- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper talks about how multi agent LLMs can get the answer correct for a given query for a collective decision making task, but the "state" might be erroneous, which the authors refer to as "off query failure". The authors introduces OffQuery, based on established datasets from medical domain and disaster preparedness domain, and evaluate three things:
- T1: identify which agent is the one with misleading info
- T2: reconstruct the shared state
- T3: answer the query correctly Based on the evaluation and diagnosis, it proposed ReGround, a framework that improves multi-agent LLMs performance on all three metrics, especially targeting at T1 and T2.
Critical strengths
- In general I like this problem the authors are targeting at, and I like how the authors are spending effort on building a evaluation framework to seperate state reliability from task accuracy. This is especially clear with the modeled shared state C*, where the benchmark can score what survives collaboration independently of the answer. And this is validated with LLM judges plus human annotators. But I also have some concerns about this part, see the weaknesses part below.
- I also like the arc from problem formulation, to benchmark construction, to evaluation and diagnosis, and to mitigation. Especially the last two phaes on that arc, where we can see clear interventional evidence that state quality drives downstream accuracy.
Critical weaknesses
- I'm quite iffy about how the benchmark is actually testing the failure. The paper’s main concern is errors that goes unnoticed in the shared state because the query never touches them. But in OffQuery, the false information is generated specifically to push the model toward the wrong answer on the query. So if I understand correctly, when the model still answers correctly, it has resisted a direct attack, and it's not just simply the misleading info is left unused. When I skim through the appendix, though, I do see one small experiment (E.2) actually checks whether errors show up on a different question later. I understand the benchmark tries to isolate what it wants to tests, which I value, but the setup also makes the false information quiet easy to catch, as there is always exactly one misleading agent among truthful ones, its claims are always contradicted by someone else, and it may even sound different from the others in style (e.g. the very example you show in the teaser figure in intro). Thus the main results do show how well models handle obvious, targeted misinformation, but not really whether hidden errors survive in realistic shared states, which is the paper’s central claim. I'm willing to hear more about the authors' thoughts here. Please correct me if I'm wrong.
- The gains produced by ReGround is a bit "inflated" by weak baselines. I was surprised to see the massive gains of T1 by 309%. But the standard collaboration baseline scores about the same as random guessing on this task as you measured. This is likely because it only sees the agents’ opinions after discussion, by which point the misleading agent’s input has been mixed in with everyone else’s. T2 has a similar issue, where the baseline agents are told to share information “briefly”, so much of their low score comes from leaving out true facts, not from keeping false ones. Since the paper is about false information surviving in the shared state, these two problems make it quite hard to tell how much ReGround actually helps with that.
- A minor issue: I understand ReGround is very specific to this task as it's informed directly by the failure diagnosed. Right now it's so closely coupled with this task setting, that I can barely see its generality to more general settings. I understand how the arc progressed, and thus I suggest the author frame ReGround as evidence of the benchmark, which is the main contribution, and the diagnosis are actionable, instead of framing this framework as a standalone main thing on its own.
- And a even more minor issue: GPT-4o is doing a lot of work in this paper. I encourage the authors to switch to newer models, as we don't want model's capability as judges for example, to cause issues.
Questions for authors
These questions of mine are coupled with the first two weaknesses I pointed out:
- On weakness 1: (a) As I talked about earlier, the false information in OffQuery is designed to change the answer to the current question, so it isn’t really “off-query.” Could you add an experiment where the false information is unrelated to the current question, and then test whether it causes errors on a later, different question? (b) In that E2 experiments I pointed out, 43% of plans were judged inconsistent. How often exactly are the two plans built from clean states judged inconsistent? Without that comparison, it’s hard to tell how much of the 43% is caused by the false information.
- On weakness 2: (a) As I mentioned earlier, the standard baseline identifies the misleading agent at about the rate of random guessing. How does ReGround compare to a fairer baseline that sees the agents’ original information instead of their post-discussion opinions? Please report the absolute differences, not just percentages. (b) Could you split the T2 score into two parts: how many true facts are kept and how many false facts are removed? Only the second part measures the problem the paper is about.
If these questions are answered with good quality and we see the headline results still hold, I will consider raising my score.
5 comments
Reply (2/3)
W2 & Q2(a). Fairer verification baselines and original evidence access
How does REGROUND compare with a baseline that accesses agents' original information? Please report absolute differences.
Our submission already includes a centralized T1 verifier with direct access to all agents' original private observations (§D.3, Table 1 in submission).
For GPT-4o, the T1 results are:
| Dataset | Standard | ReGround | Centralized verifier |
|---|---|---|---|
| Healthcare-S | 15.2% | 56.7% | 65.3% |
| Healthcare-I | 14.0% | 54.5% | 68.0% |
Relative to Standard, ReGround improves T1 by 41.5/40.5 percentage points on Healthcare-S/I. The centralized verifier achieves 65.3/68.0%, confirming that direct access to original private evidence can substantially improve source identification.
To further examine the impact of original evidence access on state reconstruction and downstream reasoning, we conduct additional comparisons with two variants:
- Retain: Preserve all agents' original observations, including those of the source identified as misleading in T1.
- Remove: Exclude the T1-identified misleading source before subsequent reasoning.
Both Standard and ReGround retain their respective reasoning procedures.
| Setting | H-S FRR | H-S T3 | H-I FRR | H-I T3 |
|---|---|---|---|---|
| Standard, Original | .211 | .669 | .242 | .627 |
| Standard, Retain | .174 | .495 | .166 | .476 |
| Standard, Remove | .196 | .612 | .199 | .574 |
| ReGround, Original | .657 | .766 | .668 | .661 |
| ReGround, Retain | .648 | .633 | .631 | .617 |
| ReGround, Remove | .683 | .888 | .705 | .792 |
The results reveal three patterns:
- Original evidence access alone provides limited benefits. Retaining all observations decreases T3 accuracy for both methods, indicating that unfiltered evidence can introduce additional interference.
- Source filtering improves downstream reasoning. For ReGround, Remove improves T3 over Retain by 25.5/17.5 percentage points on Healthcare-S/I, alongside FRR gains of 3.5/7.4 points.
- ReGround's advantages persist under matched evidence-access conditions. In the Remove setting, ReGround exceeds Standard by 48.7/50.6 points in FRR and 27.6/21.8 points in T3 on Healthcare-S/I.
These comparisons demonstrate that ReGround's improvements in misinformation removal and task resolution persist when both methods operate under the same original-evidence access protocol. They further highlight the importance of explicit verification and state reconstruction. We add these results and analysis in Table 14, §5.2 (lines 403-418), and §D.9 in updated manuscript.
W2 & Q2(b). Separating true-fact retention and false-fact removal
Could you split T2 into how many true facts are retained and how many false facts are removed?
Yes. Our existing atomic-fact annotations (§B.5 in submission) separately identify reliable and misleading facts, allowing us to decompose T2 into:
- True-fact Retention Rate (TRR): TP / (TP + FN)
- False-fact Removal Rate (FRR): TN / (TN + FP)
We report the results for GPT-4o and Gemini-3.5-Flash across all three datasets here and complete results in Table 11 in updated manuscript:
| Model / Method | H-S TRR | H-S FRR | H-I TRR | H-I FRR | Disaster TRR | Disaster FRR |
|---|---|---|---|---|---|---|
| GPT Standard | .335 | .211 | .331 | .242 | .653 | .513 |
| GPT ReGround | .807 | .657 | .787 | .668 | .777 | .845 |
| Gemini Standard | .472 | .422 | .465 | .413 | .664 | .656 |
| Gemini ReGround | .840 | .794 | .825 | .801 | .799 | .910 |
ReGround consistently improves both true-fact retention and false-fact removal across the evaluated models and datasets.
On Healthcare-S/I, FRR improves by 42.6–44.6 percentage points for GPT-4o and 37.2–38.8 points for Gemini.
The improvements are also substantial in Disaster Response:
- GPT-4o: FRR improves by 33.2 points (51.3% vs. 84.5%), alongside a 12.4-point TRR gain.
- Gemini: FRR improves by 25.4 points (65.6% vs. 91.0%), alongside a 13.5-point TRR gain.
These results directly address the concern that low T2 scores may primarily reflect omitted reliable facts. ReGround achieves substantial improvements in removing misleading information while preserving more reliable facts. The gains therefore reflect improved retained-state reliability across both dimensions. We include the complete decomposition in §D.5 and Table 11 in updated manuscript.
Reply (3/3)
W3. Contribution and framing of ReGround
I suggest framing REGROUND as evidence that the benchmark identifies actionable failures.
We agree that OffQuery and its diagnosis are the central contributions. The benchmark provides an evaluation framework for identifying retained-information failures that current-task accuracy misses. ReGround demonstrates that the diagnosed failures can be mitigated through explicit evidence verification and shared-state reconstruction.
We will revise the introduction and contribution summary to emphasize this relationship and clarify ReGround's role as an actionable mitigation of the failure identified by OffQuery.
W4. Reliance on GPT-4o
I encourage the authors to switch to newer models.
Our main experiments already cover newer models, including GPT-5, GPT-5-mini, and Gemini-3.5-Flash (Table 1 in submission), alongside additional Qwen models.
For T2 evaluation, §B.6 independently validates GPT-4o-based judgments using Gemini-3.5-Flash and Qwen3-235B-A22B. All three judges produce identical aggregate rankings across the evaluated configurations. §B.7 in submission further reports human validation with 0.88/0.84 agreement between GPT-4o and human annotators.
GPT-4o is retained as a fixed model for controlled benchmark construction, while the cross-model and human evaluations assess the reliability of the resulting fact-level scoring.
Reply (1/3)
Thank you for the constructive review and for recognizing the value of evaluating shared-state reliability independently of task accuracy. We have conducted additional experiments addressing your specific questions and clarify the existing evidence below.
W1 & Q1(a). Does OffQuery capture genuine off-query failures?
The original answer-targeted construction ensures that misleading information can affect downstream reasoning if accepted. Whether this information influences the current decision depends on which facts the model uses.
Our existing analyses provide evidence of this distinction:
To further examine the role of answer-targeted corruption, we introduce an additional question-blind misinformation setting. We sample 100 instances from Healthcare-S and Healthcare-I and generate misleading information using GPT-4o conditioned on the reliable patient context, without providing the original downstream question. The generated statement is then incorporated into the patient context. We evaluate the same instances under the original answer-targeted setting for comparison.
We evaluate GPT-4o and Gemini-3.5-Flash under both settings:
Two findings are particularly relevant:
We additionally apply the broader clinical planning evaluation from §E.2 in submission to the question-blind setting. Using GPT-4o, the plan inconsistency rates are:
These results show that unreliable retained information continues to affect broader clinical reasoning under question-blind corruption. ReGround substantially reduces these downstream inconsistencies. We add these results in §6 (lines 447-474) and Table 4 in updated manuscript.
W1 & Q1(b). Clean-clean control for broader clinical planning
We conduct an additional clean-clean control to quantify natural variation in clinical plan generation. We sample 100 instances where T2 and T3 are correct and generate two plans using clean states under the same evaluation protocol.
The clean-clean inconsistency rate is 12%, compared with 43% for the corrupted-clean comparisons reported in §E.2 in submission.
The 12% result provides an empirical reference for natural plan-generation variability. The observed corrupted-clean inconsistency rate is 31 percentage points higher. These analyses use separately selected samples, so we interpret the comparison as supporting evidence. Our controlled state-repair experiments (§E.3 in submission) provide complementary interventional evidence that improving shared-state reliability improves downstream reasoning. We add these results and analysis in §6 (lines 441-446) in updated manuscript.