💬 NLP

From Propositional to Perceptual Asymmetry: Extending Frictive Policy Optimization to Asymmetric Partial Information Dialogue

This paper extends Frictive Policy Optimization to address perceptual asymmetry in dialogue by demonstrating that friction serves as a critical epistemic signal for grounding, revealing that successful communication relies more on an informed single perspective than omniscient access to all contexts, and proposing specific annotation refinements to better capture these dynamics.

Original authors: Yifan Zhu, Kyeongmin Rim, James Pustejovsky

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yifan Zhu, Kyeongmin Rim, James Pustejovsky

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you and a friend are trying to solve a puzzle together, but you are both looking at different halves of the same picture. You can't see what your friend sees, and they can't see what you see. This is the core problem the paper tackles: how do two people build a shared understanding when they are working with different pieces of information?

The authors, Zhu, Rim, and Pustejovsky, are studying how AI agents (and humans) talk to each other when they have "asymmetric" views of the world. They propose a new way to measure and fix the "friction" (confusion, misunderstandings, and repairs) that happens during these conversations.

Here is a breakdown of their ideas using simple analogies:

1. The Old Way vs. The New Way

The Old Way (Propositional Asymmetry):
Imagine you and a friend are standing in the same room looking at a red block on a table. You both see the same block, but you argue about whether it is "red" or "orange." The old AI models assumed everyone sees the same scene; the problem was just that they interpreted the facts differently.

The New Way (Perceptual Asymmetry):
Now, imagine you are looking at a map of a city, and your friend is looking at a different map of the same city. You see a "church" on your map, but your friend's map has a "school" in that exact spot. You aren't arguing about the color of a building; you are arguing because you are literally looking at different things. The authors call this "Perceptual Asymmetry."

2. The "Friction" Detector

The paper introduces a concept called Frictive Policy Optimization (FPO). Think of "friction" not as a bad thing to be eliminated, but as a warning light on a car dashboard.

Old View: If the dashboard light flickers, it's just noise. Ignore it and keep driving.
New View: The flickering light is a signal that your internal map doesn't match your friend's. It's a chance to stop, check, and fix the route before you crash.

The authors argue that to fix these misunderstandings, an AI shouldn't try to be "all-knowing" (seeing both maps at once). Instead, it needs to act like a person who only sees their own map.

3. The "All-Knowing" Trap

This is the paper's most surprising finding. The researchers tested AI models with two different "eyes":

The Omniscient Eye: The AI sees both maps at once.
The Perspective Eye: The AI sees only the map belonging to the person speaking.

The Result: The AI with the Perspective Eye was actually better at spotting misunderstandings than the one with the Omniscient Eye.

The Analogy: Imagine you are a referee watching a game.

If you stand on a high tower and see the whole field (Omniscient), you might get confused by all the players moving around and miss the specific moment a player stepped out of bounds.
If you stand right next to the player (Perspective), you see exactly what they see. You know immediately if they are confused because you are limited to their view.

The paper found that when AI tried to see everything, it got "distracted" by the extra information and missed the subtle clues that a misunderstanding was happening. When it was forced to look only through one person's eyes, it became much sharper at detecting trouble.

4. The "Silent Crash"

The researchers found a specific type of misunderstanding that is very dangerous: The Silent Divergence.

Scenario: You say, "Go to the big meadow." Your friend has two meadows on their map. They pick one, and you say, "Great!"
The Problem: You both think you are talking about the same thing, but you aren't. You are "aligned" on the surface, but you are actually heading to two different places.
The Finding: These "silent crashes" happen most often when there are multiple similar things (like two meadows or two churches). The AI needs to learn to spot these specific "multiplicity" traps, not just general confusion.

5. What Should We Change?

Based on this, the authors suggest two simple upgrades for how we train and label AI conversations:

Don't just say "It's wrong." When a conversation stalls, we need to know why. Is it because the listener is confused? Or because the speaker used a vague word? The paper suggests breaking down "pending" (unsolved) states into smaller, specific categories.
Spot the "Silent Agreement." Sometimes people agree without actually fixing the problem. The authors suggest we need to distinguish between "negotiated alignment" (where we argued and fixed it) and "accommodated alignment" (where one person just gave up and pretended to understand).

Summary

The paper argues that in collaborative tasks where people have different information, trying to be all-knowing makes AI worse at spotting confusion. To build better AI teammates, we should design them to reason from a single, limited perspective—just like a human does. By embracing the "friction" of limited views, AI can catch misunderstandings earlier and fix them before the conversation goes off the rails.

Technical Summary: From Propositional to Perceptual Asymmetry

Problem Statement
Collaborative dialogue relies on the construction of common ground, a process inherently prone to friction (misalignment, misunderstanding, and repair). While Frictive Policy Optimization (FPO) reframes this friction as an epistemic signal essential for robust alignment, existing implementations (e.g., FAAF) operate under a propositional asymmetry assumption. In this regime, participants share a perceptual context (the same scene) but arrive at different beliefs.

This paper identifies a limitation in applying FPO to perceptual asymmetry, where participants hold asymmetric partial information (e.g., different maps or views). In such settings, identical surface expressions can yield different denotations depending on whose information state grounds the reference. The core problem is that standard FPO formulations, which assume a shared perceptual context, fail to account for the fact that the "interaction history" ( $h$ ) is participant-specific ( $h^{(i)}$ ). Consequently, omniscient evaluation of friction may miss critical misalignments or misclassify them as noise.

Methodology
The authors extend FPO to a perspectival formulation and evaluate it through cross-corpus analysis and Large Language Model (LLM) probing.

Theoretical Extension:
- The authors redefine the frictive state as a configuration where the friction functional $F^-(h^{(i)}, y) > 0$ for at least one plausible response, evaluated relative to a participant's private information horizon $h^{(i)}$ .
- They decompose the four epistemic failure modes of $F^-$ $F^{-}$ (Uncertainty, Contradiction, Hazard, Value Conflict) into role-specific diagnostics:
  - Speaker (Giver): Primarily detects Value Conflict (production underspecification) and Uncertainty (within-context ambiguity).
  - Addressee (Follower): Primarily detects Contradiction (cross-perspective grounding conflict) and Hazard (confident miscommitment).
- This formulation treats common ground not as a presupposition but as an active construction reconciling perceptually disjoint evidence.
Datasets:
- HCRC MapTask: The primary evaluation target. The authors utilize two annotation layers: the original NXT dialogue act tags and the GMMT (Grounded Misunderstandings in MapTask) dataset, which provides per-perspective annotations of intended vs. interpreted landmarks.
- OneCommon: Used as a secondary cross-corpus comparison point to observe referential asymmetry in a different task structure (dot selection).
Experimental Design:
- Corpus Analysis: Cross-tabulation of landmark discrepancy types (Identical, Lexical, Existence, Multiplicity) against GMMT pending subtypes and alignment trajectories.
- LLM Probing: Three instruction-tuned LLMs (GPT-5, Qwen3.6, Llama-4-Scout) were tested under three conditions:
  - Omniscient: Access to both participants' maps.
  - Giver Perspective: Access only to the giver's map.
  - Follower Perspective: Access only to the follower's map.
- Repair-Productive Evaluation: A "RepairScore" metric was used to determine if detected frictive states corresponded to misalignments that were subsequently repaired in the dialogue.

Key Contributions

Perspectival Frictive States: The paper formalizes a role-specific decomposition of friction where the speaker incurs friction from within-frame ambiguity and the addressee from cross-frame conflicts.
Empirical Validation of Perspective-Bound Evaluation: Through LLM probing, the authors demonstrate that restricting a model to a single participant's view improves the detection of referential misalignment compared to omniscient access.
Annotation Refinements: The authors propose two specific refinements for grounding annotation schemes:
- Subtype Decomposition: Distinguishing between pending subtypes (e.g., ungrounded vs. quantificational) to capture distinct failure modes.
- Accommodation-Aware Alignment: Differentiating between alignment achieved through explicit negotiation and alignment achieved through silent accommodation.

Results

Discrepancy-Dependent Friction: Friction is not uniform; different discrepancy types produce qualitatively distinct grounding failure profiles. Multiplicity discrepancies (where a landmark appears multiple times on one map) drive a disproportionate share of misunderstandings (12.0% rate vs. 0.2% for identical landmarks) and are characterized by high rates of ungrounded pending states.
Trajectory Analysis: Multiplicity chains exhibit high rates of non-monotonic state trajectories (38.9%), where alignment is achieved locally but decays globally as subsequent references reintroduce ambiguity. This supports the Hazard component of FPO (trajectories that appear successful but silently diverge).
LLM Probing Performance:
- Perspective Superiority: Across all models, the Giver-perspective condition significantly outperformed the Omniscient condition in detecting friction (F1 gaps of +0.47 to +0.69).
- Omniscient Failure: Omniscient models often failed to detect misalignment, treating cross-map contradictions as noise or defaulting to "yes" (alignment) due to the presence of shared entities. Restricting the view to a single perspective surfaced the structurally relevant subset of information, improving detection.
- Repair Productivity: The perspectival formulation successfully identified all 13 gold-standard repair events in the test set, whereas the omniscient baseline missed several. This confirms that perspectival signals are not just detectable but are repair-productive.
Accommodation vs. Negotiation: Silent accommodation (flipping interpretation without friction moves) was rare (0.2% of aligned REs). The dominant pattern was negotiated alignment, suggesting participants effectively deploy friction when interpretations shift, but the bottleneck lies in tracking structural ambiguities that resurface across the dialogue chain.

Significance and Claims
The paper claims that collaborative dialogue is not merely a "shared-context problem" solvable by larger context windows or more powerful models. Instead, when agents coordinate under private information, friction resides in the gap between their information horizons.

The authors argue that more context can obscure the signal: having access to both participants' perspectives (omniscience) can lead to underestimating asymmetric knowledge and missing critical misalignments. A "smaller" perspectival model, reasoning from a single information horizon, outperforms an "omniscient" frontier model in detecting and repairing misalignment.

The work establishes that FPO's friction functional is empirically valid only when evaluated from within each participant's information horizon. This challenges the default assumption in agent design that misalignment is a coverage problem, proposing instead that effective alignment requires reasoning from a perspective, not over all perspectives simultaneously. The paper concludes that for multi-agent systems and human-AI teaming, friction-aware systems must be designed to reason from specific, private information states.