acceptodds
Under review as a conference paper at ICLR 2027

From System Symptom to Agent-History Diagnosis: A Three-Coordinate Protocol for LLM Population Drift

Abstract

Behavioural change in multi-agent LLM populations may reflect shared time, heterogeneous agent histories, or visible interaction evidence. We introduce DriftClock, a three-coordinate protocol separating shared time, private exposure, and visible history. In naming games, round and exposure correlate at ; exact same-round contrasts instead localize the signal to agent histories: switching after disagreement declines with exposure, future exposure is null, a 64-population Mistral study replicates exposure-indexed dynamics, and long-horizon decline transfers across models and objectives. Matched memory-window interventions reveal a finite-context breakpoint: response falls as histories fill and partially rebounds when old evidence leaves longer windows. Discovered in Qwen, this breakpoint passes a frozen Mistral test and an independent-seed Qwen replication; social-update rules separately show that exposure signatures need not imply autonomous ageing. Standardized challenges preserve eventual adaptation, while post-hoc common-support analyses locate the raw speed gap in measured visible evidence. Paired history branches separate local source effects from propagation: history differences reach other agents while action differences remain sparse and model-dependent. DriftClock moves from population symptom to history localization, formation, functional consequence, and transmission, establishing finite-history turnover as a reproducible boundary condition on cumulative-age explanations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent LLMs, behavioural drift, agent-history diagnosis, finite-history turnover, in-context adaptation, emergent coordination, source localization, evaluation and monitoring

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

A multi-agent logging mechanism that acknowledges wall-clock as well as exposure-time for agents, allowing for analysis of drift and other multi-agent behavior as it unfolds.

Critical strengths

The proposed approach seems useful.

Critical weaknesses

The approach seems trivial and obvious and contributions are not clear. The language is hard to follow and at times obviously AI-authored, reducing its clarity.

Questions for authors

I did not understand the significance of the paper and the motivation behind the multitude of experiments in the body and appendix.

Reviewer_sMZ3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DriftClock, a diagnostic protocol for separating behavioral drift in multi-agent LLM populations into three independent coordinates: shared time (t) for synchronized environmental changes, private exposure (e) for heterogeneous agent updates, and visible history (z) for observable agent-agent interactions. The authors validate their framework on naming games and demonstrate that this decomposition enables precise localization of drift sources to specific agent histories. The work addresses an important but underexplored problem: understanding where and why multi-agent LLM systems behave differently over time.

Critical strengths
  1. Well-motivated problem: Behavioral drift in multi-agent LLM systems is an important and under-explored problem. The paper identifies a genuine gap: drift reflects multiple independent processes (shared updates, heterogeneous exposure, interaction history).

  2. Clever decomposition: The three-coordinate protocol (t, e, z) is non-obvious and provides a principled way to attribute behavioral changes to specific causes. The separation of shared time, private exposure, and visible history is mathematically well-founded and could generalize beyond LLMs.

  3. Clear validation on controlled setting: Naming games experiments are well-designed and convincingly demonstrate utility. The ability to recover ground-truth drift sources from simulated populations is compelling.

Critical weaknesses
  1. Limited validation scope: Main validation is on synthetic naming games with known ground-truth sources. No validation on realistic multi-agent LLM systems (e.g., web agents, dialogue systems). Unclear whether DriftClock scales to real environments or fails in complex settings.

  2. Scalability not addressed: Paper does not discuss computational complexity, wall-clock time, memory requirements, or how method scales with population size and history length. These are critical for practical deployment.

  3. Incomplete agent heterogeneity treatment: Framework assumes (t, e, z) characterization is sufficient, but ignores intrinsic agent differences (parameter variance, architecture differences). Limited discussion of how to handle private unobserved state beyond visible history z.

  4. Weak theoretical foundation: Lacks formal guarantees on identifiability of t, e, z. When is decomposition uniquely recoverable? Are there drift sources outside the three categories? Sensitivity to misspecification not analyzed.

Questions for authors
  1. Can you validate DriftClock on realistic multi-agent LLM tasks beyond naming games (e.g., web-based collaboration, dialogue systems, instruction following)?

  2. What is the computational complexity of DriftClock? How does inference time scale with population size N, history length L, and number of agents A? Please provide wall-clock timing.

  3. How would you extend DriftClock to handle agents with intrinsic parameter variation (different model sizes, architectures) that are independent of t, e, z?

  4. In practice without ground truth, how would you measure t, e, z? What are minimal data requirements (number of agents, history length, timesteps)?

  5. How robust is drift diagnosis to: (i) misspecified history length, (ii) incorrect agent models, (iii) non-orthogonal coordinates? Please provide sensitivity analysis.

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

This paper studies behavioral change in multi-agent LLM populations and presents DriftClock, a three-coordinate protocol separating shared time, private exposure, and visible history. Experiments apply DriftClock in two regimes, with three LLMs (Qwen2.5-7B, Mistral-7B, and Yi-6B) evaluated.

The paper is very hard to follow. AI-generated content piles up and has to be cleaned to improve the presentation quality.

Critical strengths

S1. Monitoring behavioral change in multi-agent LLM populations is an interesting research direction.

S2. Five linked acts are evaluated.

S3. The evaluated models cover three LLM families.

Critical weaknesses

W1. The paper is very hard to read. I'm not against AI-generated papers, but I doubt the authors have carefully read the generated content. For example, the abstract is basically a report of experimental results. Other indicators of AI-generated content include excessive use of semicolons and italic scripts, unnecessary defensive writing (e.g., "not a causal or equivalence test", "must not be read as ..."), as well as many SHA-256 checksums (which should appear in the repo or supplementary file instead).

W2. The context of the studied problem is unclear. It is suggested this is elaborated in the beginning of the abstract or the introduction.

W3. Only 6-7B models were evaluated in the experiments.

W4. The discussion of the results is too brief.

Questions for authors

Q1. I do not have any specific question. A refinement of the AI-generated content is necessary to make the paper easier to read.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.
From System Symptom to Agent-History Diagnosis: A Three-Coordinate Protocol for LLM Population Drift | acceptodds