Debate Training Reduces Reward Hacking in RLAIF — AI agent app

Zachary Kenton, Lili Janzer, Rory Greig, Tian Huey Teh, Kirill Tyshchuk, Jonah Brown-Cohen, Harri Edwards, Senthooran Rajamanoharan, Noah Y. Siegel, Natasha Jaques, Rohin Shah/Debate Training Reduces Reward Hacking in RLAIFUnknown

We demonstrate that RL finetuning an LLM using debate, a two-player adversarial game between a generator and a critic adjudicated by a weaker LLM judge, reduces reward hacking compared to a reinforcement learning from AI feedback (RLAIF) baseline. Reward hacking is a central obstacle in RLAIF: as training progresses, the policy learns to exploit systematic errors in its AI judge, degrading task performance, a problem that worsens precisely when the judge is weaker than the policy, the setting most relevant to overseeing increasingly capable AI systems. We study mathematics tasks, where final-answer correctness is verifiable, allowing us to measure reward hacking dynamics. We train a Gemini~2.5 Flash-class policy with a frozen, weaker Gemini~2.5 Flash Lite judge, comparing a single-player RLAIF baseline against debate. While the baseline quickly hacks the judge, debate maintains judge performance throughout training, leading to a higher peak validation accuracy (45\% performance gap recovered) that persists through many RL steps. Additional experiments show that: 1) further weakening the judge leads to faster hacking, but this can be compensated by adding an additional debate round; 2) debate incentives override prompted misalignment; 3) RL using an LLM judge has a smaller train/validation reward gap than RL from verifiable rewards; 4) learning to critique to convince the judge using ground truth labels is possible but slow. Taken together, our results are a positive update on the feasibility of debate, while highlighting that balancing multi-agent training is critical: without player constraints, adversarial training risks defaulting to critic judge-hacking. We show that critique word limits (effective up to 150 words) successfully balance the game and avoid judge hacking, though this introduces a trade-off by restricting critic expressive clarity.

agent app

Stars0

Forks0

HF Downloads—30d

Last commit—

Refreshed19h ago

Project healthUnknownNo activity data.

Production readinessResearch / EarlyBest for exploration and prototyping.

Risk notesUnknown licenseVerify license before production use.

AgentHub Score

55 / 100

Composite score from 6 signals. How we score →

Active project

55Score

Growth

40C

Activity

30C

Documentation

70C+

Maturity

45C

Community

42C

Production

58C

GitHub stars · 9 days observed0 not enough history

snapshots

Repository activity · 9 days observedReal snapshots from pushed_at

inactivepushed

2026-08-202026-08-30

Practical assessment

Should you use it?

✓ Best for

Research and experimentation
Prototype development
Learning agentic patterns

◎ Strengths

Active community
Open source
Well-documented API

✕ Not ideal for

Untested at scale without validation
Teams without AI/ML expertise

⚠ Watch-outs

Review changelog before updating
Verify license for commercial use

Technical details

What's inside

Language—

License—

Sourcearxiv

Open source✗ No

Commercial use—

Docs—

Demo—

PaperarXiv ↗

AgentHub Score

Score 55/100

Below average

Alternatives

OccuReward: LLM-Guided Occupant-Centric Reward Shaping for Demographic Equity in Grid-Interactive Buildings

0 · agent app

Don't Overthink, Don't Underthink: Toward Adaptive Reasoning in Agentic AI

0 · agent app

Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning

0 · agent app

Rethinking Agent Security as a Networking Problem

0 · agent app

55 Compare all →

Recent activity

Latest commit ——

Indexed by AgentHub crawler19h ago

Monitor for new releasesongoing