Future work should focus on developing smarter debate protocols that weight expertise, discourage blind agreement, and reward critical verification of reasoning. We need to move beyond the naive assumption that 'more talk = better outcomes. (10/10) arxiv.org/abs/2509.05396

We suspect RLHF training creates sycophantic behavior, models trained to be agreeable may prioritize consensus over critical evaluation. This suggests current alignment techniques might undermine collaborative reasoning.

Stronger agents were more likely to change from correct to incorrect answers in response to weaker agents' reasoning than vice versa. Models showed a tendency toward favoring agreement over critical evaluation, creating an echo chamber instead of an actual debate.

However, we still observed performance gains on math problems under most conditions, suggesting debate effectiveness depends heavily on the type of reasoning required.

The impact varies significantly by task type. On CommonSenseQA—a dataset we newly examined—debate reduced performance across ALL experimental conditions.

Even when stronger models outweighed weaker ones, group accuracy decreased over successive debate rounds. Introducing weaker models into debates produced results that were worse than the results when agents hadn’t engaged in discussion at all.

We tested debate effectiveness across three tasks (CommonSenseQA, MMLU, GSM8K) using three different models (GPT-4o-mini, LLaMA-3.1-8B, Mistral-7B) in various configurations.

We found that multi-agent debate among large language models can sometimes harm performance rather than improve it, contradicting the assumption that more discussion can lead to better outcomes.

My lab members Harsh Satija and Andrea Wynn and I have a new preprint examining AI multi-agent debate among diverse models, based on our ICML MAS 2025 workshop.

These roles will shape the conversation on AI and provide the opportunity for rich, interdisciplinary collaboration with colleagues and researchers in the Department of Computer Science and the School of Government and Policy.
Please spread the word in your network! 5/5
gillianhadfield.org/jobs/

We're recruiting for a Postdoctoral fellow with a track record in computational modeling that explores AI systems and autonomous AI agent dynamics, and experience with ML systems to investigate the foundations of human normativity, and how to build AI systems aligned with human values. 4/5

We're hiring an AI Communications Associate to craft and execute a multi-channel strategy that turns leading computer science and public policy research into accessible content for a broad audience of stakeholders. 3/5

We're hiring an AI Policy Researcher to conduct in-depth research into the technical and policy challenges in AI alignment, safety, and governance, and to produce high-quality research reports, white papers, and policy recommendations. 2/5

Our report is now out, chock-a-block with new ideas including insurance partnerships, government oversight of private regulators, building a robust ecosystem, and fostering trust and investment. Check it out here: srinstitute.utoronto.ca/news/co-desi...

destabilize or harm our communities, economies, or politics.Together with @djjrjr.bsky.social and @torontosri.bsky.social we held a design workshop last year with a stunning group of experts from AI labs, regulatory technology startups, enterprise clients, civil society, academia,and government.2/3