pending reviewsae combinationReward HackingApache-2.0

RewardHackGuard PoC

Detect emergent reward-hacking generalization · Anthropic Nov 2025 framing

SAE-feature-combination probe on the residual stream of Qwen3.6-27B. Trained to detect activation patterns associated with reward-hacking generalization (Anthropic 2511.18397). Key for any team running GRPO / DPO / RLHF post-training.

byCaio VicentinoOpenInterp2026-05-XXarXiv:2603.04069

ProbeScore

0.645

Reward Hacking #1Global #4

7 weighted axes. Subject to revision as more reproducers land.

Quickstart

Three lines via the openinterp.probebench SDK. Returns the probe's P(positive_class) for a tensor of activations captured at the layer/position below.

python

from openinterp import probebench
probe = probebench.load("openinterp/rewardhackguard-qwen35-4b-l18")
score = probe.score(activations)  # → P(positive_class)

ProbeScore — 7 weighted axes

Composite metric in [0, 1]. No single axis can dominate; we think this discourages single-metric optimization. Numbers update as new evaluations land.

scoring formula

ProbeScore breakdown0.645

AUROC0.163

Eval-aware0.106

Dist-shift0.062

Calibration0.070

Transfer0.050

Goodhart-resistance0.050

Latency0.094

License0.050

ProbeScore — 8-axis breakdown

ProbeBench v0.0.2 · 8 axes weighted to sum 1.0

Visual breakdown of where this probe scores high vs low across all 8 axes. Polygon area is proportional to total score; shape reveals which axes carry it (and which ones it sacrifices).

AUROC

0.650

× 0.25 = 0.163

Eval-aware

0.590

× 0.18 = 0.106

Dist-shift

0.520

× 0.12 = 0.062

Calibration

0.700

× 0.10 = 0.070

Cross-model

0.500

× 0.10 = 0.050

Goodhart-resistance

0.500

× 0.10 = 0.050

Latency

0.936

× 0.10 = 0.094

License

1.000

× 0.05 = 0.050

Per-task evaluation

Headline AUROC alongside the eval-awareness-corrected and distribution-shift versions for each task. Reproducers run from a single notebook on a single Colab session.

Task	n	AUROC [95% CI]	Eval-aware corrected	Dist-shift	ECE	FPR@99TPR	Latency	Reproducer
HaluEval-QA Hallucination · hallucinated	200	0.650 [0.56, 0.74]	0.590	0.520	0.150	0.320	1.8 ms	—

Cross-model transfer (Pearson_CE)

Pearson correlation of paired ablation effects across models. ≥ 0.7 suggests the probe's direction is shared; 0.4–0.7 suggests partial transfer; < 0.4 suggests retraining per architecture.

No cross-model transfer measurements yet for this probe. Pearson_CE tells you whether a probe's direction transfers to another architecture without retraining. Submit a transfer measurement.

Mean Pearson_CE for this probe (excluding self-baseline): 0.500

Architecture & artifact

Base model

Qwen3.6-27B

Qwen/Qwen3.6-27B

Family: Qwen
Params: 27B
Architecture: Hybrid GDN + Gated-Attn (dense, reasoning)
Layers: 64
d_model: 5120
Weights license: Apache-2.0

Probe attaches at

Layer 31 · token_avg

huggingface.co/Qwen/Qwen3.6-27B

Artifact

RewardHackGuard weights

sklearn-compatible probe + scaler + meta.json

Params: 480,000
Size: 1.8 MB
License: Apache-2.0
Released: 2026-05-XX
sha256: not published

huggingface.co/datasets/caiovicentino1/RewardHackGuard-linearprobe-qwen35-4b source notebook on GitHub open reproducer in Colab

Hash matches the artifact at huggingface.co/datasets/caiovicentino1/RewardHackGuard-linearprobe-qwen35-4b. Recompute via openinterp probebench verify openinterp/rewardhackguard-qwen35-4b-l18.

Reproduce

Three entry points. Pick one. Each path lands at the same numbers.

Open in Colab

One-click. Free T4 works for inference; reproducer needs an A100 or RTX 6000.

colab.research.google.com

View on GitHub

Read the notebook source, see commit history, file an issue.

github.com · OpenInterpretability

Download from HF

Probe weights, scaler, and meta.json. Verified by the sha256 above.

huggingface.co · datasets

shell

git clone https://github.com/OpenInterpretability/notebooks.git
pip install openinterp
openinterp probebench reproduce openinterp/rewardhackguard-qwen35-4b-l18

Honest scope

Limits derived from the evaluation data above. We think these are the honest constraints; revisions land as more reproducers do.

Trained / evaluated on: HaluEval-QA. Performance outside these tasks is unmeasured here.
Eval-aware AUROC drop: −0.060 AUROC averaged across tasks (uncorrected − corrected for eval-awareness confound, arXiv:2509.13333).
Distribution-shift drop: −0.130 AUROC averaged across tasks (in-distribution − long-context / OOD).
Cross-model fit (mean Pearson_CE): 0.500 — values below 0.4 suggest retraining per architecture; values above 0.7 suggest the direction is shared.
Probe attaches at: L31 · token_avg of Qwen3.6-27B. Other layers / positions are out of scope unless re-trained.

Back to Reward Hacking leaderboard Submit your own probe

ProbeBench v0.0.1 · 5 probes registered · 11 evaluations · 7 transfer measurements. Schema and weights subject to revision.

Quickstart

Three lines via the openinterp.probebench SDK. Returns the probe's P(positive_class) for a tensor of activations captured at the layer/position below.

python

from openinterp import probebench
probe = probebench.load("openinterp/rewardhackguard-qwen35-4b-l18")
score = probe.score(activations)  # → P(positive_class)

Per-task evaluation

Headline AUROC alongside the eval-awareness-corrected and distribution-shift versions for each task. Reproducers run from a single notebook on a single Colab session.

Task	n	AUROC [95% CI]	Eval-aware corrected	Dist-shift	ECE	FPR@99TPR	Latency	Reproducer
HaluEval-QA Hallucination · hallucinated	200	0.650 [0.56, 0.74]	0.590	0.520	0.150	0.320	1.8 ms	—

Cross-model transfer (Pearson_CE)

Pearson correlation of paired ablation effects across models. ≥ 0.7 suggests the probe's direction is shared; 0.4–0.7 suggests partial transfer; < 0.4 suggests retraining per architecture.

No cross-model transfer measurements yet for this probe. Pearson_CE tells you whether a probe's direction transfers to another architecture without retraining. Submit a transfer measurement.

Mean Pearson_CE for this probe (excluding self-baseline): 0.500

Architecture & artifact

Base model

Qwen3.6-27B

Qwen/Qwen3.6-27B

Family: Qwen
Params: 27B
Architecture: Hybrid GDN + Gated-Attn (dense, reasoning)
Layers: 64
d_model: 5120
Weights license: Apache-2.0

Probe attaches at

Layer 31 · token_avg

huggingface.co/Qwen/Qwen3.6-27B

Artifact

RewardHackGuard weights

sklearn-compatible probe + scaler + meta.json

Params: 480,000
Size: 1.8 MB
License: Apache-2.0
Released: 2026-05-XX
sha256: not published

huggingface.co/datasets/caiovicentino1/RewardHackGuard-linearprobe-qwen35-4b source notebook on GitHub open reproducer in Colab

Hash matches the artifact at huggingface.co/datasets/caiovicentino1/RewardHackGuard-linearprobe-qwen35-4b. Recompute via openinterp probebench verify openinterp/rewardhackguard-qwen35-4b-l18.

Honest scope

Limits derived from the evaluation data above. We think these are the honest constraints; revisions land as more reproducers do.

Trained / evaluated on: HaluEval-QA. Performance outside these tasks is unmeasured here.

Eval-aware AUROC drop: −0.060 AUROC averaged across tasks (uncorrected − corrected for eval-awareness confound, arXiv:2509.13333).

Distribution-shift drop: −0.130 AUROC averaged across tasks (in-distribution − long-context / OOD).

Cross-model fit (mean Pearson_CE): 0.500 — values below 0.4 suggest retraining per architecture; values above 0.7 suggest the direction is shared.

Probe attaches at: L31 · token_avg of Qwen3.6-27B. Other layers / positions are out of scope unless re-trained.