Prospective Interpretation Risk: Principled Communication Control Between LLMs

Wanrong Yang, Rehan Deen, Julian Ma, Yuheng Fan, Yaoyu Jin, Taher Jafferjee, Ziquan Liu, Dominik Wojtczak, Yalin Zheng, David Henry MguniarXiv 2026 · arXiv

TL;DR

This paper introduces Prospective Interpretation Risk (PIR), a calibrated metric that predicts how likely a receiver LLM is to misinterpret a sender’s message before it is even sent. By separating interpretation failure from downstream execution failure, the authors enable "receiver-aware" communication where agents can proactively rewrite messages or selectively query the receiver to ensure they are understood correctly.

What Problem Does This Paper Address?

As we move from single LLMs to complex multi-agent systems (like AutoGen or MetaGPT), communication becomes a major point of failure. The core issue is heterogeneity: different models (with different architectures, prompts, or fine-tuning) can interpret the same natural language instruction in wildly different ways.

For example, a planner might tell a triage agent to "prioritize urgent cases." One model might interpret "urgent" as symptom severity, while another sees it as time sensitivity. Current uncertainty methods usually focus on whether a model is confident in its own answer, but they rarely estimate how a specific receiver will reconstruct a task from a message. This paper aims to turn communication from a "send and pray" process into a controlled, decision-theoretic one.

Methods

The authors model communication as a sender-receiver problem with a latent receiver "type." Their framework consists of three main pillars:

  1. Risk Separation: They distinguish between Interpretation Risk ($Y^I$—did the receiver understand the task?) and Capability Risk ($Y^A$—can the receiver actually perform the task?). This is crucial because a message rewrite can fix a misunderstanding, but it can't fix a model's lack of underlying ability.
  2. Black-Box Probing: To make measurement scalable, they use independent probes. Instead of analyzing free-form text, they ask the receiver to pick the intended task from a set of candidates. This provides a clear behavioral label for training.
  3. Value of Interpretation Information (VoII): Querying a receiver for clarification is costly (latency, tokens). The authors introduce VoII, a decision-theoretic rule that only triggers a query if the expected improvement in communication outweighs the cost. This prevents "query flooding" in large networks.
Figure 1: Framework overview. Top: receivers read one message as different tasks (1), PIR separates interpretation from capability risk (2), and the sender infers the receiver and queries only if worthwhile (3). Bottom: offline probes train the PIR model, which picks the verified rewrite to send.

Datasets

The researchers evaluated their framework using six knowledge-base question-answering (KBQA) datasets:

  • FreebaseQA, WebQSP, GrailQA, SimpleQuestions-Wikidata, PopQA, and Mintaka.

They tested these against a pool of five diverse open-weight receiver models: Nemotron-3 Nano, Gemma 4, gpt-oss-20b, Mistral Small 3.2, and Qwen3.6.

Key Insights and Ideas

  • Interpretation is Receiver-Specific: Interpretation failure rates aren't just a property of the message; they vary by 4–13× across different receiver models. A message that is perfectly clear to Qwen might be baffling to Nemotron.
  • Execution is a Poor Proxy for Understanding: The authors found that 32% of misread tasks still "passed" the final answer check by accident, while 56% of correctly understood tasks failed due to lack of capability. You cannot rely on the final output to know if your agent was understood.
  • Information $\neq$ Decision Value: Standard "Information Gain" (IG) queries to identify a receiver's type often don't help. VoII is superior because it recognizes that identifying the receiver is only useful if it actually leads the sender to choose a different, better message.
Figure 2: Interpretation failure is separate from execution fail- ure and depends on the receiver (test split). (a) Y A after a correct read (◦) and a misread (•) of the probe. (b) Y I per receiver. Bands span the lowest to the highest receiver.

Key Results

  • Better Prediction: Their receiver-conditioned PIR model reduced calibration error by 68% compared to receiver-agnostic baselines.
  • Reduced Failure: PIR-guided message revision reduced interpretation failures by 44% relative to the original message and 40% relative to generic "make this clearer" rewrites.
  • Efficient Querying: The VoII strategy outperformed both random querying and pure Information Gain. It lowered interpretation failure while only querying in 18.2% of episodes, effectively spending its "query budget" only where the receiver's identity truly mattered for the message choice.

Limitations and Future Work

  • Narrow Task Scope: The study focused on short KBQA questions. Real-world agent handoffs (like passing code or multi-step plans) are much more complex and offer more ways to be misread.
  • Closed-Set Receivers: The current model assumes the receiver belongs to a known pool of models. Future work needs to address "open-set" modeling for unknown or custom-tuned receivers.
  • Static Interaction: The experiments were largely single-turn. The authors suggest that multi-round "query-revise-query" loops and longer communication chains (multi-hop) are the next logical steps for this research.

Summary updated

Abstract

Large language model (LLM) agentic systems increasingly rely on models communicating with one another, yet existing uncertainty and multi-agent methods rarely estimate how a particular receiver will interpret a message before it is sent. This matters in heterogeneous systems, where capable receivers can reconstruct different tasks from the same message. We model this as a sender-receiver problem with a latent receiver type and define prospective interpretation risk (PIR): the probability that a receiver reconstructs a task other than intended. Rather than model an LLM's full input-output behaviour, we use black-box probes relating messages, intended tasks, and receiver-specific reconstructions, yielding scalable supervision while separating interpretation from downstream capability failure. Offline, heterogeneous frozen receivers provide supervision for receiver-conditioned risk and the effects of predefined mutable message features. At deployment, history induces a posterior over receiver types, guiding message revision and selection. We introduce value of interpretation information (VoII), querying for receiver information only when its expected communication benefit exceeds its cost. Our theory characterises when receiver information has decision value and bounds such queries. Empirically, interpretation-failure rates vary by 4-13x across receivers. Receiver information reduces PIR calibration error by 68% relative to a receiver-agnostic predictor, largely by correcting receiver-specific risk levels. PIR-guided revision reduces interpretation failure by 44% relative to the original message and 40% relative to a generic rewrite, mostly through a repair that helps every receiver. VoII outperforms information-gain and random querying at matched cost on the interpretation objective it optimises, lowering interpretation failure from 3.84% to 3.79% while querying 18.2% of episodes.
BibTeX
@misc{yang2026prospective,
  author = {Wanrong Yang and Rehan Deen and Julian Ma and Yuheng Fan and Yaoyu Jin and Taher Jafferjee and Ziquan Liu and Dominik Wojtczak and Yalin Zheng and David Henry Mguni},
  title = {{Prospective Interpretation Risk: Principled Communication Control Between LLMs}},
  year = {2026},
  eprint = {2609.33885},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.33885},
}
Figure