acceptodds
Under review as a conference paper at ICLR 2027

LLM Agent Collaboration Needs Trust Framing

Abstract

Trust is a cornerstone of human cooperation. Does it matter similarly when large language model (LLM) agents collaborate? We raise this question because of a surprising pattern in sequential multi-agent systems: receivers often ignore partner messages that contain exactly the information they lack. We call this pattern information neglect. In multi-hop question answering (QA) with distributed evidence, it accounts for of audited failures. We first tried to improve the message itself. Rewriting, compressing, structuring, and routing partner messages did not reliably help, and several variants hurt performance. This points to the receiver as the source of the problem. We therefore introduced Importance Framing (IF), which increases the weight a receiver gives to the partner message. IF helps on clean messages. Averaged across corrupted-message conditions, however, it falls slightly below the unframed baseline. We call this the importance dilemma. It arises because importance is a first-order notion: it scales attention to the whole message uniformly, so it amplifies correct and incorrect content alike. Drawing on social-science accounts of trust, we argue that trust is second-order. It governs how attention should be allocated by judging both whether the source is reliable and which parts of its message merit reliance. We operationalize this view in Trust Framing (TF), a training-free protocol that places source and task-context cues immediately before a partner message and requires no extra LLM calls. TF improves F1 by on multi-hop QA and task-completion rate by on the cooperative game Collab-Overcooked (both ). In a factorial ablation, the trust cue alone accounts for of the full gain. TF also outperforms the unframed baseline under every message-corruption condition tested. These results suggest that the concept of trust itself shapes how LLM agents use partner information: effective collaboration requires appropriate reliance on partner messages, not just more attention to them.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language model agents, Multi-agent systems, Agent collaboration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hErFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This work studied information neglect in LLM multi-agent systems. A receiving agent may fail to use relevant information communicated by another agent, even when that information is necessary to complete the task. The authors propose Trust Framing (TF), a fixed preamble that introduces the partner’s message using source and reliability cues, optional task-specific guidance, and an instruction to integrate the message with the receiver’s own information.

The method is evaluated on three question-answering benchmarks with distributed evidence and on the cooperative environment Collab-Overcooked. The authors report aggregate improvements of 9.5 percentage points in QA F1 and 7.6 percentage points in task-completion rate. Through component ablations and experiments with corrupted messages, they argue that effective collaboration requires appropriate reliance on partner information.

Critical strengths
  1. The motivation and beginning of the paper are quite interesting. I found the introduction compelling, particularly the sociological discussion on trust. Additionally, In relation to the where information or capabilities are distributed across agents, the failure to integrate a partner’s contribution can undermine the purpose of collaboration is also interesting. The human audit substantiates this problem and distinguishes receiver-side neglect from sender errors and other plausible failures.

  2. The paper reports substantial and significant aggregate gains for a simple intervention. The QA evaluation covers all 36 directed combinations of six model endpoints, and the evaluation in Collab-Overcooked provides evidence of past question answering tasks. The reported aggregate improvements are statistically significant, and the component ablations characterize which parts of the adaptor contribute to the performance. The method is also straightforward and simple to implement and requires no model training or additional search.

Critical weaknesses
  1. I find the presentation of the experiments difficult to assess. Although the setup is described across the main text and appendix, it takes substantial effort to mentally reconstruct a complete trial and understand the differences among the main experiment, component ablation, and corruption studies. In particular, the paper defines the corruption categories but does not show examples of original and corrupted messages. These examples are important for judging the severity of the corruptions and understanding what successful selective reliance is. The Collab-Overcooked results would similarly benefit from a worked example showing the kitchen state, each agent’s observations and available actions, the exchanged messages, and the resulting state transitions. Table 11 provides a short textual example, but it does not make the comparison sufficiently transparent. A consolidated summary of the experimental settings and a matched rollout visualization would substantially improve the presentation.

  2. The evidence for the narrow interpretation in terms of trust could be more clearly demonstrated. The “trust” component combines several elements, including partner provenance, exclusive access to information, and a reliability claim. The factorial ablation supports the usefulness of this combined block, but it does not fully distinguish the proposed interpretation from alternatives such as clearer source attribution or clarification that the partner holds missing evidence. More closely matched controls separating these elements, or narrower wording of the mechanistic claims, would make the conceptual contribution clearer.

  3. A minor presentation issue is the residual narrow text wrapping at the top of page 6, apparently carried over from Table 1 on the preceding page.

Questions for authors
  1. Could the authors provide representative clean and corrupted messages for each condition, together with the relevant construction and validation rules? It would also help to provide the complete condition-by-condition results for the separate primary corruption experiment, since Table 2 reports the exploratory PT/IF/TF comparison.

  2. Could the authors clarify how the task and evaluation budget are matched in Table 11? The passthrough example reports an 80-step timeout, while the TF example reports 14 completed orders over 623 steps. A comparison with the same initial state, recipe, action budget, and success criterion would make the coordination improvement easier to interpret.

  3. What motivated using Qwen3.5-35B-A3B as the sole adaptor model for the L1–L3 baselines? Wouldn’t Gemma or OLMo models have different trust capacity? How sensitive might the negative results for message rewriting be to that choice? Please also explain the selection of the three endpoints used for Collab-Overcooked.

  4. What evidence most directly distinguishes the proposed trust mechanism from neutral source attribution or an explicit reminder that the partner has access to missing information?

  5. It would be interesting to test whether prompt or context optimization using task-success feedback—for example, GEPA https://arxiv.org/abs/2507.19457 or CORE https://arxiv.org/abs/2605.28742—can discover similar instructions for using partner information without being initialized with TF. Evaluating the resulting prompts on held-out tasks and model pairs would help determine whether TF can emerge naturally through optimization and whether it provides additional benefit after optimization.

Reviewer_NpjLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper identifies an issue among existing work that they do not separate correct claims from incorrect ones within a message and proposes to include additional prompt blocks to alleviate this issue. The authors first verify that simply re-writing the message does not address the issue, and then propose their TF method including trust, family hint, and integration scaffold. Experiments show that TF increases the accuracy and task success to some extents across QA and agent tasks.

Critical strengths
  1. The motivation of "Existing methods do not separate correct claims from incorrect ones within a message" makes sense.
  2. The observation that simply re-writing partner message is not helpful from analytic experiments is valuable.
Critical weaknesses
  1. The presentation is not good. There are several descriptions, claims, and conclusions not concentrated, explained or justified. See comments below.
  2. The motivation claims that "Existing methods do not separate correct claims from incorrect ones within a message" but I do not see how TF resolves this problem.
  3. It is not convincing that the gain comes from TF rather than the priviledged information like the task structure and knowledge about peers holding relevant information.
Questions for authors
  1. The summary and introduction are not informative, repeating claims with lots of terms and details defined only in later sections.
  2. L213: TF is not introduced clearly. The notation looks confusing.
  3. It is mentioned that "IF increases the weight a receiver gives to the partner message" but the current implementation looks like only some kind of rewriting. How is it confirmed that the weight is increased?
  4. It is mentioned that "TF adopts fewer than 1% of injected wrong candidates." but I did not find any relevant evidence.
  5. Sec 4.1: The content is not concentrated and relevant to the main logic. If the automated categorization is neither accurate nor aligned with humans, then only results of the human evaluation should be included.
  6. Table 1: Why are the numbers of pairs different? Where can I find evidence for the following conclusion (specifically the +4.7/+6.7 results)? What is the insight behind this? If the overall results are negative here, why is it necessary to mention that they are positive at small scale?

compressed and profile-adapted initially appear positive at that scale but reverse upon scaling ...

  1. Minor: the wraptable for table 1 is not working correctly.
  2. L295: To make the results comparable, the hint-free variant should be the main version. Also the integration scaffold to some extent provides priviledged information.
  3. How are the results in Table 5 obtained? Is it from automatic catogerization or human examination?
  4. The writing around Table 2 (starting L369) is strange. It is hard for me to realize that those contents are about Table 2, which may be the most important parts that attempt to verify the motivation. Furthermore, more quantitative (e.g. Table 5) and qualitative evidence is needed to show that TF helps in distinguishing correct and misleading evidence in the same piece of message.
Reviewer_fBbjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper looks at two-agent collaboration where each agent has different evidence. It identifies 'information neglect,' where the receiver ignores a partner's message that contains exactly the evidence it lacks. In a human audit of 460 failed QA cases, this accounts for 51.3%. Rewriting, compressing, routing, or memory-based adaptation of the message doesn't solve the problem. Importance Framing helps with clean messages but not with corrupted ones. Drawing on sociological work on trust, the authors propose Trust Framing (TF). TF is a training-free preamble of about 100 tokens that signals to the receiver that the message comes from a partner with exclusive access to the data and asks the receiver to incorporate it. TF improves F1 by 9.5 across 36 model pairs on three multi-hop QA datasets and improves task completion by 7.6 on Collab-Overcooked. The trust sentence on its own accounts for 76% of the gain.

Critical strengths
  • Simple and low-cost method. TF is just a fixed string prefix.
  • Broad evaluation across model pairs. The experiment covers 36 directed pairs across six models. The experiments also include Collab-Overcooked, which is a different type of task than QA.
Critical weaknesses
  • The method doesn't implement the theory. Trust is defined as second-order, with per-claim acceptance dependent on source and context. The actual preamble, however, simply says "Every claim below is verified," which is a message-level cue with no instruction to examine the claims in any other way. Under the paper's own definitions, it's actually a much stronger first-order reliance cue.
  • Mechanism claims are not tested. The signal-detection argument is made but not measured. It's claimed that importance shifts the operating point and that trust improves discriminability. For instance, there is no indication of how often PT, IF, and TF actually adopt the correct content versus the injected wrong content.
Questions for authors
  1. Can you report, for PT, IF, and TF, how often the receiver adopts the injected wrong candidate versus the correct answer in the explicit-conflict and false-answer conditions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.