deliberation

This paper introduces InfoDelphi, a framework that uses information asymmetry (partitioning evidence into shared public and disjoint private subsets) to improve multi-agent LLM deliberation and forecasting. On the PolyGym benchmark, it outperforms single-agent and multi-agent baselines by 12-18% in Brier score and 4-8 percentage points in accuracy, demonstrating that diverse evidence is key to effective multi-agent reasoning.

This paper investigates multi-agent deliberation methods for legal reasoning tasks using LLMs, introducing two novel frameworks inspired by courtroom procedures. The experiments show that multi-agent systems achieve comparable overall performance to monolithic LLMs but produce distinct answers and can solve cases that baselines fail, highlighting the potential of multi-agent approaches for legal AI.

This paper dissociates difficulty registration from deliberation allocation in large reasoning models (LRMs) and humans, finding that LRMs spend more tokens on problems they get wrong while humans spend less time on failures, revealing opposite within-item patterns despite similar cross-item difficulty correlations.

This paper investigates whether reasoning models' thinking tokens genuinely improve safety alignment, finding that safety outcomes are predictable from early hidden representations and that deliberation is largely superficial, with current safety interventions causing over-refusal.

This paper models multi-agent LLM deliberation as a closed-loop dynamical system where each agent has a hidden internal belief (anchor) that continually pulls its opinion, and shows how this anchor can be recovered from deliberation data alone, explaining phenomena like opinions escaping the convex hull of initial beliefs.

Introduces Think-Before-Speak (TBS), an interval-based multi-agent simulation framework that separates agents' private internal evaluation from public utterance generation, enabling analysis of the pathway from internal states to public expression in social simulations.

The paper introduces the Belief Engine, an auditable belief-update layer for LLM agents that makes stance changes in multi-agent deliberation configurable and inspectable by treating belief as an evidential state with explicit update rules.

This paper introduces CIG (Conversational Information Gain), a framework for measuring how utterances advance collective understanding in deliberative dialogues by tracking evolving semantic memory and scoring utterances on novelty, relevance, and implication scope. The authors demonstrate that memory-derived dynamics correlate better with human-perceived dialogue quality than traditional heuristics and develop LLM-based predictors for information-focused conversation analysis.