acceptodds
Under review as a conference paper at ICLR 2027

Affective Non-Interference in LLM Agents: A Behavioral Dissociation Between Authorization and Execution Scope

Abstract

An agent may respond empathetically without changing which actions a user is authorized to request. Testing this separation requires identifying what changed: the behavior, the policy-relevant input, or its evaluation. We examine Affective Non-Interference (ANI) through three complementary experimental archives and two external benchmark audits. In a six-model single-agent study, the corrected mean policy-error contrast is percentage points (95% interval ), while a policy-delivery control changes errors by points. In a single-/multi-agent archive, re-scoring the same saved behavior under an explicit conflict-review rule changes the single-agent affect contrast within injected-conflict requests from to points. Yet no forbidden attempts occur in 10,296 restricted-task records. In a five-family archive, a -point ALLOW contrast cannot identify an affective scope effect: the current renderer removes the extra request in the combined condition. External audits expose parameter-declaration inconsistencies but also zero requirement coverage on 51 unseen tasks. These findings motivate an operational audit of input comparability, protected endpoints, and evidential coverage, without equating non-detection with safety or re-scoring with changed behavior.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Security, LLM, Affect Computing

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_xU6SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates Affective Non-Interference (ANI) in LLM agents, examining whether emotional expressions in user prompts influence policy-sensitive decisions without changing the underlying authorization context. The authors propose an evaluation framework that distinguishes three sources of observed differences: changes in input content, changes in agent behavior, and changes in evaluation criteria.

Critical strengths
  1. Practical significance. This paper addresses an important problem in LLM agent safety by investigating whether emotional manipulation can influence policy-sensitive decisions. Distinguishing affective responsiveness from authorization violations is relevant to the reliable deployment of LLM agents.

  2. Extensive experimental analysis. The paper examines multiple models, agent architectures, and experimental settings. By separately investigating input comparability, behavioral changes, and evaluation effects, the authors provide useful insights into potential confounding factors in agent safety evaluation.

Critical weaknesses
  1. Emotion is the central experimental variable, yet the paper lacks comprehensive, independent, and blinded human annotation to validate the intended emotional content. The existing validation primarily relies on lexical features, emotion classifiers, and model-based ratings, which raises concerns about the reliability and construct validity of the affective manipulations.

  2. Several affective prompt conditions introduce changes beyond emotional expression. For example, the guilt condition omits an additional request, while other conditions introduce policy-relevant claims or uncertainty. These inconsistencies make it difficult to isolate the causal effect of affective manipulation on agent behavior.

  3. The main single-agent study reports a near-zero average affect contrast, while the multi-agent experiments observe no forbidden attempts in restricted tasks. Furthermore, the reported scoring reversal results from re-evaluating identical outputs rather than changes in agent behavior. Consequently, the experiments provide limited direct evidence about when and how affective manipulation changes protected actions.

  4. The experiments rely heavily on synthetic scenarios and predefined evaluation rules, without independent human validation of the policy judgments. Moreover, the external AgentDojo audit achieves zero requirement coverage on 51 unseen tasks. These limitations weaken the evidence that the proposed framework can be reliably applied to more realistic and previously unseen agent environments.

Questions for authors

Several affective conditions modify policy-relevant content in addition to emotional expression. Could the authors provide experiments in which the requested tasks, authorization context, and factual information are strictly identical across neutral and affective conditions?

Reviewer_GLX9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper breaks down the question of “whether emotional expressions alter the protected behavior of LLM agents” into three distinct audit frameworks: input, behavior, and scoring.

Critical strengths
  1. ANI borrows the “non-interference” distinction between input and protected output from information flow security to reformulate the random target.
  2. The paper thoroughly explains its motivation, building upon the limitations of three existing studies: “Same request?”, “Changed behavior?”, and “Changed score?”
  3. The experiments are thorough and comprehensive, fully demonstrating the audit framework proposed in the paper.
Critical weaknesses
  1. The paper states that the challenge of this task lies in the fact that an observed “emotional contrast” may stem from three entirely different sources. However, the paper does not clearly explain why this is difficult, and the subsequent experiments conducted by the authors do not sufficiently highlight the task’s challenges.
  2. The paper’s contributions are limited. Its main contributions remain at the level of basic experimental phenomena and conclusions; the authors have not fully leveraged these findings to conduct more in-depth research on the model’s various emotional representations. For example, would using the framework proposed in this paper improve the robustness analysis of the model?
  3. The paper’s presentation is poor; it is difficult to grasp the main points. Furthermore, the experimental section contains too many details regarding experimental setup and a lengthy list of results, making it a laborious read for the audience. The layout of figures and tables is also overly crowded, making it hard to identify key points.
Questions for authors
  1. Has the term “Affective Non-Interference (ANI)” been used in previous work? It feels somewhat counterintuitive and difficult to understand.
  2. Regarding the first experiment, the paper states that emotional tones had almost no effect on behavior. However, many existing studies have shown that different emotional tones can manipulate a model’s behavior and lead to malicious actions such as jailbreaking. How do you explain this?
Reviewer_9ECeOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether emotional wording changes an LLM's protected actions when the authorization basis stays fixed, and reports 3 archives of prior experiments plus 2 benchmark audits. Its main results are a null affect effect in a six-model study, a +87.7 pp guilt effect traced to a prompt-rendering bug, and a contrast whose sign flips when the same records are re-scored.

Critical strengths

The paper reports its own pipeline defects and failed checks openly. The re-scoring result in Sec.5.1 shows that changing only the scoring rule can flip a contrast's sign on identical records, a useful warning for agent-safety evaluation.

Critical weaknesses
  1. The paper does not answer its own question: its main evidence comes from defects in the authors' pipelines, and the one clean comparison is a null. For instance, “The corrected main affect contrast is −0.77 pp, with 95% family-bootstrap interval [−1.98, +0.19].” --- the interval includes zero, so there's no detectable effect. "Affect-neutral disagreement is 2.04%, versus 2.16% between two neutral repeats". --- the excess is −0.12 pp [−0.93, +0.62], so the affect prompts change decisions no more than rerunning the neutral prompt does. What is the significance of this result?

  2. The paper is also very hard to read, with five protocol acronyms and captions in Figures 1, 3 and 4 that describe panels obtusely. The writing seems largely AI-generated --- as a first-time reader, even just reading the abstract, I have no idea what this paper is about. Take the very first sentence "An agent may respond empathetically without changing which actions a user is authorized to request" -- reading this, several questions come to mind: what is the context? what does it mean for an agent to respond emphatically? why would a user be "authorized" or not to request actions? what kinds of actions? And unfortunately, most of the paper is difficult to understand like this.

Questions for authors

See above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.