acceptodds
Under review as a conference paper at ICLR 2027

Out of Character: Per-Agent Conformal Screening for Byzantine LLM Committees

Abstract

Committees of language-model agents rely on redundancy to catch reasoning errors, but correlated exploits such as prompt injections can subvert multiple models simultaneously; when corrupted agents gain a majority, standard majority voting collapses from an uncorrupted accuracy of to . Without labels, no confidence statistic of a single decision exposes an adversary that picks a wrong action while preserving the agent's confidence profile, and peer-consensus defences fail because the adversary forms the majority. We resolve this dilemma with Sentinel, a peer-free screen that tests each agent's confidence and action choices over a window against its own clean history, bounding the probability of falsely rejecting an honest agent by a budget in finite samples under exchangeability, regardless of how many peers are corrupted. Across 5 models and 3 benchmarks, Sentinel recovers accuracy at 60% corruption and at 80% within the screened window; on free-form answers under real prompt injection it regains – of the lost accuracy; and it restores live task completion from to of tasks under a destructive attack; live under the exact null, honest rejection is – per screen with no attack, against a budget. We further prove that an adversary preserving its own action law can change the action symbol on only a bounded fraction of decisions — a bound on symbol alteration, not on task failure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-agent LLMs, Byzantine robustness, LLM safety, conformal prediction, adversarial agents, robust aggregation

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_smbVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies LLM committees under a coordinated Byzantine majority. Sentinel screens each agent against its clean calibration history using a Kolmogorov-Smirnov distance on confidence complements and total variation distance on selected answer symbols, then conformally trust-weights survivors. Theorem 2 controls honest-agent rejection under exchangeability, independently of the number of corrupted peers; Proposition 1 and Theorem 3 characterize single-decision indistinguishability and the action-law-preserving attacker that marginal screens cannot detect. Experiments cover five LLMs, five-agent committees, MMLU, TruthfulQA, ARC-Challenge, GPQA-Diamond, ALFWorld, Mind2Web, and τ-bench, with 20 seeds in the principal suite plus live and free-form prompt-injection evaluations. At three corrupt agents of five, accuracy is 0.682 versus 0.094 for majority vote; on injected GSM8K answers, Sentinel recovers 73 to 76 percent of lost accuracy.

Critical strengths
  1. The paper targets a clear failure mode: a correlated majority can subvert peer consensus and standard robust aggregators. The two-channel self-history screen is a technically motivated response.

  2. The formal claims are carefully scoped: Proposition 1 establishes the one-decision limit, while Theorem 3 gives a concrete blind spot and bounds symbol alteration without implying task-level security.

  3. The evaluation spans five model families, four multiple-choice benchmarks, two agentic domains, live τ-bench, and free-form GSM8K injection. Table 1 reports 20 seeds, intervals, paired tests, and baselines; 0.682 mean accuracy substantially exceeds the 0.094 majority-vote result.

  4. Table 5 shows decision-level resampling breaks the budget on dependent agentic traces, while episode-level resampling restores it. The appendices also document failure cases and calibration alternatives.

Critical weaknesses
  1. The guarantee requires honest test windows to remain exchangeable with clean calibration. Corrupted agents can shift honest members' context: Table 24 reports 0.242 to 0.280 honest rejection in attacked live retail sessions, above the 0.10 budget. Warm-up calibration in Table 25 helps, but the guarantee does not cover these shifts.

  2. An attacker preserving its action marginal remains a substantive blind spot. Theorem 3 allows every selected symbol to change when no symbol exceeds half the clean marginal; Table 8 reports 0.305 accuracy against 0.705 clean for a couple-law coupling. This disclosed limit narrows robustness to adaptive harmful agents.

  3. Although BFT is cited in related work, it is not compared experimentally, leaving the advantage over current confidence-based BFT defenses unclear.

  4. Live task utility remains modest: Table 6 reports 4/30 retail and 3/30 airline completions under destructive attack, versus 3/30 clean in each domain. Table 24 reports 2/25 screened retail completions for the distinct-model committee. Security improves more clearly than end-task success; code and cached outputs are promised only upon acceptance.

Questions for authors
  1. How does attack-induced context shift affect exchangeability and honest-agent rejection after calibration on representative attacked trajectories?
  2. Can you evaluate CP-WBFT under the paper's fixed-shape, permutation-mimic, and prompt-injection attacks using matched models and task splits?
  3. Can you report live τ-bench uncertainty intervals and compare completion against an oracle screen over more episodes?
Reviewer_dvdwOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Sentinel, a framework for detecting compromised agents in multi-agent LLM systems, particularly when attackers control the majority of agents. The method compares each agent's confidence and action distributions against its own historical behavior using conformal screening, without relying on peer agreement or test-time labels. The paper provides theoretical guarantees for false rejection under exchangeability and evaluates the method across multiple models, benchmarks, and attack settings. The results show improved robustness against coordinated attacks compared with existing aggregation and detection approaches.

Critical strengths
  • The paper addresses an important challenge in multi-agent LLM security, where traditional voting and consensus methods can fail when most agents are compromised. The proposed per-agent screening approach offers an interesting alternative that does not depend on an honest majority.
  • The paper provides theoretical analysis and extensive experiments across different models, benchmarks, and attack strategies. The combination of finite-sample guarantees, adaptive attack evaluations, and live-agent experiments provides meaningful support for the proposed approach.
Critical weaknesses
  • The theoretical guarantees depend on clean calibration data and exchangeability between calibration and deployment. These assumptions may be difficult to maintain in real-world settings, especially when agent behavior changes or attacks alter the environment. The paper reports cases where false rejection exceeds the intended budget, raising concerns about practical reliability.
  • The live-agent evaluation remains limited, with relatively few tasks and low overall completion rates. In addition, the method requires a window of decisions before screening, leaving the system vulnerable to early or short-lived attacks. These limitations make it difficult to assess its effectiveness in more realistic deployments.
Questions for authors
  • How would Sentinel maintain reliable screening when agent behavior changes over time or clean calibration data are unavailable?
  • Can the authors provide stronger evidence of effectiveness in live multi-agent systems with higher baseline task-completion rates?
  • Can the authors evaluate the proposed anytime-valid extension and clarify how effectively it detects early or short-lived attacks?
Reviewer_Vn5iOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It appears is about trying to catch (or defend?) a fraction of adversarial LLMs, among a committee of LLMs that are trying a to solve a decision problem. That is the extent I can understand from this paper, given the heavily unsupervised AI written text.

Critical strengths

unable to tell.

Critical weaknesses

The abstract and intro are not good. Look at the first sentence of the paper in the abstract for instance:

"Committees of language-model agents rely on redundancy to catch reasoning errors, but correlated exploits such as prompt injections can subvert multiple models simultaneously; when corrupted agents gain a majority, standard majority voting collapses from an uncorrupted accuracy of 0.705 to 0.094."

A very long sentence, which looses its meaning and context as it progresses (if it had any to begin with!), and ends in two random numbers out of the blue:) what are those numbers? are they from a specific experiment ran on a particular dataset?

This is a prime example using AI, without reading the output properly. Ironically, this is the very first sentence of the paper, which means how little the authors cared about the quality of the text in this manuscript.

Then it goes:

"Without labels, no confidence statistic of a single decision exposes an adversary that picks a wrong action while preserving the agent’s confidence profile, and peer-consensus defenses fail because the adversary forms the majority."

It just keep getting worse.

What is "labels"?

"no confidence statistic of a single decision exposes an adversary..." I can not understand a single word!!

The abstract is supposed to communicate what is the problem that is studied in this paper in the first few sentences. What is the problem? I cant tell. what I understand is this: It involves a committee of LLm agents and there are "some adversaries" "somewhere" doing "something" and we dont have "some labels" which creates a dilemma...

This paper needs a heavy "human" revision on the authors side. I refrain to be the first "human" to rigorously read this paper fully, as that is not the job of a reviewer in the peer review system.

Questions for authors

none.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.