Beyond Local Surprise: Grounded Dialogue as Selective Belief Revision under Referential Uncertainty

Ziming Liu; Bhanu Chaitanya Jasti; Ziyang Xu; Hongyu Wu; Yi Wu; Jiqun Liu

cs.CLAug 26, 2026

Beyond Local Surprise: Grounded Dialogue as Selective Belief Revision under Referential Uncertainty

Authors: Ziming Liu, Bhanu Chaitanya Jasti, Ziyang Xu, Hongyu Wu, Yi Wu, Jiqun Liu

Organizations: School of Computer Science, University of Oklahoma · School of Library and Information Studies, University of Oklahoma · School of Information Studies, University of Wisconsin–Milwaukee

Abstract

When a speaker refers to a scene that the listener cannot directly see, the listener must decide whether to preserve its current understanding or revise it as new utterances arrive. Many language systems treat local mismatch as a cue for updating: divergence from the current understanding encourages adjustment. Yet conversational understanding may be more conservative, interpreting mismatching evidence relative to prior understanding rather than immediately revising it. We introduce a controlled, data-driven framework for turn-by-turn preserve/revise decisions in dialogue, where competing revision policies are learned under otherwise identical conditions. We compare four theory-driven revision strategies, each reflecting a different assumption about when listeners should preserve or revise. Two findings stand out. First, a mismatch-driven policy that updates solely based on local divergence reacts strongly to mismatch but destabilizes grounding and degrades retrieval. Second, an uncertainty-sensitive policy extends mismatch-based updating with accumulated evidence, preserving coherent understanding while maintaining strong retrieval performance. Surprisingly, coherent understanding emerges from a counterintuitive pattern: local mismatch promotes preservation, whereas accumulated uncertainty promotes revision, suggesting that listeners maintain prior understanding despite local mismatch and revise only when uncertainty sufficiently accumulates. This pattern is consistent with conceptual pact theory.

Figures & tables

Figure 1: Overview of the preserve-or-revise framework for grounded dialogue. Grounding evolves over dialogue turns by integrating local referential evidence with prior belief through a preserve-or-revise mechanism. The underlying architecture is fixed while revision policies vary: pure-corr , fixed , dcp-b , and multi-c .

	Retrieval					Grounding Dynamics
Regime	R@1	R@5	R@10	MRR	MedR $\downarrow$	Pres. $\downarrow$	Rev. Sens. $\uparrow$	Mean $\rho$
Random	0.0011	0.0056	0.0112	0.0083	447	—	—	—
clip-only	0.0873	0.2066	0.2630	0.1450	73	—	—	—
pure-corr	0.118 $\pm$ .005	0.314 $\pm$ .005	0.424 $\pm$ .006	0.219 $\pm$ .006	16.0 $\pm$ 1.0	0.050 $\pm$ .000	—	0.100
fixed	0.117 $\pm$ .006	0.315 $\pm$ .003	0.426 $\pm$ .006	0.218 $\pm$ .007	16.0 $\pm$ 1.0	0.049 $\pm$ .001	—	0.100
dcp-b	0.009 $\pm$ .006	0.052 $\pm$ .032	0.086 $\pm$ .048	0.038 $\pm$ .019	184.0 $\pm$ 121.3	0.001 $\pm$ .001	0.900 $\pm$ .015	0.203

Table 1: Retrieval performance and intrinsic grounding dynamics across policies (mean

\pm

std, three seeds). pure-corr and fixed remain within one standard deviation of multi-c on all retrieval metrics. dcp-b exhibits the strongest intrinsic dynamics, including large mismatch-triggered switches (Switch KL

=12.31\pm 0.88

), yet substantially underperforms on retrieval.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Condition	R@1	R@5	R@10	MRR
Full	0.112 $\pm$ .004	0.303 $\pm$ .011	0.418 $\pm$ .009	0.212 $\pm$ .002
Img-Zero	0.000 $\pm$ .000	0.000 $\pm$ .000	0.000 $\pm$ .000	0.002 $\pm$ .000
Img-Random	0.001 $\pm$ .000	0.004 $\pm$ .003	0.010 $\pm$ .004	0.008 $\pm$ .001
Img-Mean	0.000 $\pm$ .000	0.000 $\pm$ .000	0.000 $\pm$ .000	0.002 $\pm$ .000
Txt-Zero	0.000 $\pm$ .000	0.000 $\pm$ .000	0.000 $\pm$ .000	0.002 $\pm$ .000
Txt-Random	0.003 $\pm$ .002	0.006 $\pm$ .002	0.013 $\pm$ .001	0.010 $\pm$ .001

Appendix

Table 3: Inference-time modality ablation for multi-c (mean

\pm

std, three seeds). Replacing either modality collapses retrieval to near-random, confirming grounding depends jointly on both modalities.

Belief Revision

Belief revision studies how agents change, retain, or retract beliefs when new evidence conflicts with what they already accept.

See all 29 papers

Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations
Oct 10, 2026Alessandro Morosini, Wei Sun, Georgia PerakisState Tracking Language Model Error Detection
Massachusetts Institute of Technology · IBM Research
Mind the Gap: Theory-of-Mind-Grounded Friction for Epistemic Alignment
Aug 31, 2026Yifan Zhu, Kyeongmin Rim, James PustejovskyLLM Alignment Pragmatic Reasoning in Language Models
Brandeis University, Waltham, MA, USA
Using Semantic Uncertainty to Estimate Transition Relevance in Turn-taking
Sep 10, 2026Muhammad Umair, Jan P. de RuiterSpoken Dialogue Systems Turn-Taking Prediction
Department of Computer Science Tufts University Medford, Massachusetts, USA

Oct 10, 2026J/K move · Enter open · S save

Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations

Alessandro Morosini, Wei Sun, Georgia Perakis

Massachusetts Institute of Technology · IBM Research

State Tracking Language Model Error Detection LLM Self-Correction

Open paper AskPDF arXiv

Large language model (LLM) assistants are increasingly used in multi-turn conversations, where the user's intent is often unclear at first and revealed only gradually over time. The assistant must therefore interpret each exchange and form inferences about the user's intent, using them to update its working state. These updates usually refine the assistant's understanding, but a misinterpretation can instead corrupt its working state by ruling out the true intent, influencing subsequent turns and ultimately causing the conversation to fail. Efficient recovery then requires locating and revising the misinterpretation without discarding information that remains valid. We introduce Loci (Localization Of Corrupted Inferences), a framework for probabilistically localizing misinterpretations and selectively repairing the assistant's working state. Given an observed failure, Loci reasons over possible user intents and assigns each update a probability of having excluded the true intent. These probabilities rank suspected updates and quantify the model's confidence in the localization of the underlying misinterpretations. Loci then uses this information in a repair policy that decides whether to retain, verify, or remove suspected updates, enabling recovery from corrupted inferences while preserving useful information. To evaluate Loci's ability to identify misinterpretations, we first test it in a controlled setup with injected errors. We find that Loci improves localization over the baselines, with its advantage becoming larger when multiple errors occur. We then evaluate it in free-form conversations, where misinterpretations arise naturally, and measure the interactions needed to complete the task after an initial failure. Using localization confidence to guide repair, Loci reduces mean recovery cost by 46-79% relative to the strongest baseline across multiple domains.

Abstract

Belief Revision

Explore similar work

Where Did We Go Wrong? Localizing Errors to Repair Multi-Turn Conversations

Mind the Gap: Theory-of-Mind-Grounded Friction for Epistemic Alignment

Using Semantic Uncertainty to Estimate Transition Relevance in Turn-taking