Multi-Agent Debate for Belief Optimization in Defeasible Domains
An arXiv paper finds multi-agent debate causes confidence escalation and martingale belief trajectories rather than better reasoning.
arXiv cs.LG — Machine Learning·28 Oct 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
An arXiv paper finds multi-agent debate causes confidence escalation and martingale belief trajectories rather than better reasoning.
Researchers introduced a benchmark exposing limitations in Vision-Language Model explainability caused by text biases and redundancy.
Research demonstrates network traffic side-channel attacks on cloud LLMs, inferring models and tasks from packet patterns despite encryption.
Researchers propose ReCoPII, a relation-grounded framework for generating contextual PII masking datasets using synthetic augmentation.
Researchers propose FinLLM-KG, combining knowledge graphs with LLMs for explainable financial sentiment analysis in algorithmic trading.
Researchers propose a low-rank multilinear decomposition method for detecting hallucinations by analyzing internal LLM hidden states.
Researchers introduce a framework to evaluate task handoffs and information preservation between collaborative AI agents.
Researchers propose a quantitative approach to measuring Asch conformity in frontier LLMs by repeatedly sampling a single question.
A preprint analysis examines whether token confidence scoring in reasoning language models merely reflects chain length.
Researchers propose a harness to reduce over-defensive writing in AI agents caused by optimizing against adversarial reviewer prompts.
Researchers introduced DocSkill, an agentic framework for multilingual multi-hop document question answering across complex layouts.
Researchers evaluate LLM ability to predict user non-compliance by comparing verbal promises against decision records in human-agent chats.
Researchers study decision models for identifying knowledge-conflict types in RAG retrieval sources to improve generator accuracy.
Researchers introduce TrawMem, an agent memory framework framing LLM memory as task-oriented role-adaptive workspaces rather than static retrieval.
Researchers introduced DocXEval, a benchmark framework for evaluating explainability in multimodal document understanding models.
An audit of 51 multilingual benchmarks reveals wide but thin coverage and methodological gaps across 219 languages and 242 datasets.
A research paper evaluates decision models as faster, lower-cost judges for hallucination detection across six benchmark datasets.
Researchers introduced UCVE, a graph-structured retrieval framework designed to improve automated coverage and compliance validation.
A new arXiv study audits pretraining disclosure across 41 models from 17 families using Foundation Model Transparency Index indicators.
Researchers propose a metacognitive controller for diffusion language models to dynamically adapt token generation and compute usage.
Researchers propose GEPA-Curator, a method to filter noisy learning signals for self-improving multi-turn LLM agent workflows.
Researchers propose SCOPE, a multi-agent system for profiling news media outlets regarding bias, factuality, and credibility.
An arXiv paper examines self-distillation methods for open-weight models doing financial quantitative reasoning under data residency.
A new research paper proposes a control map for evaluating test-time language model updates beyond simple update counts.
Researchers propose a decision-theoretic framework to resolve knowledge conflicts between retrieval context and parametric memory in language models.
Researchers introduced PRISM-BENCH, a benchmark evaluating multimodal model reasoning groundedness, accuracy, and inference costs.
Researchers introduce QESChunker, a question-evidence-supervised chunking method optimizing RAG preprocessing indices.
Researchers propose batching reflective text optimization loops to run parallel proposals, cutting wall-clock execution time.
Researchers propose GUI-JEPA, a decoder-free mobile GUI agent policy using Joint-Embedding Predictive Architectures to improve speed.
Researchers introduced RubricAtlas to extract type-level rubrics from expert reports for evaluating deep research agents.