Co-lead with @yuweiyin.bsky.social

Huge thanks to @veredshwartz.bsky.social, Peter West, Giuseppe Carenini
Paper: huggingface.co/papers/2509....
Code will be released soon!

Our findings highlight that:
👉 Social reasoning in LLMs cannot be achieved through optimizing their performance on general reasoning benchmarks alone!
👉It requires explicit modeling of mental states to enable safe, fair, and effective interactions with humans.

We also examine mental states (beliefs, desires, intentions, emotions, knowledge).

🔹 ToMA prioritizes intentions > emotions (other dimensions remain similar)
🔹 Uses +5.6% more 1st-order belief than bases, even when both are prompted equally for 0th/1st order states.

We analyze 4 scenario types: cooperation, negotiation, persuasion, and conflict.

ToMA outperforms the base under all settings. Its reasoning is more strategic (e.g., compromise, accommodation). Even in failures, ToMA shows more active engagement (e.g., failed persuasion).

ToMA adapts effectively to long conversations, sustaining strategic dialogue. When paired with diverse partners, it improves both its own goal completion and its partners’ success.

ToMA generates latent mental states and utterances optimized for social interaction goals using dialogue simulation signals. On Sotopia, it improves performance by +18.9% with Qwen2.5-3B and +6.9% with Qwen2.5-7B, while remaining competitive with a GPT-5 nano baseline.

👋

Also, consider presenting a poster showcasing any ongoing projects or previously presented works from recent conferences. It will be a great chance to get feedback and promote your work!