weight-steering/README.md

Weight Steering

Fork of Fierro & Roger 2025. Train two PEFT adapters on contrastive personas (POS vs NEG), merge into base, take dW = θ_pos − θ_neg, add α·dW at inference.

We test whether weight-space steering (dW) competes with hidden-state steering and prompting on a directly comparable Authority↓ benchmark. For dataset, persona pairs, calibration recipe, and baseline methods, see steering-lite (sl). ws shares the persona pairs, vignettes, and 1-nat KL budget so rows below drop into sl's tables. (ws = this repo; sl = steering-lite, hidden-state steering baselines.)

Results: Authority↓ on Qwen3.5-4B (iso-KL=1.0)

We ask three questions:

1. Does dW move Authority in the right direction?
2. Does dW beat hidden-state steering and persona-prompting?
3. Does dW have lower uncertainty than hidden-state steering?

Glossary

• dW = θ_pos − θ_neg: weight-space contrast from two PEFT adapters.
• α: steering strength, calibrated so worst-5%-KL hits 1 nat.
• ΔAuth: mean change in logit P(is_wrong) on Authority vignettes, paired by (vignette, condition). Negative = Authority↓ achieved.
• axis_Δ = −ΔAuth (positive = correct direction, persona-aligned).
• SI(Auth): bidirectional Surgical Informedness on Authority. High means the method moves Authority without breaking other foundations. Definition: steering-lite eval.
• prompt_only: baseline that injects the POS persona as a system prompt, no steering vector.

Δlogit and uncertainty (Auth ↓ target, Care = off-target effect)

Authority is the target (move down). Care is one off-target effect: surgical methods should leave it near zero, broadly-suppressing methods drag it down with Authority. Full 7-foundation table in out/authority/.../foundations_dlogit.csv. Bold = best per column (most-negative ΔAuth, lowest std, closest-to-zero ΔCare).

method	ΔAuth ↓ (mean ± std)	ΔCare → 0 (mean ± std)
sl:engineered_prompt	−2.98 ± 1.20	−1.64 ± 1.03
sl:sspace_ablate	−2.89 ± 0.86	−2.79 ± 0.92
sl:sspace	−2.78 ± 0.93	−2.57 ± 0.90
sl:angular_steering	−2.67 ± 0.89	−2.49 ± 0.84
sl:cosine_gated	−2.08 ± 0.64	−1.88 ± 0.61
sl:directional_ablation	−1.94 ± 1.22	−1.80 ± 1.24
sl:mean_diff	−1.93 ± 1.11	−1.72 ± 1.09
sl:mean_centred	−1.80 ± 1.17	−1.63 ± 1.14
sl:spherical	−1.44 ± 0.89	−1.21 ± 0.71
sl:pca	−1.36 ± 1.50	−1.30 ± 1.36
sl:topk_clusters	−1.18 ± 0.97	−1.12 ± 0.91
ws:delora*	−0.89 ± 0.58	−0.49 ± 0.60
sl:linear_act	−0.83 ± 0.67	−0.70 ± 0.52
sl:chars	−0.45 ± 0.61	−0.40 ± 0.54

*ws:delora calibrated at p95=0.5, not kl=1.0 — expect larger effect after re-calibration.

Surgical Informedness (headline, ↑ better)

SI(Auth), SI_fwd, SI_rev, Auth_sep, and pmass²×100 all higher is better. Bold = best in column. sl rows from sl's published Qwen3.5-4B run. ws:delora is at p95=0.5 budget (kl=1.0 re-run queued with lora/dora).

method	SI(Auth) ↑	SI_fwd ↑	SI_rev ↑	Auth_sep ↑	pmass²×100 ↑
sl:directional_ablation	52.90	0.32	+1.00	+2.05	80.1
sl:super_sspace	47.71	0.67	+0.40	+1.99	88.8
sl:sspace	45.67	0.64	+0.85	+0.69	61.0
sl:mean_diff	32.81	0.34	+1.00	+1.65	49.0
sl:mean_centred	32.72	0.29	+1.00	+1.56	50.6
sl:topk_clusters	31.34	0.13	+0.72	+1.55	73.9
sl:sspace_ablate	24.11	0.74	+0.02	+0.59	63.6
sl:linear_act	20.24	−0.19	+1.00	+0.83	49.9
ws:delora	19.03	0.02	+0.37	+0.76	99.9
sl:engineered_prompt	17.36	0.50	−0.02	+1.90	71.7
sl:cosine_gated	8.92	0.09	+1.00	+2.00	16.4
sl:angular_steering	7.00	0.55	−0.38	+0.32	80.6
sl:spherical	4.98	0.16	n/a	+0.85	30.3
sl:pca	−0.92	0.03	−0.08	+0.85	39.0
sl:chars	−9.16	−0.26	+0.00	+0.50	68.3

TL;DR

1. Did dW replicate? Yes. ws:delora ΔAuth = −0.89 (sign correct) and SI(Auth) = 19.03 — verdicts do flip in the right direction.
2. Did dW beat steering and prompting? Partially. SI = 19.03 beats the engineered-prompt baseline (17.36) and 5 other sl methods, but is below 8 hidden-state methods. ΔAuth std = 0.58 is the lowest in the table (lower uncertainty than all sl methods).
3. Did dW have lower uncertainty? Yes. ws:delora std = 0.58, lowest in the table (sl best: chars 0.61).

Open: lora and dora training queued (pueue 141-144); ws:delora is at p95=0.5 budget, not yet at sl's kl=1.0 — expect SI to shift after re-calibration. Full 4-adapter table pending.

How to run

# 1. generate persona-conditioned data
uv run python -m ws.data --behavior authority --model-id Qwen/Qwen3.5-4B
# 2. train all adapters (dW = merged_pos - merged_neg)
uv run python -m ws.run_sweep --behavior authority --model Qwen/Qwen3.5-4B
# 3. iso-KL calibrate α
uv run python -m ws.kl_calibrate --behavior authority --model Qwen/Qwen3.5-4B
# 4. eval on tinymfv airisk
uv run python -m ws.scripts.eval_tinymfv_calibrated --behavior authority --model Qwen/Qwen3.5-4B
# 5. rebuild README tables
uv run python -m ws.scripts.readme_tinymfv_table --behavior authority

Outputs go to out/authority/<adapter>/. Smoke test on a tiny model: just smoke.

Cite

@article{FierroRoger2025,
  author    = {Constanza Fierro and Fabien Roger},
  title     = {Steering Language Models with Weight Arithmetic},
  journal   = {arXiv preprint arXiv:2511.05408},
  year      = {2025},
  url       = {https://arxiv.org/abs/2511.05408}
}

Related

• steering-lite: hidden-state steering, sister project, source of all baseline rows above
• tinymfv: vignette dataset
• PEFT: adapter library
• RepE (Zou et al. 2023): hidden-state steering precursor