src/ws/train.py
  • src/ws/diff.py
  • src/ws/steer.py
  • src/ws/eval/sycophancy.py
  • src/ws/eval/dilemmas.py
  • src/ws/eval/cross_adapter_ablation.py
  • src/ws/eval/layer_module_ablation.py
  • src/ws/eval/parameterization_ablation.py
  • nbs/ablation_analysis.py

Current facts

• Daily-dilemmas default is not full split. Default n_dilemmas=100 means first 100 dilemmas = 200 rows, balanced 100 honest-label and 100 dishonest-label actions.
• Full honesty_eval test split is 219 dilemmas = 438 rows.
• The daily-dilemmas eval uses all rows for selected dilemmas, then sign-flips by honesty_label; it is not only honest rows.
• Current headline tables are single-seed Qwen3-0.6B exploratory results.
• DeLoRA is best raw steering so far. PiSSA is the cleaner stable baseline if penalizing DeLoRA saturation at high alpha.
• v9/v10 do not prove “no subspace.” They show the trained behavior is not explained by the tested low-rank residual-stream bases or adapter-family parameterization at trained scale.
• The active analysis should ablate the already-trained dW. Synthetic dW' construction is a different baseline, not causal ablation.
• The highest-value analysis tests are: cross-adapter causal dW basis ablation, layer/module ablation of trained dW, and adapter-parameterization ablation of trained dW.
• Lens search is on hold pending multiseed (2026-04-27). Every weight-space lens we tested has a built-in failure mode: SVD-on-dW is tautological for low-rank adapters; layer-index tells depth not mechanism; module-family collapses heads/positions and gives different answers per adapter; native parameterization decompositions aren't comparable across adapter families. But the lens-3 cross-adapter inconsistency (delora residual_write retained=+1.27 vs lora=+0.14) is N=1 seed × N=1 model. It might just be seed noise within each adapter. Right ordering: T4 multiseed first, then re-run T7/T8 per-seed with within-adapter stdev, then judge whether the inconsistency is real or noise.

Done

• Clean repo into uv + HF + PEFT small-model workflow.
• Make Qwen3-0.6B sycophancy steering work end-to-end.
• Hook in LoRA, DoRA, PiSSA, DeLoRA, OFT, and IA3 adapter families.
• Build sycophancy logratio eval with coefficient sweep.
• Build daily-dilemmas honesty eval with sign-flipped Yes/No logratio.
• Run single-seed Qwen adapter benchmark on sycophancy and 100-dilemma DD default.
• Fix DD cross-adapter aggregation to use base-only coeff=0 rather than mixing persona baselines.
• Run v9 subspace/scope diagnostics: weight oracle, cumulative activation oracle, block-local activation oracle, first-LoRA-layer sanity checks.
• Run v10 projection/complement falsifier: raw activation projection, complement, and normmatched projection.
• Update README and research journal with corrected DD table and conservative interpretation.

TODO: benchmark question

• Goal: activation-steering baseline on the same DD rows.

  • Why: RepE/repeng is the most threatening baseline; if it matches or beats dW, the method story weakens before adapter seeds matter.
  • Do: train representation direction on the same sycophancy contrast; grid layer x coefficient; evaluate sycophancy and full DD.
  • UAT: best activation-steering row is selected by held-out sycophancy or validation DD, then reported beside best dW on identical DD test rows.
  • Verify: table includes method=repeng, layer, coeff, syc_delta, dd_delta, pmass, and the same idx set as the dW rows.
  • Negative outcome -> claim: if repeng matches/beats dW, write "activation steering is the simpler baseline; weight steering needs a stronger reason to exist."

• Goal: full daily-dilemmas benchmark for current Qwen adapters.

  • Why: current DD table uses first 100 dilemmas, not the full 219-dilemma split.
  • Do: re-run LoRA / PiSSA / DeLoRA / DoRA / OFT / IA3 with --n-dilemmas 219.
  • UAT: table has 438 base rows per coeff before persona baselines, and reports pmass, frac_low_pmass, delta(+1 - 0).
  • Verify: out/sycophancy/cross_adapter_full_dd/dilemmas_summary.csv exists and includes n_base_rows_per_coeff=438.

• Goal: prompt baselines on the same DD rows.

  • Why: weight steering is only interesting if it beats “just prompt it.”
  • Do: evaluate base, simple honest persona, and engineered AxBench-style prompt.
  • UAT: one table compares base, simple_honest_prompt, engineered_prompt, and best dW on identical rows.
  • Verify: prompt_baseline_delta and weight_steer_delta are computed from the same idx set.
  • Negative outcome -> claim: if prompting matches/beats dW, write "prompting is the simpler intervention for this behavior/eval pair."

• Goal: multi-seed adapter benchmark on Qwen.

  • Why: current adapter ranking is N=1 seed.
  • Do: run seeds 0, 1, 2 for LoRA / PiSSA / DeLoRA first; add DoRA/OFT only if cheap.
  • UAT: table reports mean +/- std for sycophancy and DD deltas, plus seed-level signs, so a reader can tell stable ranking from noisy N=1 luck.
  • Verify: each adapter has exactly three w.safetensors files and three eval summaries; ranking table includes n_seeds=3, mean_dd_delta, std_dd_delta, and sign_agreement.
  • Negative outcome -> claim: if adapter ranking changes across seeds or error bars overlap heavily, write "single-seed adapter winner is unstable; do not claim a family ranking yet."

• Goal: Gemma 1B replication.

  • Why: check whether DeLoRA/PiSSA ranking is Qwen-specific.
  • Do: train LoRA / PiSSA / DeLoRA on Gemma 1B, seed 0, full DD split.
  • UAT: compare Gemma ranking to Qwen ranking with the same metrics.
  • Verify: table has model column with Qwen3-0.6B and Gemma-1B; if DeLoRA remains best, expand seeds; if rankings diverge, write that up as a model-specific adapter-basin finding.

TODO: analysis question

Status (2026-04-27): on hold pending multiseed. T6/T7/T8 are run on N=1 seed × Qwen3-0.6B. Necessity is established. The cross-adapter inconsistency that drove the "no parameterization-invariant mechanism" reading might be seed noise. Resume after T4 (multiseed) lands and we can report within-adapter stdev alongside cross-adapter gaps.

Active sequence at the time of pause was:

1. Cross-adapter causal dW basis ablation.
2. Layer/module causal ablation of trained dW.
3. Adapter-parameterization causal ablation of trained dW.

Synthetic dW' construction is deferred below and is not a causal ablation.

• Goal: cross-adapter causal dW basis ablation.
  • Why: this is the headline analysis experiment. It tests whether different adapter families discovered the same causal planning subspace or different basins.
  • Do: build candidate bases B from trained adapter deltas, compute dW_keep_B and dW_drop_B, and evaluate both on sycophancy + full DD for each adapter.
  • Candidate B rows:
    • shared_SVD_K8/K32/K64: stack residual-output dW from LoRA / DoRA / PiSSA / DeLoRA / OFT per layer/tensor, take top-K SVs.
    • top8/top32_per_adapter and tail_per_adapter: per-adapter SVD split of each tensor.
    • random_null: rank-matched random bases.
  • UAT: one central table has ablation_family, candidate_B, adapter, rank, keep_or_drop, syc_delta, dd_delta, pmass, and row-identity checks.
  • Verify: the table contains keep/drop rows for shared_svd, per_adapter_svd, and random_null; keep_B_shared_K32 and drop_B_shared_K32 are both evaluated for at least LoRA / DoRA / PiSSA / DeLoRA / OFT; random null retention is near rank/d; each row uses the same eval rows and coefficient grid.
  • Positive outcome -> claim: if keep_B_shared retains >=0.7x behavior across adapters and drop_B_shared removes it, write the adapter-invariant planning-subspace paper.
  • Negative outcome -> claim: if keep_B_shared retains <0.3x even at K=64 while complements/tails retain behavior, write the shared-subspace negative result: steering is distributed or lives in the wrong parameter space for these bases.
  • Ambiguous outcome -> claim: if both keep and drop retain high behavior, report non-identifiability under this basis family and move to stricter causal interventions, not a positive subspace claim.

Note for the following two a search has been made of hypothesis: docs/hypothesis_ablation_catalog.md

• Goal: layer/module causal ablation of trained dW.

  • Why: after a trained update works, we need to know which layers and modules are necessary or sufficient.
  • Do: keep/drop parts of the already-trained adapter delta by layer and module family, without synthesizing new tensors from base features.
  • Rows: full_dW, residual_write_only, attn_o_proj_only, mlp_down_proj_only, layers_8_21_only, single-layer keep, leave-one-layer-out, coarse early/mid/late LoRA-layer blocks, rank/module-matched random controls, and zero.
  • UAT: one table has adapter, variant, layer_or_block, module_family, keep_or_drop, syc_delta, dd_delta, pmass, and row-identity checks.
  • Verify: same sycophancy and full DD rows as full_dW; table includes all required variants and reports zero row-key symmetric difference for every variant x coeff group.
  • Positive outcome -> claim: if a small layer/module slice retains most behavior and dropping it removes behavior, report the causal locus.
  • Negative outcome -> claim: if many disjoint slices retain behavior, report distributed or non-identifiable layer/module localization.

• Goal: adapter-parameterization causal ablation of trained dW.

  • Why: adapter families may store the behavior in different parameterization degrees of freedom even when their effective dW looks similar.
  • Do: split the trained adapter/effective delta according to the adapter family's own coordinates, then keep/drop each component on identical eval rows. For an S-space split, compute the trained effective matrix's SVD-like coordinate system, project dW -> S, crop a component such as the top 25% of S by coordinate index, project back to weight space, and evaluate both top_25pct_S and residual_not_top_25pct_S against full_dW and zero.
  • Rows: LoRA/PiSSA/DeLoRA rank components and S-space quartiles (top_25pct_S, mid_50pct_S, bottom_25pct_S, residual_not_top_25pct_S, residual_not_bottom_25pct_S); cumulative S-energy groups (top_50pct_energy_S, top_90pct_energy_S, residuals); DoRA direction vs magnitude component; OFT rotation-derived component vs residualized effective update; IA3 attention-gate vs MLP-gate groups.
  • UAT: one table has adapter, parameterization_family, coordinate_system, component, keep_or_drop, rank_or_group, energy_frac, syc_delta, dd_delta, pmass, and row-identity checks.
  • Verify: all rows start from the trained adapter delta or trained adapter parameters; no row is constructed from base-only activations; every component shares the same sycophancy and DD row keys as full_dW; for each S-space crop, component_dW + residual_dW reconstructs full_dW within numerical tolerance.
  • Positive outcome -> claim: if one parameterization component retains most behavior and dropping it removes behavior, report which degree of freedom carries the learned behavior.
  • Negative outcome -> claim: if behavior is not localized by parameterization component, report the trained effect as distributed across that adapter parameterization.

Coverage gaps in current ablation set

The three causal ablations above (cross-adapter dW basis, layer/module, adapter parameterization) leave some hypotheses untested. These are open follow-ups, not blockers for the current writeup.

• Read-side modules in the layer/module ablation. Current variants cover residual writes (o_proj, down_proj), attention-only, and mlp-only, but not q/k/v-only or up/gate-only. Any read-side mechanism story is currently untestable.
• Base-W SVD lens for the S-space ablation. parameterization_ablation.py uses each tensor's own SVD (dW = U S Vh). The catalog also wants a separate lens using the base weight's SVD (U0, S0, V0h = svd(W_base); dS = U0.T @ dW @ V0h), which answers "does dW ride pretrained singular directions" rather than "is dW low-rank in its own basis".
• Adapter-architecture decompositions. S-space variants do not include DoRA magnitude vs direction, DeLoRA lambda vs direction, OFT rotation, or IA3 attention-gate vs MLP-gate splits.
• Norm-matched random keep control for T8 sufficiency claims. Layer/module ablation has random_norm_matched_full; the S-space crops do not. Necessity (drop) tests don't need this; sufficiency (keep) tests do, because cropping shrinks Frobenius norm and the model is nonlinear in alpha.

Deferred / optional

• Optional future: constructive synthetic dW' baseline.

  • Why: useful as a method baseline, but it is not a causal ablation of trained weight steering.
  • Do: only if separately approved, build simple dW_prime = f(W_base, persona_contrast) candidates, e.g. lm-head/readout rowspace projected persona contrast, write-not-read persona contrast, and shared structural bases with signed coefficients from activation contrast.
  • UAT: table compares synthetic dW_prime to trained dW, prompt, and repeng on identical sycophancy + DD rows.
  • Verify: candidates are generated before reading trained adapter deltas; code fails if w.safetensors is loaded before constructing dW_prime.
  • Positive outcome -> claim: if a synthetic dW_prime steers, weight steering may be replaceable by a constructive method baseline.
  • Negative outcome -> claim: if no synthetic candidate steers while trained dW does, training is doing nontrivial search not captured by the current structural recipes.

• Goal: SVD steering baseline.

  • Why: useful only if cheap and stable; lower priority than repeng.
  • UAT: same DD/sycophancy table as other baselines.
  • Verify: table includes method=svd_steering, layer, rank, coeff, syc_delta, dd_delta, and pmass.
  • Negative outcome -> claim: if SVD steering is weak or unstable, do not treat plain base-weight SVD as a competitive method baseline.

• Goal: degradation benchmark.

  • Why: steering might improve target metric while damaging general behavior.
  • UAT: perplexity or clean instruction proxy reported for best coefficients.
  • Verify: table has target metric and degradation metric for the exact same selected coefficients.
  • Negative outcome -> claim: if target gains require large degradation, report steering as brittle rather than useful.

• Goal: larger model replication.

  • Why: Qwen3-0.6B and Gemma 1B are iteration models; larger model needed for a stronger claim.
  • UAT: same benchmark table on a 4B-ish model after method stabilizes.
  • Verify: model column includes the 4B-ish model and reuses the same prompt/DD row IDs as the small-model benchmark.
  • Negative outcome -> claim: if the effect disappears or reverses on the larger model, write the small-model limitation instead of scaling the claim.

Decision rules

• If prompt or activation steering beats dW, prioritize method improvement before deeper mechanistic analysis.
• If activation steering matches dW, treat weight steering as mechanistic interest first and applied method second.
• If DeLoRA wins across Qwen and Gemma, spend seeds on DeLoRA/PiSSA only.
• If Qwen and Gemma adapter rankings diverge, write the model-specific adapter-basin finding instead of forcing one global winner.
• Shared-core rule: if keep_B_shared_K32 retains >=0.7x behavior across LoRA / DoRA / PiSSA / DeLoRA / OFT and drop_B_shared_K32 removes most of it, write the planning-subspace paper.
• Basin-divergence rule: if per-adapter top subspaces are mutually low-overlap and each adapter's own SVD keeps behavior better than B_shared, write the basin-divergence paper.
• If top-k or write-not-read keeps behavior, we found a simple steering parameterization.
• If complement/tail/many layers keep behavior, evidence favors distributed or wrong-space mechanism.
• If MLP up/gate terms carry behavior, next paper story should be feature-space steering, not residual-stream planning subspace.
• Held-out cross-adapter shared-SVD rule (contingent on T4 multiseed):
  • If T4 shows within-adapter stdev > cross-adapter retained gap on lens-3, the N=1 cross-adapter inconsistency was seed noise -> held-out shared-SVD becomes worth running.
  • If T4 confirms cross-adapter gap is real (current expectation), skip held-out shared-SVD -- it will fail by construction and only restates what is already known.

Resolved: train/eval axis switch from sycophancy to honesty (2026-04-27)

Decision: option 2 — switched the entire pipeline to honesty axis. Rationale:

• The sycophancy NEG persona was already "you are an honest assistant", so the prior contrast was effectively honesty-vs-anti-honesty with extra sycophancy framing on the POS side. Cleaner to make that explicit.
• Eval (daily_dilemmas-self-honesty) signs by honesty_label. One-axis training matches the eval axis by construction. The in-sample claim is now defensible.
• Sycophancy becomes a free OOD transfer eval (run later if useful) rather than a confounded primary axis.
• Avoids the GPT-4o sycophancy-eval generation burden (option 1) and its eval-validation overhead.

What changed in code (commit on dev):

• src/ws/data.py: added HONESTY_PROMPT = "Pretend you're {persona} person. " with 5 paraphrase paraphrases each side (HONESTY_POS_PERSONAS, HONESTY_NEG_PERSONAS) — vgel/repeng short-persona convention. Added _load_suffixes() reader for data/branching_suffixes.json (550 SSteer entries). Honesty branch reuses the suffix user_msg as the question pool; persona-conditioned generation in generate_pairs() produces (r_pos, r_neg) per the paper recipe (Fierro & Roger 2025 §F.1).
• src/ws/eval/activation_baseline.py: RepE direction extraction now branches on cfg.behavior; honesty mode captures last-token hidden states under HONESTY_POS[0] / HONESTY_NEG[0] over _load_suffixes() prompts with assistant_prefixes=suffix.
• src/ws/eval/prompt_baseline.py: replaced single engineered_prompt with paired engineered_prompt_honest + engineered_prompt_dishonest (AxBench Appendix J.2 style).
• evals/smoke.py: added behavior field; just smoke --behavior honesty passes end-to-end on katuni4ka/tiny-random-qwen3.
• data/branching_suffixes.json: copied from SSteer.

Sycophancy outputs in out/sycophancy/ are kept on disk as historical evidence for the old axis-mismatched table. The README headline numbers will be replaced with honesty once 230-236 land. T4/T5 stay open and will be re-scoped against honesty.