CounterSteer/sycophancy-artifacts

CounterSteer steering-vector artifacts — sycophancy

Curated steering directions and frozen configurations from the diversified-probe sycophancy experiment. One probe is trained on all three sycophancy forms (answer, pushback, feedback), 512 rows split 171 / 171 / 170. It excludes model weights, credentials, generated responses, judge caches and superseded runs.

Sources: CAA and CounterSteer-LoRA from code/experiments/sycophancy_diversified_probe_data (results table in that folder's reports/GATHERED_RESULTS.md); CounterSteer-GRAD and LAT-PCA from code/experiments/training_free_nas_sycophancy, whose LAT vectors were extracted by code/experiments/more_baselines/RepE/extract_sycophancy_directions.py. Each arm's audit.json names the exact run folder and direction file it came from.

Included artifacts

Base model	Method	Artifact	Frozen configuration
Qwen/Qwen2.5-3B-Instruct	CAA	Layers 18-35	layer 33, alpha 64
Qwen/Qwen2.5-3B-Instruct	CounterSteer-LoRA	Layers 18-35	layer 28, alpha 32
Qwen/Qwen2.5-3B-Instruct	CounterSteer-GRAD	Layers 18-35	layer 20, alpha 4
Qwen/Qwen2.5-3B-Instruct	LAT-PCA	Layers 18-35	layer 33, alpha 32
Qwen/Qwen2.5-7B-Instruct	CAA	Layers 14-27	layer 17, alpha 16
Qwen/Qwen2.5-7B-Instruct	CounterSteer-LoRA	Layers 14-27	layer 20, alpha 32
Qwen/Qwen2.5-7B-Instruct	CounterSteer-GRAD	Layers 14-27	layer 14, alpha 4
Qwen/Qwen2.5-7B-Instruct	LAT-PCA	Layers 14-27	layer 17, alpha 16
meta-llama/Llama-3.1-8B-Instruct	CAA	Layers 16-31	layer 25, alpha 1
meta-llama/Llama-3.1-8B-Instruct	CounterSteer-LoRA	Layers 16-31	layer 16, alpha 1
meta-llama/Llama-3.1-8B-Instruct	CounterSteer-GRAD	Layers 16-31	layer 16, alpha 1
meta-llama/Llama-3.1-8B-Instruct	LAT-PCA	Layers 16-31	layer 26, alpha 2
google/gemma-2-9b-it	CAA	Layers 21-41	layer 21, alpha 128
google/gemma-2-9b-it	CounterSteer-LoRA	Layers 21-41	layer 21, alpha 128
google/gemma-2-9b-it	CounterSteer-GRAD	—	not run for this model
google/gemma-2-9b-it	LAT-PCA	—	not run for this model
Qwen/Qwen2.5-3B-Instruct	Random (control)	Layers 18-35	layer 28, alpha 32
Qwen/Qwen2.5-7B-Instruct	Random (control)	Layers 14-27	layer 20, alpha 32
meta-llama/Llama-3.1-8B-Instruct	Random (control)	Layers 16-31	layer 16, alpha 1
google/gemma-2-9b-it	Random (control)	Layers 21-41	layer 21, alpha 128

Every model folder carries the same five arm directories, four methods plus the random/ control. A folder holding only a README means that arm was not run for that model: gemma-2-9b-it has no CounterSteer-GRAD and no LAT-PCA — the sycophancy experiment was never run for it on either.

CounterSteer-GRAD convention

v_grad = -mean_j grad_u L_j with H~ = H + u, per-example gradient weighting, summed over token positions without a length divisor, then normalised. Both weightings were swept; per-example is the reported arm on every model.

Llama-3.1-8B's frozen configuration is alpha=1, not the objective default alpha=2. The default regressed the answer form (agreement 15.0 -> 16.2) while winning on pushback; the report selects alpha=1 as the configuration competitive on all three forms.

LAT-PCA convention

PC1 of the recentred within-pair differences with each pair's order shuffled — RepE's own recipe — sign fixed by RepE's get_signs, read out at the same position as CAA.

Llama-3.1-8B's selection comes from a declared sensitivity variant of the feasibility rule (degenerate allowance max(1 pp, 50% of baseline) = 1.89 pp). Under the frozen +1 pp rule that arm had no feasible configuration at all — 0 of 40 coarse jobs. lat-pca/metadata.json records this in selection_note; the variant is reported alongside the frozen result, never instead of it.

All layer indices are zero-based decoder-block indices. Every frozen configuration was selected on validation only, over the full latter-half layer range at the strength grid {1, 2, 4, 8, 16, 32, 64, 128}; each published bundle contains every layer that was searched, not only the selected one.

Random control convention

random/direction.pt is a standard Gaussian per layer, drawn once with seed 42, in the same raw/unit/norm schema as the methods. It is a control, not a method, and is not ranked in the results tables.

It is not searched. It inherits the layer and alpha that CounterSteer-LoRA selected on validation, so the two arms differ only in the direction steered along — which is what makes it a control for that arm, and why its frozen_selection.json records an inherited configuration rather than a selected one. The source runs hold a second control at the CAA configuration (random_at_caa) sharing the same seeded file; only the CounterSteer-LoRA-matched arm is published here.

Inference rule

Every arm here uses the same rule, with no hidden-state-norm scaling:

h' = h - alpha * unit(v)

applied at one decoder block, to all prompt and generated token positions. The direction points toward the sycophantic behaviour and inference subtracts it.

How the directions are built

Arm	Construction	Readout
CAA	mean(h[sycophantic reply] - h[contrast reply]) over same-prompt pairs	final token of the templated prompt plus reply
CounterSteer-LoRA	mean(adapter_on - adapter_off), the shift a short LoRA fine-tune toward sycophancy induces	prompt-final block output
CounterSteer-GRAD	-mean_j grad_u L_j with H~ = H + u, per-example weighting, position sum with no length divisor	gradient w.r.t. an additive bias at the block output
LAT-PCA	PC1 of the recentred within-pair differences, pair order shuffled, sign from RepE's get_signs	final token of the templated prompt plus reply
Random (control)	standard Gaussian per layer, seeded (seed 42); carries no behavioural orientation	not applicable

Provenance and what was verified

Unlike the refusal release, the sycophancy pipeline did not record a SHA-256 of the direction file in its test runs, so byte identity cannot be established from a stored digest. Each published bundle was instead tied to its held-out test by recomputation and cross-reference:

• per-layer ||raw|| recomputed from the published file matches the audit written at extraction time;
• unit equals raw / ||raw|| at every layer;
• the per-layer CounterSteer-LoRA-versus-CAA cosines recomputed from the published files match cosine_nas_caa.json from the same extraction (CAA and CounterSteer-LoRA only);
• the selected layer and alpha match the frozen validation selection, the test plan, the test arm's own record, and the job recorded in every generation manifest under that arm, including both pushback turns.

For CounterSteer-GRAD and LAT-PCA, added later, the same checks were run plus a direct one the earlier arms could not use: each published tensor is bit-identical at every layer to the direction file its test run read, so no digest is needed. Concretely, for all six of those bundles

• direction.pt equals the source directions_per_example/grad.pt or <model>_lat/directions/lat_shuffled.pt tensor for tensor across the whole swept range;
• the frozen selection, jobs_test.json, and the job recorded in all six generation manifests (answer, feedback, and both pushback turns) agree on layer, alpha and direction filename;
• the CounterSteer-GRAD source is confirmed to be the per-example weighting by its own grad_direction_report.json (weighting: per_example), which is a materially different vector from the mean-weighted one — they differ at cosine 0.45 to 0.57 at the first swept layer.

A sha256 for each file as published is recorded in <arm>/metadata.json and in manifest.json.

Qwen2.5-7B ships the epoch-5 probe, which is the reported arm. Gate A selected epoch 8; the two tie within noise and epoch 5 was chosen for the headline. The epoch-8 run remains in the source tree at runs/qwen2.5-7b, and its CAA file is byte-identical to the one published here. See reports/qwen2.5-7b/epoch5_epoch8_comparison.md.

Loading

import torch

bundle = torch.load(
    "qwen2.5-7b-instruct/cs-lora/direction.pt",
    map_location="cpu",
    weights_only=True,
)
direction = bundle["unit"][20]   # selected layer; alpha 32

Run python validate.py after downloading. It verifies every published checksum and, when PyTorch is installed, checks tensor keys, shapes, norms, and that each selected layer is present in its bundle.

File guide

• config.json: the experiment configuration, including pinned model revision and decoding settings.
• <arm>/direction.pt: per-layer bundle with raw, unit and norm dictionaries.
• <arm>/metadata.json: public metadata — model, layers, construction, inference rule, frozen selection, search budget and verification references.
• <arm>/audit.json: extraction-time record — layer list, per-layer norms, construction counts.
• <arm>/frozen_selection.json: the validation-selected operating point, the search plan and the unsteered baseline it was compared against. For random/ it records the inherited CounterSteer-LoRA configuration instead, since the control is never searched.
• <arm>/README.md, where an arm folder holds only that file: states why the arm has no vector. Present for gemma-2-9b-it's cs-grad/ and lat-pca/ only.
• manifest.json: byte size and SHA-256 for every release file except the manifest itself.

The base models remain subject to their respective upstream access and license terms.