The paper investigates whether language models that produce the same probabilistic predictions after different types of evidence (textual or activation interventions) actually share the same underlying belief states. Using a Bernoulli inference task and an analytic Bayesian reference, the authors compare four textual routes and two activation-level interventions across three open-weight models, measuring both next-outcome continuations and explicit probabilistic estimates. They find that the format of evidence influences which readouts reflect its content, and that the effective evidence weight—how many new observations a belief is worth—varies widely across routes, with some elicitation questions altering inferred weights by orders of magnitude. In a one-shot steering scenario, an intervention that matches a natural record’s immediate prediction loses most of its effect when new evidence is introduced, demonstrating that agreement in current predictions does not guarantee equivalent belief updating.
By Weihan Li, Junhao Wu, Xinlei Chen, Tianshi Zheng
arXiv:2606. 30850v1 Announce Type: new Abstract: Large language models (LLMs) are typically deployed in multi-turn conversations, where each turn provides new evidence that should reduce epistemic uncertainty about their environment.
By Ankur Samanta, Akshayaa Magesh, Tal Lancewicki, Ayush Jain, Youliang Yu, Paul Sajda, Kaveh Hassani, Aditya Modi, Daniel R. Jiang, Yonathan Efroni
Large language models (LLMs) are increasingly used for clinical reasoning, yet their ability to revise judgments as patient evidence changes is uncertain. In this study, researchers evaluated longitudinal belief updating using real intensive‑care patient trajectories and found that conditioning on prior judgments often increased prediction error. Controlled experiments revealed two failure modes: models were more responsive to worsening than improving respiratory evidence, and they shifted estimates significantly when prior risk was altered, indicating a causal influence of prior beliefs. Prompting did not improve reliability, and an Evidence‑Validated Longitudinal Update (EVLU) approach produced fewer but more trustworthy revisions, highlighting a reliability‑coverage trade‑off.
By Min Zeng, Rui Zhang
Faithfully understanding the internal process that occurs as models form their chain of thought is an important problem with the rise in model capabilities. However, this remains challenging due to th...
The paper introduces a method for analyzing how interventions in a prompt affect a language model’s reasoning by using behavioral counterfactuals. It trains a module to edit the residual stream so that the next-token distributions for a fixed response match between a source and a counterfactual prompt, outperforming standard activation reconstruction. The resulting attribution maps and visualizations help identify causally important locations for counterfactual recovery and guide hypothesis testing with activation patching, as demonstrated in a case study on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B.
By Shawn Im, James Oldfield, Adel Bibi, Sharon Li
The study investigates whether prompting language models to provide five weighted estimates in a single response can reduce bias in their judgments. Across various bias tasks—anchoring, hindsight, base‑rate, conjunction, and status‑quo/default—the authors compare direct numerical answers with a multi‑estimate format, testing different source framings, random cues, and probability‑consistency constraints on three models (gpt‑6.1‑sol, claude‑sonnet‑5.5, and Qwen3‑8B). Results show that while direct elicitation reduces irrelevant‑anchor indices and status‑quo effects in two configurations, multi‑estimate elicitation does not consistently debias or attenuate responses, and the observed effects are limited by noise and lack of equivalence.
By An-Sheng Jhang
The paper introduces anchored factual recall, where language models must choose an object for a subject–relation pair conditioned on an explicit anchor such as a time. Using causal interventions on four open‑weight models, the authors identify two parallel computational routes: a subject‑mediated route that aligns the subject representation with the queried year, and a direct route that lets the final position attend to the anchor and uses upper‑layer MLPs to bias the answer. These routes combine to determine the model’s answer, showing that factual recall involves selecting among context‑dependent alternatives rather than retrieving a single stored object.
By Ayana Niwa, Masahiro Kaneko, Kentaro Inui
arXiv:2606. 22030v2 Announce Type: replace Abstract: We investigate when belief-based memory actually improves large language model (LLM) agents.
By Pranav Singh
arXiv:2606. 17832v1 Announce Type: new Abstract: Large language models (LLMs) are often hypothesized to perform implicit Bayesian inference, yet a key coherence condition, the martingale property of predictive beliefs, has been shown to fail in controlled synthetic in-context learning settings.
By SongEun Kim, Seungyoo Lee, Edwin Fong, Hyungi Lee, Juho Lee
arXiv:2603. 01437v2 Announce Type: replace Abstract: As chain of thought (CoT) has become central to scaling reasoning capabilities in large language models (LLMs), it has also emerged as a promising tool for interpretability, suggesting the opportunity to understand model decisions through verbalized reasoning.
By Kyle Cox, Darius Kianersi, Adri\`a Garriga-Alonso
arXiv:2610.33237v1 Announce Type: new
Abstract: How can a language model know whether it has enough evidence to answer? Language models tend to answer even when the evidence they have is insufficient...
By Jiayi Li, Dennis L. Shung, Bradly C. Stadie
Process Reward Models (PRMs) are crucial for scaling large language model reasoning at test time, but they suffer from two types of uncertainty: PRM uncertainty from limited supervision and imperfect training, and policy uncertainty from mismatched data collection policies. The paper introduces BayesianPRM, which treats process rewards as random variables, discretizes the reward space with an ensemble, and applies two belief updates—reliability‑aware updating for PRM uncertainty and conservatism‑aware calibration for policy uncertainty—to form a Bayesian belief whose expectation yields a more accurate reward estimate. Experiments on multimodal reasoning benchmarks show that BayesianPRM improves reward accuracy and test‑time scaling performance.
By Zhongjian Qiao, Zhixiang Zhou, Jiafei Lyu, Peisong Wang, Qiaosheng Zhang, Siyang Gao, Shuang Qiu