acceptodds
Withdrawn from ICLR 2027

Does Human Group Psychology Transfer to Agent Collectives?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A team's emotional state bears on its output: one member's mood travels to the others, and the resulting composition predicts what the team produces. Collectives of language model agents now occupy much the same arrangement, and practitioners already configure their affect by assigning roles or by instructing agents to be skeptical. Whether the human regularities govern such collectives is nonetheless unestablished, since an instruction to feel anxious admits no graded intensity, while expressed text discloses only what an agent chooses to express. We therefore set each agent's state by intervening on its internal representations at calibrated intensities, and read every member's state from its activations on each round of deliberation. Affect does propagate, in proportion to the intensity applied, although the resulting shift is absent from the text the recipients produce and consequently invisible to a judge that reads it. What propagates, however, is frequently not what was injected: the targeted emotion leads the recipients' affective profile only within a bounded range of intervention strengths, and that range narrows as models grow. No configuration we examined improved group accuracy on any benchmark, whereas every emotionally configured group proved markedly more willing to adopt a confidently asserted wrong answer that unsteered groups disregard entirely. Affect in these collectives is accordingly a liability to be monitored rather than a lever to be pulled.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Agent, Emotional Dynamics

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_x3nYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates if the findings identified from the research on human teams' composition, interaction, and performance can be reused also to understand and govern the collective behaviors of groups of language models. The authors found that affect propagates within groups of language models, and it does in proportion to the applied intensity. Interestingly, the resulting shift is absent from the textual content the recipients produce and thus it's invisible to a judge that verifies. Again, steered language models does not improve the group accuracy on a significant amount of benchmarks, while every emotionally steered group proved more willing to adopt a wrong answer when it was asserted in a confident way.

Critical strengths
  • The idea of studying if and how findings characterizing human groups transfer to the collective behavior of language models is of of great interest and of paramount relevance given the growth in the adoption of multi-agent systems in real-world tasks.

  • The authors place their research hypotheses on several works and findings in group and social psychology, although sometimes they accept some claims from these fields as evident while are more controversial. For example, not always the emotions spread following a contagion effect. Some studies, indeed, have shown that emotions can spread following a mimicry strategy vs. a repulsion strategy: individuals can thus in some situations adapt to the emotions of their peers and in other situations they can react in the opposite way.

  • Some findings are interesting, for example the fact that affect seems to propagate covertly and that emotional steering seems to not increase accuracy in solving downstream tasks.

Critical weaknesses
  • Several findings in group psychology literature are taken for granted and treated more as axioms than as hypotheses or previous findings with some supporting evidence (see the example cited above of contagion vs. repulsion effects).

  • In pages 4 and 5 the formalizations of these group psychology findings are often not well defined. A lot of variables in the mathematical formulation are left completely undefined. This makes difficult to properly understand and evaluate these formalizations. See for example Equation 1: it seems just taken from opinion dynamics literature without any explanation (or adaptation) of the terms.

  • Several parts of the paper are written with a difficult to understand writing style. In particular, several paragraphs in the Introduction and in other parts of the paper are very long, unclear and require a rephrasing (see for example, the paragraph starting with "Because all 25 dimensions ..." on page 2, and the one starting with "Neither result translates ..." again on page 2).

Questions for authors

A couple of suggestions are to:

  • improve mathematical presentation;
  • simplifying and clarifying significantly the writing;
  • evaluating the transcripts not just using LLMs as judges but also human evaluators.
Reviewer_en8hOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Whether four psychology of human groups regularities (contagion: P1, affective composition: P2, arousal and task difficulty: P3, susceptibility to a confident member: P4) hold in LLM multi agent groups is the main question of the paper. Affect is set by adding mean-difference emotion directions (i.e., representation engineering) to the residual stream, calibrate intensity per emtion, and read agents' "emotional state" from activations after every round. Four agents interact round-robin fashion for five rounds on datasets such as MATH500 or HumanEval. Main results they claim are:

  • Steering one agent shifts its teammates' activation, even though an LLM judge rates the steered agent's messages as neutral.
  • The emotion that spreads is often not the one that was injected.
  • No clear relationship between emotional configuration and task performance.
  • Steered groups are much more likely to go along with a planted wrong answer from an 'artificially' confident agent.
Critical strengths

The experiment seems careful and report looks honest, very thorough, technically detailed (way beyond typical human written papers). The experiment calibrates intensity, compares against the full emotion profile, reports how strong the representing engineering turned out to perform, and states what didnt work such as Gemma setup and diversity manipulation.

Critical weaknesses
  1. Writing

Honestly, this paper is extremely, extremely hard to understand, because of convoluted super long sentences which human will have significant difficulty in understanding. I am highly confident that the paper is AI-written with (almost) no human editing. For example, the first sentence of Section 3, which is the beginning of the paper after Introduction and Related Work section, starts with:

"The source literature states these claims verbally, and each admits more than one operationalization, so we state them formally."

This is not how human paper writers begin their paper's very first part of the main section without proper and sufficient explanation of the problem context. My concern is not limited to this specific line; I would strongly recommend authors to do testing by asking someone to read a portion of the paper (e.g., Section 5.7: Checks carried out) and evaluate whether it sounds human written.

  1. One concern about the experimental design is that the paper don't show these effects come from emotion in particular. There is no control baseline which steers agents with a random vector direction, or with a non-emotional concept. Without this control, it is hard to tell whether steered agents (models) adopt the wrong answer more often because of the stated affect, or simply because pushing on the residual stream makes a model reason a bit worse compared to the no-intervention (i.e., original model weight).
Questions for authors

Authors noted that "We have not used generative AI tools for research ideation or experimental design". I would like to ask a clarifying question, whether no coding agents / models are involved in the execution of the experiments and all are done by human authors.

Reviewer_y7PdOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper translates four claims from human group psychology—emotional contagion, affective composition, arousal and task difficulty, and susceptibility to a confident wrong member—into testable quantities. Using residual-stream steering and unsteered forward-pass probes, the authors measure internal representations, answer rates, accuracy, and adoption of a planted wrong answer in four-agent multi-round deliberation. The main finding is that representational changes propagate but do not reliably improve accuracy, and some emotion configurations markedly increase adoption of the wrong answer.

Critical strengths

1.The problem definition and evidence chain are careful: 25 probe dimensions, per-emotion calibrated injection strength, paired problems and random seeds, 500 paired trials, and factorization of answer production from accuracy-given-an-answer, which avoids conflating non-response with capability decline (Sections 4–5, Tables 1–2). The unsupported inverted-U and composition-diversity hypotheses are reported as failed manipulation checks rather than forced into interpretation. 2.The safety-relevant conclusions are subjected to strong restrictive tests: the scripted wrong member is excluded from the majority vote, and adoption of the wrong answer survives excluding off-task messages and incomplete answers.

Critical weaknesses

1.The construct of "internal emotion" is still intervened on and measured largely with directions extracted from the same emotion narratives, creating a circular validation risk between probe and manipulation. The authors use unsteered forward passes, orthogonal readouts, and a second text discriminator to weaken the simple echo explanation, but the lexical classifier still reaches AUC 0.84, indicating non-obvious textual differences exist. The current evidence more safely supports "representational and linguistic signal transmission" rather than a strict transfer of human psychological mechanisms (Section 4, Section 5.2, Appendix G). 2.The causal test of composition-to-performance did not hold: configured diversity correlates negatively with realized final-round diversity (r = −0.33), so P2 cannot be tested as pre-registered. Models, the four-agent five-round protocol, and task types limit extrapolation; the blanket wording that "all configurations increase wrong-answer adoption" should be qualified to the configurations/models actually tested (Section 5.6).

Questions for authors

1.Can you reproduce the dose–response and "target emotion identity" effects using an independently trained emotion readout or a blind-labeled behavioral task, while controlling for message length, task engagement, and lexical-classifier-detectable cues? 2.Please provide the complete pre-registration and deviation record, and list paired sample sizes and multiple-comparison handling per configuration for the adversarial experiment. If realized group diversity were actually manipulated, would the P2 conclusion change?

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

Related papers

Discussion (0)

Sign in to comment.