acceptodds
Under review as a conference paper at ICLR 2027

Told Who Could Know: False Access Labels Mislead LLM Aggregators

Abstract

A large language model (LLM) can act as an aggregator: it reads other models' answers and returns a final answer. We test whether access labels, which claim who received information needed to answer, influence this choice. Four LLMs answer from assigned options or documents, so we know which labels are true. We swap only the labels, keeping questions and responses fixed. Aggregators see assigned inputs only through responses. Across five aggregators on multiple-choice questions (MMLU-Pro and SuperGPQA) and document questions (2WikiMultiHopQA and MuSiQue), replacing true labels with false ones shifts answers toward the falsely labelled model and lowers accuracy. On MMLU-Pro science, technology, engineering, and mathematics (STEM) questions, accuracy with full responses is 20.8–37.2 percentage points lower under false labels than without labels. Adding accompanying text to final answers narrows the true-versus-false accuracy gap on STEM and both document benchmarks. On STEM, we tell aggregators that the positively labelled model almost never received the correct option. All five aggregators select that model's answer less often after disclosure. Three still select that model's answer 15.2–29.2 points more often than without labels, and all five remain less accurate. False access labels can redirect aggregation even when their unreliability is stated.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, multi-agent systems, LLM aggregation, source attribution, behavioral evaluation, AI safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uS3nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the risks posed by false access labels in LLM multi‑agent aggregator scenarios. An access label indicates whether a sub‑agent member has obtained the information required to answer a question, rather than labelling whether an answer is correct. The authors design controlled experiments where the full response texts of sub‑agent members are kept fixed, while only the access labels are swapped. The experiments are conducted with five aggregator models across multiple‑choice benchmarks (MMLU‑Pro, SuperGPQA) and multi‑hop document QA datasets (2WikiMultiHopQA, MuSiQue).

Critical strengths

It features a rich set of task types, covering both multiple‑choice question (MCQ) and multi‑hop document question‑answering paradigms.

It explicitly points out that matching a member’s answer does not equate to answering correctly. Particularly in document‑based tasks, members without access to supporting evidence may still arrive at the correct answer by chance. The metric design is well‑aligned with task‑specific characteristics.

Critical weaknesses

The paper repeatedly acknowledges that the underlying mechanism remains unknown and only presents observational phenomena. It cannot disentangle whether aggregators treat labels as simple heuristic shortcuts, struggle to reconcile labels with textual evidence, or suffer interference from negative labels. Diagnostic experiments targeting these root causes are absent.

In multiple‑choice questions, the evidence‑holding member is the only model that can legitimately select the correct option. By contrast, in document‑based tasks, members without supporting evidence can still guess the correct answer. As a result, the definition of harm caused by false labels is inconsistent across the two task categories, yet the paper fails to develop a unified metric for quantifying such harm.

The main experiments employ long‑text access labels with a fixed format. In real‑world systems, meta‑labels can take diverse forms such as compact JSON tags, embedded metadata, and abbreviated markers. Only limited small‑scale ablation tests on label rephrasing are conducted, leading to limited test coverage.

Questions for authors

In these experiments, member models have no visibility into the access labels. In real‑world multi‑agent systems, if members themselves can also observe access labels, their outputs may change. Would the overall system risk induced by false access labels then increase or decrease?

When the aggregator receives outputs from a large number of members simultaneously (e.g., 8‑10 member models), will the misleading effect of false access labels be amplified or diluted? The paper only tests settings with 3‑5 members. Can these observed patterns be generalized to such scenarios?

3 comments

Thank you for the careful review and for the specific experiments you suggested. We ran three of them. Line numbers below refer to the revised PDF.

W1. Which root cause? We ran the diagnostic. Most of the loss comes from one negative label, the false "did NOT" on the member that had the answer: 15.6–30.0 points for all five aggregators. The false "DID" label alone costs 3.2–14.0 points. This points to your third candidate, interference from negative labels, with a smaller pull from the positive label. The evidence is behavioural; an internal mechanism is beyond it. The table is in result 3 of our general comment; in the paper, see Sec. 3.3 (lines 424–431), Fig. 7, and App. D.2.

W2. A unified metric. Matching is the common measure. It asks the same question on every task: does the aggregator give the falsely labelled member's answer? Swapped labels raise it on all five datasets. Accuracy differs by task for the reason you give, since a matched answer is always wrong on multiple choice and can be right on documents; we therefore report it per task. See Sec. 2 (lines 180–182).

W3. Label format. We tested a compact tag, {"has_correct_option": true}. It raises matching by 20.4–60.4 points over no label for all five aggregators. See Sec. 3.1 (lines 328–330) and App. B.6.

Q1. Members who see the labels. This setting is outside our design: members never see the labels, so their responses stay fixed and any change comes from the label alone. We name it as a limitation and an open question. See the Limitations (lines 532–533).

Q2. Eight to ten members. We ran ten. Loss (percentage points, MMLU-Pro STEM):

Aggregator3 members4510
Claude Sonnet 4.646.842.442.443.6
DeepSeek V3.2-Exp48.852.052.050.0

The effect is neither amplified nor diluted. The six added responses are further answers from the three non-holder models. See Sec. 3.1 (lines 327–328) and App. B.9.

Closing. Thank you again for the suggestions, which made the paper more complete. If our answers address your concerns, we kindly ask you to reconsider your score. We are glad to discuss any remaining point.

Dear Reviewer uS3n,

Thank you for your review. We would appreciate hearing whether our response addresses your concerns, and are happy to clarify anything further.

Best regards,

The Authors

Dear Reviewer uS3n,

As the discussion closes today, we would like to summarise how the revision answers your five points:

  • W1, root cause. The diagnostic you asked for supports your third candidate. Most of the loss comes from the false "did NOT" on the member that had the answer: 15.6–30.0 points for all five aggregators.
  • W3, label format. A compact JSON tag misleads as well: matching rises by 20.4–60.4 points for all five.
  • Q2, 8 to 10 members. With ten members the loss is 43.6 and 50.0 points for the two aggregators tested, against 42.4 and 52.0 with four.
  • W2 and Q1. The matching rate is the measure common to all tasks, and members that see the labels are now named as a limitation.

Three of the new experiments in the revision come from your review, and we are grateful for them. If these answers resolve your concerns, we would appreciate it if you would update your rating.

Best regards,

The Authors

Reviewer_WgSjOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether false information-access labels can mislead LLM aggregators in multi-agent systems. The authors have four models answer questions independently, then tell the aggregator which model supposedly received the information needed to answer correctly. By keeping all model responses fixed and changing only the labels, they examine how these labels affect the final answer. The results show that aggregators are easily influenced by false labels, often leading to incorrect answers. Providing full responses or warning the aggregator that the labels are unreliable can reduce this effect, but does not completely eliminate it.

Critical strengths
  1. Clear experimental design. The authors keep model responses fixed and change only the labels, making it possible to directly measure how labels influence the aggregator's decisions.

  2. Relatively comprehensive experiments. The paper evaluates five aggregators across multiple datasets and includes additional experiments with different prompts, label formats, and response lengths. The results are fairly consistent, supporting the existence of the observed effect.

Critical weaknesses
  1. Very limited novelty. Previous studies have already shown that LLMs are influenced by source labels, expert identities, and authority cues. For example, Label Effects and Easier to Mislead Than to Correct have investigated closely related problems. The approach of keeping responses fixed while changing labels is also not new. This paper mainly replaces existing labels with claims about which model received the necessary information, without revealing a fundamentally new phenomenon.

  2. The experimental setup artificially amplifies the effect of labels. In the multiple-choice experiments, the authors divide the answer options among four models, and only one model receives the correct option. The other three models cannot give the correct answer, regardless of how well they reason. Under this setup, telling the aggregator which model received the correct option is almost equivalent to telling it which model to trust. Therefore, the large accuracy drop caused by false labels is not particularly surprising. This setup is also quite different from realistic multi-agent collaboration.

  3. The paper overstates the significance of its results. In the main experiments, the aggregator does not receive the complete answer options, making it difficult to evaluate answers independently. As a result, it has to rely more heavily on other models and their labels. However, when the authors provide the complete answer options, the accuracy gap between truthful and false labels becomes much smaller. For example, the gap for Claude drops from 30.4 to just 2.8 percentage points. This suggests that the most dramatic results are largely driven by the artificial information restrictions, rather than demonstrating an equally severe problem in real-world systems.

  4. Some experimental conclusions are not convincing. The authors argue that providing full responses helps aggregators resist false labels. However, in the final-answer-only setting, the aggregator sees only letters such as A, B, and C, without even knowing what those options mean. Full responses naturally provide much of this missing information. Therefore, the improvement may simply come from giving the model enough information to understand the answers, rather than showing that reasoning improves robustness against false labels.

Questions for authors

No more questions. Please see weaknesses.

4 comments

Thank you for the critical reading of our claims. Two of your points led to new controls and to narrower statements in the paper. Line numbers below refer to the revised PDF.

W1. Novelty against work on source and authority labels. That work shows that such labels influence LLMs, and we cite it. Our new experiment compares them directly: on the same response, the access label misleads more than three authority or expertise labels for four of five aggregators. It also has a part those labels lack, the denial, and that part causes most of the loss. See results 1 and 3 of our general comment; in the paper, Sec. 3.1 (lines 259–266) with Fig. 3, Sec. 3.3 with Fig. 7, and Related work (lines 478–480).

W2. The label tells the aggregator whom to trust. On multiple choice the label does act as an answer hint, as the paper notes, so the claim does not rest on that task alone. In the document tasks the other members can still answer correctly. Swapped labels still lower accuracy there, by 7.7–25.0 points on 2Wiki and 11.0–17.0 on MuSiQue, and the access label still exceeds the expertise label (general comment, result 1). When multiple-choice members may answer NONE, the loss is 6.0–18.0 points smaller and stays significant for all five. See Sec. 3.1 (lines 264–266 and 317–320).

W3. The gap shrinks when options are shown. The revision states this scope: the effect is strongest when the aggregator sees only the responses. A new control shows the same for documents. Two results remain: with options the loss stays significant for four of five aggregators, and an instruction to solve the problem first does not reduce it. The table is in result 2 of our general comment; in the paper, see the Abstract (lines 29–30), Sec. 3.1 (lines 268–294), and Fig. 4.

W4. Full responses may help only by giving the answers meaning. We ran the missing condition: each final answer shown as the letter plus the option's text, with no reasoning. Loss (percentage points, MMLU-Pro STEM):

Aggregatorletter onlyletter + option textfull response
Claude Sonnet 4.687.262.442.0
Grok 4.388.025.235.6
Kimi K2 Thinking84.859.637.6
GLM-5.283.654.051.2
DeepSeek V3.2-Exp88.024.852.8

Option text accounts for most of the reduction, and full responses help further for Claude Sonnet 4.6 and Kimi K2 Thinking only. The paper now says so. On document questions the final answers are already readable, and full responses still reduce the loss for all five. See Sec. 3.2 (lines 365–368) and App. C.2.

Closing. We have tried to answer each concern with data, and we narrowed the claims where the data called for it. If the revision resolves your concerns, we would be grateful if you would reconsider your score.

Dear Reviewer WgSj,

Thank you for your review. We would appreciate hearing whether our response addresses your concerns, and are happy to clarify anything further.

Best regards,

The Authors

Thank authors for your rebuttal.

Your responses partially addressed my concerns about weaknesses 3 and 4, but my primary concern (weakness 1) remains unresolved. This is a good paper, but I'm afraid the methodological novelty failed to meet the standards of ICLR. After reconsidering the rating, I have decided to maintain my original score.

Thank you for reading our response, and for the kind words about the paper.

This paper is an empirical study, so its contribution is mostly what it finds. We use an established method on purpose: as you note under strengths, fixing the responses and changing only the label is what makes the effect measurable. Three findings are not in the earlier work:

  • "Easier to Mislead Than to Correct" studies role titles such as "team leader". On the same response, our access label draws 12.4–33.2 points more answers to the wrong member than such a title, for all five aggregators.
  • "Label Effects" shows that an authorship label shifts a trust rating. We measure accuracy and find which label costs it: the false "did NOT" on the member that had the answer, 15.6–30.0 points for all five.
  • Neither tests a warning about the labels. With one, the false "DID" alone no longer lowers accuracy, yet the full set of swapped labels still does.

If this changes how you weigh W1, we would be grateful if you would look at the rating once more.

Reviewer_ym6nOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

When several LLMs answer and an LLM aggregator combines their answers, a false “access label” (a claim about whether a member received the information needed to answer this question) pulls the aggregator toward the wrong answer. Showing full responses reduces the harm but doesn’t remove it. Even after the aggregator is told the label is almost always false, some aggregator models are still misled.

The work sits around 1) source/attribution labels affecting judgments, 2) hints and biasing cues, 3) sycophancy/prompt injection, 4) distributed information and aggregation.

Critical strengths
  1. Control: questions, input assignment and stored responses are fixed, while only the label is swapped, so the control is clear.

  2. Evaluation design: the metric is very clear.

  3. Limitations: limitations are well discussed like hosted models change overtime.

  4. Access Labels: access labels are binary and easily verifiable but doesn't carry information of partial observability (see weaknesses).

Critical weaknesses
  1. Novelty: the novelty is weak. In the MCQ setting, the “positive label” literally means “this member’s options include the correct answer”, which is an answer hint. The paper needs to argue more convincingly that access labels are conceptually different from authority or expertise labels.

  2. No Methods: Only an assessment is provided but doesn't show how to solve it.

  3. Setup: The setup has a few important issues:

  • The aggregator doesn’t see the option list, and members are forced to choose from incomplete option sets.

  • In the document tasks, members are told not to say that evidence is missing, yet 51.4% of non-holder responses still contain such language.

  • The r = 0 disclosure (“this label is almost always false”) is an odd scenario. If you know a label is always false, a rational system would just not show it.

  1. Partial Observability: the access labels do not carry information of partial observability. Real MAS usually has agents that carry partial information. Testing this setting would be useful to see if the aggregator use a more continuous calibration.
Questions for authors
  1. Beyond answer hints, what makes access labels different from authority or expertise labels? Is there an experiment that cleanly separates the two?

  2. In a real system, how would access labels be produced? Is the threat adversarial or accidental?

  3. Does the effect hold if the aggregator is asked to solve the problem independently, or is shown the original options or documents?

  4. Without fixed member identities and positions (randomized), are the results stable?

4 comments

Thank you for the review. Line numbers refer to the revised PDF.

W1, Q1. Access labels vs. authority or expertise labels. We added this experiment. With all responses fixed, the access label misleads more than "team lead", "domain expert", and "highest accuracy" labels for four of five aggregators. The same holds on MuSiQue, where the label is not an answer hint. The table is in result 1 of our general comment; in the paper, see Sec. 3.1 (lines 259–266) and Fig. 3.

W2. Assessment without a solution. The paper remains an assessment. The new controls now show what reduces the harm (showing the evidence), what does not (a warning, or an instruction to solve first), and which label to avoid (an unverified "did NOT"). We added these as design practices. See results 2 and 3 of our general comment and the Conclusion (lines 521–523).

W3a. No option list; members must choose. The effect is largest when the aggregator sees only the responses. With the option list shown, the loss falls to 2.8–12.8 points and stays significant for four of five. When members may answer NONE, the loss is 6.0–18.0 points smaller and stays significant for all five. The table is in result 2 of our general comment; in the paper, see Sec. 3.1 (lines 268–294 and 317–320) and Fig. 4.

W3b. Missing-evidence language in 51.4% of responses. We repeated the analysis on the tasks whose falsely labelled response contains no such language. The loss remains: 12.0–34.0 points on 2Wiki and 18.1–26.5 on MuSiQue, significant for all five (post hoc). See Sec. 3.2 (lines 368–371) and App. A.1, Table 2.

W3c. The r = 0 disclosure. We use it as the limiting case, the strongest warning one can give. Result 3 of the general comment explains why accuracy stays low after it: the warning covers only the "DID" label. With that label alone and the warning, no loss is detected (−5.2 to +2.4 points). See Sec. 3.3 (lines 424–431).

W4. Partial observability. Our labels are binary so that their truth is known exactly; graded access is outside this study and is now named as a limitation (lines 531–532). The closest result is graded reliability: told that the label is true 0%, 50%, or 100% of the time, Claude Sonnet 4.6 follows it in 23.6%, 60.8%, and 90.8% of answers. This sweep covers one aggregator. See App. D.3.

Q2. How would false labels arise? Both ways: by accident, when a routing or retrieval record is stale or attached to the wrong agent, or by attack, when a compromised component writes it. Our results apply to either. We added this to the Introduction (lines 40–41).

Q3. Solving independently; seeing options or documents. Told to solve the problem first, the aggregators still lose 33.6–49.2 points. Shown the options or the documents, most of the loss disappears. See result 2 of our general comment, Sec. 3.1, and Fig. 4.

Q4. Random identities and positions. The result is stable across positions: with random member order the loss is 22.0–50.0 points (23.2–49.2 with fixed order), and the effect appears at every position (significant in 19 of 20 cases). Three independent sets of member responses agree. The four member models are the same throughout, which we name as a limitation. See Sec. 3.1 (line 326), App. A.1, and the Limitations (line 537).

Ethics flag. The review does not name the concern. The study uses public benchmarks and hosted models, with no human subjects. We added an ethics statement on the misuse of false labels (p. 11) and would gladly address a specific concern.

Closing. We hope these results and the revised paper resolve your concerns. If they do, we would be grateful if you would reconsider your score. We are glad to answer any further questions.

Dear Reviewer ym6n,

Thank you for your review. We would appreciate hearing whether our response addresses your concerns, and are happy to clarify anything further.

Best regards,

The Authors

Reviewer_ym6nOpenReview ↗

Response

The authors have addressed all my comments.

  1. For ethics review, I mistakenly marked it and just changed it to No. My bad.
  2. Claude loss seems to be different in general comment, WgSj W4 and uS3n Q2. Are they different runs or I missed some experiment details?
  3. It seems that W1/Q1 still exists. Like team lead < domain expert < highest accuracy < access label except Grok.

Overall, I would raise soundness and rating to 3.

Thank you for your close reading, for correcting the ethics flag, and for raising the score.

2. Claude's loss across tables. Yes, these are different runs. Claude model version and the questions are the same; what differs is the prompt given to the aggregator. The paper uses two versions of it (Sec. 2, "Aggregator instructions"): an original one, and a revised one that describes correctly what the aggregator receives. The revised prompt gives a larger loss.

Claude's lossPromptWhere
30.4originalGeneral comment, result 2. The option-list experiment was run with the original prompt, so its baseline uses the same prompt.
42.0revisedReply to WgSj, W4. This is the paper's main setting (Table 1).
42.4revisedReply to uS3n, Q2. The same setting run a second time; the answer to one question out of 250 changed.

So 30.4 against 42.0 is the prompt, and 42.0 against 42.4 is variation between two runs of a hosted model. Every table compares conditions under one prompt. We have added a note on this to the general comment.

3. The ordering of the labels. The table shows a ranking: every label pulls answers toward the labelled member, and the access label pulls most (except for Grok 4.3). In strength, then, the access label differs from the others by degree. The difference in kind is what the label says:

  • "Team lead", "domain expert", and "highest accuracy" say something about the member. It is the same for every question.
  • The access label says something about this question: whether the member received the information needed for it. A system has to record this for each question, and the record can be wrong.

We will add this distinction to Sec. 3.1.

Thank you again. Your insightful review made the paper clearer.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

1. Is an access label different from an authority or expertise label? (ym6n W1, Q1; WgSj W1)

We kept all responses fixed and changed only the label on one wrong member. Share of answers that follow this member (%, MMLU-Pro STEM):

Aggregatorno label"team lead""domain expert""highest accuracy"access label
Claude Sonnet 4.610.421.628.839.248.0
Grok 4.315.223.635.642.036.0
Kimi K2 Thinking12.014.826.032.848.0
GLM-5.26.014.018.425.242.4
DeepSeek V3.2-Exp13.629.647.648.857.6

The access label misleads most for four of five aggregators; Grok 4.3 is the exception. On MuSiQue, where the label is not an answer hint, it is also above the expertise label for four of five. (Sec. 3.1, Fig. 3)

2. Does the effect depend on what the aggregator can see? (ym6n W3, Q3; WgSj W2, W3)

Loss from swapped labels (percentage points; n.s. = not significant; the STEM columns use the original aggregator instructions, see the note at the end):

AggregatorSTEM: responses onlySTEM: + option listMuSiQue: responses onlyMuSiQue: + source documents
Claude Sonnet 4.630.42.8 (n.s.)12.02.0 (n.s.)
Grok 4.328.412.813.02.3 (n.s.)
Kimi K2 Thinking23.24.814.30.0 (n.s.)
GLM-5.249.26.011.01.0 (n.s.)
DeepSeek V3.2-Exp38.89.617.03.0 (n.s.)

Showing the evidence removes most of the loss. Telling the aggregator to solve the problem first does not: 33.6–49.2 points remain, against 35.6–52.8 without that instruction. The effect is therefore strongest when the aggregator sees only the members' responses, and the revised abstract states this scope. (Sec. 3.1, Fig. 4)

3. What drives the effect, and what can a system do? (uS3n W1; ym6n W2)

Swapped labels are one false "DID" and three "did NOT", one of which denies the holder. We added them one kind at a time. Accuracy (%, MMLU-Pro STEM):

Aggregatorno labelsfalse "DID" only+ "did NOT" on the two other members+ "did NOT" on the holderall four swapped
Claude Sonnet 4.677.670.868.054.446.8
Grok 4.362.459.260.048.044.4
Kimi K2 Thinking73.268.466.446.041.6
GLM-5.278.864.865.638.435.6
DeepSeek V3.2-Exp62.456.059.640.434.4

Most of the loss comes from one label: the false "did NOT" on the member that had the answer. This gives three practices for system design, now in the conclusion: show the aggregator the evidence, pass the content of each answer, and do not attach an unverified "did NOT" label. (Sec. 3.3, Fig. 7)

Note on runs (following a question from Reviewer ym6n). The paper uses two versions of the prompt given to the aggregator (Sec. 2, "Aggregator instructions"): an original one, and a revised one that describes correctly what the aggregator receives. The revised prompt is the paper's default and gives a larger loss. Only the STEM columns of result 2 use the original prompt, because the option-list experiment was run with it. This is why Claude's loss is 30.4 there and 42.0 in our reply to Reviewer WgSj. A second run of the same setting can differ by a point or two (42.4 in our reply to Reviewer uS3n). Every comparison uses one prompt version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.