5Berkeley AI Research (BAIR) Blog·
BAIR introduces ABBEL: teaching LLMs to update beliefs for efficient long-horizon interaction
Researchers from UC Berkeley's BAIR Lab present ABBEL, a method for training LLMs to perform principled belief updating during long-horizon interactive tasks, reducing the number of clarifying questions needed while maintaining task accuracy. The approach targets a core inefficiency in current LLM-based agents: failure to maintain and revise a coherent model of user intent across multi-turn interactions. The work is positioned as a step toward more efficient human-AI collaboration in agentic settings.
Related guides (2)
Related events (8)
BeliefTrack: Benchmarking and Improving Contextual Belief Management in LLMs
This paper introduces Contextual Belief Management (CBM) as a framework for studying how LLMs should update, preserve, or ignore information across long-horizon interactions. The authors release BeliefTrack, a closed-world benchmark with symbolic verifiers enabling exact turn-level evaluation across Rule Discovery and Circuit Diagnosis tasks. Vanilla LLMs show severe CBM failures; reinforcement learning with belief-state rewards reduces failure rates by 70.9% on average, while representation-level steering achieves 46.1% reduction. Probing experiments reveal latent belief-state dynamics underlying these failures.
LLMs can extract and reuse experiential abstractions to improve reasoning, study finds
Researchers investigate whether LLMs can distill solution traces into reusable natural-language abstractions (strategies, reminders) and apply them to future problems, analogous to human experiential learning. Using the MATH training set, abstractions are extracted by either a stronger teacher model or the LLMs themselves into a retrievable library, then applied via inference-time retrieval or RL with abstraction-augmented prompts. Both modes improve performance on mathematical and logical reasoning benchmarks, with self-extracted abstractions matching teacher-extracted ones. The framework generalizes across datasets and models, suggesting a promising direction for continual self-improvement in LLMs.
Benchmark and study of LLM agents in deliberative multi-agent collaboration under partial observability
Researchers introduce a scalable benchmark for evaluating LLM agents on cooperative joint decision-making tasks where agents must exchange information under partial and asymmetric observations to reach a shared decision. A systematic evaluation of representative LLMs finds that state-of-the-art models still struggle with complex deliberative collaboration, failing in either information alignment or downstream reasoning even with external mathematical tools. Diagnostic analysis also reveals that deliberation can enable reflection and error correction, sometimes outperforming centralized baselines, offering a nuanced picture of multi-agent LLM capabilities.
LLawCo framework teaches embodied multi-agent LLMs to derive and follow cooperation laws
Researchers from MERL propose LLawCo (Learning Laws of Cooperation), a framework that enables embodied LLM-based agents to autonomously align with partners and task objectives in decentralized, partially observable environments. Agents reflect on past failures to extract misaligned behavioral patterns and derive high-level behavioral laws (e.g., 'Talk when necessary', 'Wait for partner'), which are incorporated into reasoning via supervised fine-tuning. The authors also introduce PARTNR-Dialog, a new large-scale multi-agent communicative planning benchmark, and report average success rate improvements of 4.5% on PARTNR-Dialog and 6.8% on TDW-MAT over state-of-the-art open-source communicative agent frameworks across four backbone LLMs.
AIR: Adaptive Interleaved Reasoning with Code in Multimodal LLMs via Reinforcement Learning
Researchers propose AIR, a system that trains multimodal large language models to adaptively interleave reasoning with code execution for numerical computation tasks, going beyond prior work that focused only on visual operations. The approach combines a two-stage cold-start data pipeline, RL dataset filtering, and a group-constrained reward function for tool-invocation decisions. Experiments show a 6.1 percentage point average improvement on evaluation benchmarks, with interleaved reasoning samples gaining 9.9 pp and tool-use success exceeding 95%.
Benchmark study reveals how linguistic framing of user beliefs shifts LLM context-following behavior
A new arXiv paper introduces a typology of 17 linguistically motivated expression-of-belief (EoB) types—spanning form, evidentiality, epistemic stance, and tone—to evaluate how phrasing affects whether LLMs defer to user-stated beliefs or their own prior knowledge. The authors benchmark 16 LLMs across Llama 3, Qwen3, and Gemma3 families at scales from 1B to 30B parameters, finding that larger and instruction-tuned models are systematically less context-following than smaller or base models. Specific linguistic framings (e.g., presuppositions, certainty markers) are identified as statistically more persuasive, with implications for prompt robustness and sycophancy research.
Reinforcement Learning with Metacognitive Feedback (RLMF) improves LLM calibration and uncertainty expression
Researchers introduce Reinforcement Learning with Metacognitive Feedback (RLMF), a training paradigm that refines preference optimization using a model's self-judgments of its own performance quality. The method is applied to faithful calibration — aligning a model's expressed confidence with its intrinsic uncertainty — and achieves state-of-the-art results across diverse tasks while outperforming standard RL by up to 63%. A companion technique, metacognitive data selection, uses similar self-judgments to identify high-value training examples, outperforming naive active learning baselines. The work positions metacognitive performance as a novel and effective RL signal for improving LLM reliability and alignment.
Int-Bench reveals LLM tutors over-intervene and prioritize short-term task success over learning
Researchers introduce Int-Bench, a simulation-based benchmark evaluating how LLMs intervene during student problem-solving across code debugging, mathematics, and brain teasers. The benchmark compares LLM 'teachers' to human tutors, finding that LLMs intervene more frequently, earlier, and tend to provide complete solutions rather than targeted hints. The study concludes that current LLM assistants optimize for immediate task success rather than fostering deeper reasoning and long-term learning, a meaningful finding for AI tutoring system design.