From Propositional to Perceptual Asymmetry: Extending FPO to Asymmetric Partial Information Dialogue

2026-07-20 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing & Dialogue Systems · Depth: Expert, quick

Summary

Yifan Zhu, Kyeongmin Rim, and James Pustejovsky extend Frictive Policy Optimization (FPO) to address "perceptual asymmetry" in collaborative dialogue, a scenario where participants hold asymmetric partial information. Previously, FPO assumed shared perceptual contexts, focusing on "propositional asymmetry" where friction arose from differing interpretations of shared scenes. The researchers evaluated this extension through cross-corpora analysis and LLM probing on referentially asymmetric dialogue tasks, notably the HCRC MapTask. Their findings, presented in the Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue in August 2026, indicate that FPO's friction functional is empirically valid only when assessed from each participant's individual information horizon. They observed that specific landmark configurations produce distinct grounding failure modes, with ambiguous setups causing successful-appearing but silently diverging misunderstandings. The LLM probe further confirmed that an informed single viewpoint surpasses omniscient access to all participant contexts. The work also proposes annotation refinements for pending grounding states and alignment classification.

Key takeaway

For NLP Engineers developing collaborative dialogue systems, especially those involving asymmetric information, your models must account for individual participant perspectives. This research demonstrates that an informed single viewpoint outperforms omniscient access, suggesting that agents should prioritize understanding their own information horizon rather than attempting to synthesize all contexts. Incorporate annotation refinements like subtype decomposition of grounding states to better diagnose and mitigate silent divergences in dialogue.

Key insights

Understanding dialogue friction in asymmetric information requires evaluating FPO from each participant's unique perspective, as an informed single viewpoint outperforms omniscient access.

Principles

• Friction is an epistemic signal, not noise.
• Perceptual asymmetry impacts reference denotation.
• Individual information horizons are critical for grounding.

Method

The study extends FPO to perceptual asymmetry, evaluating it via cross-corpora analysis and LLM probing on referentially asymmetric dialogue tasks like the HCRC MapTask. It also proposes annotation refinements.

In practice

• Decompose pending grounding states.
• Use accommodation-aware alignment classification.
• Prioritize informed single viewpoints in LLM design.

Topics

• Frictive Policy Optimization
• Perceptual Asymmetry
• Dialogue Systems
• Common Ground
• LLM Probing
• HCRC MapTask

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Seeing Is Not Sharing: Some Vision-Language Models Overestimate Common Ground in Asymmetric Dialogue — Vision-language models often conflate potential common ground with established mutual understanding in dialogue.
• Epistemic Stance Flexibility Probing: Measuring Prompt-Conditioned Register Shift in Large Language Models — LLM epistemic stance flexibility, distinguishing attributed vs. self-claims, is orthogonal to general model capability.
• Unlocking Proactivity in Task-Oriented Dialogue — Conditioning on user latent concerns is key to unlocking proactive behavior in task-oriented dialogue agents.
• Common Ground inconsistencies in dialogue systems: conflict patterns implied by polar question forms — Polar question forms influence human ability to track Common Ground inconsistencies in dialogue systems.
• When Should Memory Stay Silent: Measuring Memory-Use Boundaries in Memory-Augmented Conversational Agents — LLMs often inappropriately integrate sensitive user memory, requiring explicit boundary management at retrieval and generation.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.