Belief Updating in LLMs

LLM: Large Language Model

Belief updating in large language models concerns how models form and revise views as new evidence arrives, facts change, or other agents make claims. Researchers test this with controlled multi-turn benchmarks, Bayesian and human comparisons, social-persuasion tasks, and probes of internal representations; agent methods also use explicit belief states, retrieval, fine-tuning, and steering. Models can track some evidence well, yet often show primacy or recency effects, rely on stale context, or shift under social pressure, with results depending on the model and evaluation setup. A central challenge is making models resist unsupported influence without blocking legitimate evidence-based revision, while determining whether changed answers reflect changed internal beliefs or surface compliance.

Latest papers

97 papers40 papers in the last four weeks, up 82% on the four weeks before. 0.2% of all new papers.

Surprise Alone Learns Lies: Evidence-Gated Continual Learning for LLMs from Streaming Claims
Oct 29, 2026Abdelkader Metales, Miriam GrabenheinrichContinual Learning for LLMs LLM Reliability
caaise GmbH
Auditable Belief Updates as a Missing Safety Layer for LLM-Powered Multi-Agent Systems
Oct 28, 2026Zhuoyun Li, Boxuan Wang, Jinwei Hu +2AI Agent Safety LLM Agent Safety
School of Computer Science and Informatics, University of Liverpool, UK
Not One Belief: Answer Payloads, Persistence, and Dissociated Confidence Signals in Conversational Question Answering
Oct 26, 2026Duc-Vinh TranConversational QA Belief Updating in LLMs
School of Information and Communication Technology, Hanoi University of Science and Technology, Hanoi, 100000, Vietnam.
Can LLMs Trust Appropriately? Evaluating Calibrated Reliance on Imperfect Partners
Oct 25, 2026Mengdi Chu, Jintao HuangTrust in AI LLM Reliability
The Ohio State University.
The Patient Is Always Right? Evaluating Sycophancy in Medical Diagnosis with Graded Patient Beliefs
Oct 25, 2026Viswonathan Manoranjan, Andrea Mock, Myra Cheng +4LLM Sycophancy Clinical Language Model Evaluation
University of North Carolina at Chapel Hill · Vals AI · Stanford University +2
Network Topology Does Not Matter: The Breakdown of Classical Consensus in LLM Populations
Oct 24, 2026Sahil Rajesh DhayalkarMulti-Agent Consensus Belief Updating in LLMs
Arizona State University
Do "True" and "False" Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations
Oct 23, 2026Anouar Imel, Jin QiBehavioral Consistency of LLM Agents Large Language Model-Based Social Simulation
School of Information and Communication Engineering University of Electronic Science and Technology of China (UESTC) Chengdu, China
How Much Do You Trust A Dog? Quantifying Hint-Following in Reasoning Language Models
Oct 23, 2026Anxin Guo, Chirag Pabbaraju, Sara Ahmadian +2LLM Sycophancy LLM Reliability
Northwestern University, Evanston, IL, USA · Northwestern University, Evanston, IL, USA; NSF-Simons AI Institute for the Sky (SkAI), Chicago, IL, USA · Stanford University, Stanford, CA, USA +2
Invisibilization as a Source of Probability Bias in Language Models
Oct 22, 2026Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez +2Demographic Bias in Language Models Language Model Bias Evaluation
University of Birmingham · University of Maryland, College Park · University of Western Ontario +1
Large language models show systematic failures in incorporating evolving medical knowledge
Oct 22, 2026Min Zeng, Zaifu Zhan, Kai Yu +1Medical LLMs LLM Reliability
Division of Computational Health Sciences, Department of Surgery, University of Minnesota, Minneapolis, 55455, MN, USA.
From Prompts to Posteriors: Can LLM Agents Perform Bayesian Inference in Context?
Oct 21, 2026Linus Filbry, Luckeciano C. Melo, Yarin GalBelief Updating in LLMs Bayesian Inference
OATML, University of Oxford
To Manipulate or Not: Opinion Dynamics in LLM Interactions
Oct 20, 2026Mohammad Siahkamari, Phuoc Dat Nguyen, Stephen MacNeilPersuasion in Language Models Belief Updating in LLMs
Temple University, Philadelphia, PA, USA
Is Sycophancy a Form of In-Context Updating?
Oct 20, 2026Itai Shapira, Myra Cheng, Hadas OrgadLLM Sycophancy Belief Updating in LLMs
Harvard University · Stanford University · Kempner Institute at Harvard University
Hearsay or Fact? How Framing and Elicitation Shape LLM Probability Updates
Oct 20, 2026An-Sheng JhangProbability Elicitation from Language Models Belief Updating in LLMs
Dongguan University of Technology
LAMB: Latent Multi-Hypothesis Beliefs for Long-Horizon Large Language Agents
Oct 18, 2026Peien Ding, Hui Li, Yuhan Liu +1Long-Horizon LLM Agents Memory-Augmented Language Model Agents
School of Informatics, Xiamen University · MBZUAI
The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents
Oct 17, 2026Jinwei Hu, Youcheng Sun, Yi Dong +1Adversarial Attacks on LLMs LLM Agent Reliability
School of Computer Science and Informatics, University of Liverpool · Department of Computer Science, MBZUAI
Process Reward Modeling with Reliability- and Conservatism-Aware Bayesian Belief
Oct 16, 2026Zhongjian Qiao, Zhixiang Zhou, Jiafei Lyu +4MLLM Reasoning Process Reward Models
CityUHK · Shanghai Innovation Institute · Tencent +1
Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussions
Oct 15, 2026Liangjie Huang, Henry Peng Zou, Wei-Chieh Huang +3Multi-Agent LLM Systems LLM Agent Reliability
University of Illinois Chicago
From Hindsight to Foresight: Dissecting Clinical Evidence Acquisition in Large Language Models
Oct 13, 2026Ziyang LingMedical LLMs Clinical Language Model Evaluation
School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, China
Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made
Oct 13, 2026Ayoub Merimi, Souhail Bakkali, Abdeljalil SakatLLM Decision-Making Temporal Reasoning in Language Models
Upstream Influence, Downstream Re-Anchoring in LLM Agent Chains
Oct 13, 2026Chun-Yu Hsu, Lun-Wei Ku, Chung-Chi ChenMulti-Agent LLM Systems LLM Agent Reliability
Institute of Information Science, Academia Sinica, Taiwan · Human-Agent Ally Lab, National Institute of Informatics, Japan
Beyond Successful Play: Diagnosing Opponent Modeling in Large Language Models
Oct 12, 2026Hanbo Xie, Yue Zhao, Robert C. WilsonOpponent Modeling Theory of Mind
University of Sussex
Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate
Oct 12, 2026Harsh Kumar, George Eilender, Ashton AndersonSocial Bias in Language Models Belief Updating in LLMs
Department of Computer Science University of Toronto Toronto, ON, Canada
Big-Big-SycoBench: A Benchmark to Detect Sycophancy Through Model Responses to Valid vs. Invalid Arguments
Oct 11, 2026Marthe Ballon, Arne Vanhoyweghen, Seorin Kim +11LLM Sycophancy LLM Evaluation
Data Analytics Lab, Vrije Universiteit Brussel, Pleinlaan 5, 1050 Brussel, Belgium · imec-SMIT, Vrije Universiteit Brussel, Pleinlaan 9, 1050 Brussels, Belgium · School of Engineering and Applied Sciences, Harvard University, Cambridge, Massachusetts 02138, USA
Provenance-Aware Belief Updating in Human-AI Feedback Loops
Oct 9, 2026Tianyi Huang, Eric Liu, Jay Chong +2Data Provenance in Language Models Belief Updating in LLMs
Ryquo · App-In Club
EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
Oct 8, 2026Janet Wang, Jiang Guo, Yi Xiang +6World Models Belief Updating in LLMs
Tulane University · Amazon
BeliefScope: Diagnosing Evidence-Driven Revision and Pressure-Induced Shifts in Large Language Models
Oct 8, 2026Shuai Guo, Yidong CuiBelief Updating in LLMs LLM Auditing
Beijing University of Posts and Telecommunications Beijing, China
PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs
Oct 7, 2026Linghao Meng, Feng He, Xuan Yang +5Hallucination in Language Models LLM Hallucination Mitigation
National University of Singapore · Independent Researcher · Tsinghua University +2
Copies or Sources? Measuring How LLM Aggregators Count Restated Evidence in Multi-Agent Systems
Oct 5, 2026Jianxin Gao, Runze Li, Tianyi Yu +3Belief Updating in LLMs Multi-Agent Communication
China Agricultural University · Jilin University · Tianjin University of Finance and Economics +1
Learning to Simulate Individuals from Macro Social Signals
Oct 5, 2026Yining Zhao, Bushi Liu, Haofei Yu +5Human Behavior Simulation RL for Language Model Reasoning
University of Illinois Urbana-Champaign · Independent Researcher

Oct 29, 2026J/K move · Enter open · S save

Surprise Alone Learns Lies: Evidence-Gated Continual Learning for LLMs from Streaming Claims

Abdelkader Metales, Miriam Grabenheinrich

caaise GmbH

Continual Learning for LLMs LLM Reliability Belief Updating in LLMs Data Provenance in Language Models

Open paper AskPDF arXiv

Continual learning methods for large language models increasingly decide what to learn by surprise, keeping the inputs the model finds least predictable. This rule is unsafe when the stream contains lies. A planted falsehood about something the model knows has low probability under the model's belief, so it is highly surprising by construction, and a surprise-driven learner absorbs nearly every such lie, at a rate similar to random selection with the same budget. We study the problem with LoRA-adapted students (Qwen2.5-1.5B, Qwen2.5-3B and Llama-3.2-3B) that learn from a stream of true and false claims about invented people and real CounterFact facts. We report four findings. (1) A verifier key, a judge that must approve each claim, removes lies to the extent that its verifier is accurate; real LLM judges are cautious and reject unsourced truths and even source-backed updates. (2) Calibrated regression tests on unrelated knowledge slow learning down without filtering out lies. (3) The original model's own prior flags claims that contradict what it knows, without a judge. It separates lies about known facts from true new facts (AUC 0.78-0.83 across three models) but not from legitimate updates of the same kind of fact (AUC 0.53). It saves 43% of judge calls, but an attacker with access to the model can choose lies that pass it. (4) If trusted sources exist that an attacker cannot plant, routing contradicted claims to evidence recovers legitimate updates and the combined system has the lowest poison rate among learned claims; when an attacker plants matching sources for 30% of the lies, the system ends up worse than the judge alone. The acceptance signals answer different questions, and none of them decides on its own whether a claim is true.