mlsanityData to 7 Oct 2026

RH-Detect: A Unified Benchmark for Reward Hacking Detection

arXiv 2610.10947 ↗First posted 7 Oct 2026cs.LGcited by 0

On the map: LLM alignment / RLHF / preference optimisation → Reward hacking and misspecification

Abstract (arXiv)

Reward hacking, where a model exploits an evaluation signal without completing the intended task, threatens the reliability of deployed language model systems. Existing datasets use different labels, response formats, and metadata conventions, making detector results difficult to compare. We present RH-Detect, a benchmark that combines reward-hacking-relevant subsets from eleven public datasets, comprising 92,761 rows and six behavior categories, into a common schema. On 5,021 open-ended evaluation units, each comprising a task prompt and a free-form model continuation, including multi-turn tool-use trajectories, we evaluate six off-the-shelf language models from five families as reward hacking detectors without additional training. The best model achieves a pooled AUROC of 0.962, with accuracy above 93%. For the four strongest models, however, accuracy on the two multi-turn tool-use datasets, MALT and TRACE, is 10.7-15.9 percentage points lower than on the other sources at a common decision threshold, highlighting a key gap for deployment-time monitoring. We find that different input formats have different effects across models. Removing thinking raises Qwen3.5-4B AUROC from 0.779 to 0.849, but lowers Qwen Flash from 0.977 to 0.950. We also evaluate the benchmark as a training dataset for detectors. Holding out each source in turn, single-token SFT improves average AUROC on five of six held-out sources. A GRPO follow-up on that failure case yields a slight improvement in detection performance. Our results show that a single pooled score can conceal variation across data sources, detector inputs, and training procedures.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reward hacking and misspecification, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 19 of its 29 references that are in the map. First posted on arXiv 7 Oct 2026.

Builds on

19 of 29 references.

Closest ancestors · 8

Same direction, newest first

  1. 2026Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use3 May
  2. 2026Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR7 Mar
  3. 2026Monitoring Emergent Reward Hacking During Generation via Internal Activations4 Mar
  4. 2026Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis27 Jan
  5. 2025EvilGenie: A Reward Hacking Benchmark26 Nov
  6. 2025Natural Emergent Misalignment from Reward Hacking in Production RL23 Nov
  7. 2025School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs24 Aug
  8. 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation14 Mar

Other references · 11

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  2. 2023Towards Understanding Sycophancy in Language Modelscited by 988
  3. 2024Alignment faking in large language modelscited by 316
  4. 2024Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Modelscited by 158
  5. 2025ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Casescited by 51
  6. 2026Terminal Wrench: A Dataset of 331 Reward-Hackable Environments and 3,632 Exploit Trajectoriescited by 14
  7. 2026How does information access affect LLM monitors' ability to detect sabotage?cited by 12
  8. 2026AutoMonitor-Bench: Evaluating the Reliability of LLM-Based Misbehavior Monitorcited by 8
  9. 2026MonitoringBench: Semi-Automated Red-Teaming for Agent Monitoringcited by 4
  10. 2026SLEIGHT-Bench: A Benchmark of Evasion Attacks Against Agent Monitorscited by 4
  11. 2026Training Deliberative Monitors for Black-Box Scheming Detectioncited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL
  2. 2026HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models
  3. 2026Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
  4. 2026Harness-agnostic detection and immunization of reward hacking in self-evolving language models
  5. 2026Multimodal Reward Hacking in Reinforcement Learning
  6. 2024RewardBench: Evaluating Reward Models for Language Modeling
  7. 2026Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks
  8. 2024RRM: Robust Reward Model Training Mitigates Reward Hacking