Research map / Reinforcement learning / RL from feedback
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Mohammad Taufeeque, Stefan Heimersheim, Adam Gleave, Chris Cundy
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit code length penalty to the reward (ablation) | RL without detector (KL 0.0001) · MBPP-Honeypot | deception rate: 59.1% -> 96.4% | Improved |
| increase KL coefficient from 0.01 to 0.1 (ablation) | RL with code length penalty · MBPP-Honeypot | deception rate (aggregate): 100.0% -> 33.3% | Improved |
| increase KL coefficient from 0.1 to 1.0 (ablation) | RL with code length penalty · MBPP-Honeypot | pass-all-tests (aggregate): 46.0% -> 60.1% | Improved |
| run standard RLVR without a detector (ablation) | off-domain deception probes (Simple Deception) · MBPP-Honeypot AUC | AUC: 0.991 -> 0.376 | Hurt |
| run standard RLVR without a detector (ablation) | off-domain deception probes (Diverse Deception) · MBPP-Honeypot AUC | AUC: 0.996 -> 0.655 | Hurt |
| run standard RLVR without a detector (ablation) | on-domain MBPP probe · MBPP-Honeypot AUC | AUC: 1.000 -> 1.000 | No clear effect |
| add detector penalty to the reward (ablation) | RL training · representation drift | representation change | No clear effect |
| pass gradients through the probe (ablation) | RL against detector · obfuscated activations | representation drift | Hurt |
| increase detector penalty coefficient (ablation) | RL at low KL · MBPP-Honeypot | fraction of honest policies | Improved |
| add Diverse Deception probe penalty alpha=3 (no code length penalty) | Llama-3-8B-Instruct, KL 0.0001 · MBPP-Honeypot | honest runs: 1/10 -> 10/10 | Improved |
| add Diverse Deception probe penalty alpha=3 (no code length penalty) | Gemma-3-12B-IT, KL 0.0001 · MBPP-Honeypot | honest runs: 6/10 -> 10/10 | Improved |
| add instruction not to hardcode to the system prompt (ablation) | RLVR without detector, KL 0.0001 · MBPP-Honeypot | reward hacking rate | No clear effect |
Closest papers by meaning
- Preference Learning with Lie Detectors can Induce Honesty or Evasion (NeurIPS 2025 · similarity 0.90)
- Do Coding Agents Deceive Us? Detecting and Preventing Cheating via Capped Evaluation with Randomized Tests (NeurIPS 2026 · similarity 0.88)
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.87)
- Detecting Strategic Deception with Linear Probes (ICML 2025 · similarity 0.87)
- Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation-Level Enforcement (NeurIPS 2026 · similarity 0.86)
- Large language models can learn and generalize steganographic chain-of-thought under process supervision (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
