Debate Training Reduces Reward Hacking in RLAIF

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.

Human-AI Complementarity: A Goal for Amplified Oversight

Human feedback is critical for aligning AI systems to human values. As AI capabilities improve and AI is used to tackle more challenging tasks, verifying quality and safety becomes increasingly challenging. This paper explores how we can leverage AI to improve the quality of human oversight. We focus on an important safety problem that is already challenging for humans: fact-verification of AI outputs. We find that combining AI ratings and human ratings based on AI rater confidence is better than relying on either alone. Giving humans an AI fact-verification assistant further improves their accuracy, but the type of assistance matters. Displaying AI explanation, confidence, and labels leads to over-reliance, but just showing search results and evidence fosters more appropriate trust. These results have implications for Amplified Oversight -- the challenge of combining humans and AI to supervise AI systems even as they surpass human expert performance.

An Approach to Technical AGI Safety and Security

Artificial General Intelligence (AGI) promises transformative benefits but also presents significant risks. We develop an approach to address the risk of harms consequential enough to significantly harm humanity. We identify four areas of risk: misuse, misalignment, mistakes, and structural risks. Of these, we focus on technical approaches to misuse and misalignment. For misuse, our strategy aims to prevent threat actors from accessing dangerous capabilities, by proactively identifying dangerous capabilities, and implementing robust security, access restrictions, monitoring, and model safety mitigations. To address misalignment, we outline two lines of defense. First, model-level mitigations such as amplified oversight and robust training can help to build an aligned model. Second, system-level security measures such as monitoring and access control can mitigate harm even if the model is misaligned. Techniques from interpretability, uncertainty estimation, and safer design patterns can enhance the effectiveness of these mitigations. Finally, we briefly outline how these ingredients could be combined to produce safety cases for AGI systems.

Gemini: A Family of Highly Capable Multimodal Models

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of Gemini models in cross-modal reasoning and language understanding will enable a wide variety of use cases and we discuss our approach toward deploying them responsibly to users.

Improving alignment of dialogue agents via targeted human judgements

We present Sparrow, an information-seeking dialogue agent trained to be more helpful, correct, and harmless compared to prompted language model baselines. We use reinforcement learning from human feedback to train our models with two new additions to help human raters judge agent behaviour. First, to make our agent more helpful and harmless, we break down the requirements for good dialogue into natural language rules the agent should follow, and ask raters about each rule separately. We demonstrate that this breakdown enables us to collect more targeted human judgements of agent behaviour and allows for more efficient rule-conditional reward models. Second, our agent provides evidence from sources supporting factual claims when collecting preference judgements over model statements. For factual questions, evidence provided by Sparrow supports the sampled response 78% of the time. Sparrow is preferred more often than baselines while being more resilient to adversarial probing by humans, violating our rules only 8% of the time when probed. Finally, we conduct extensive analyses showing that though our model learns to follow our rules it can exhibit distributional biases.