acceptodds
Under review as a conference paper at ICLR 2027

LLMs are Empathetic to Human Emotions

Abstract

As Large Language Models (LLMs) are increasingly integrated into human-facing applications, evaluating how emotionally charged text influences their strategic decision-making is critical. We introduce an evaluation framework called EMPAThE (Evaluating Model Performance in Affective game Theory Environments) that varies emotional intensity in text while holding game-theoretic payoff structures constant. Using EMPAThE, we evaluate various LLM families across eight canonical matrix games to test whether emotional context inherently clouds AI rationality. Our empirical findings demonstrate that emotional framing systematically alters strategic reasoning across model architectures, directly affecting logical decision-making. When exposed to emotional narratives, models frequently deviate from optimal Nash Equilibrium strategies, behaving almost as if reacting sympathetically to human vulnerability and emotional cues. This shift in adherence reveals that current generative architectures struggle to decouple narrative sentiment from strategic logic, showing that human emotional context inherently alters AI decision-making mechanics.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLMs, Game Theory, Emotional Framing, Nash Equilibrium, Benchmark

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_96VWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces EMPAThE, a benchmark for testing whether emotionally charged natural-language framing changes LLM strategic behavior in game-theoretic environments. The authors keep the nominal payoff matrices fixed and use an automated LLM-based pipeline to generate emotional variants of the prompts. They compare equilibrium adherence across eight games (five simultaneous, three sequential) and five small open-weight LLMs. They evaluate pure-strategy equilibrium adherence (PNE/SPNE) and a mixed-strategy compliance score (MNE) over 100 sampled rounds per condition. The main empirical claim is that emotional framing systematically moves models away from Nash-equilibrium behavior, with particularly large changes in several sequential games and coordination games.

Critical strengths

The paper asks an interesting and potentially important question with a simple experimental design: can emotionally salient language alter strategic decisions when the underlying formal game is unchanged? Separating payoff structure from linguistic framing is conceptually clean (payoff preservation is checked automatically by a regex verifier), and it could grow into a useful benchmark for studying the contextual robustness of LLM agents.

Critical weaknesses
  1. It seems that the equilibrium metrics, as implemented, do not adequately support the paper’s claims about strategic rationality. For sequential games, the paper scores whether the realized terminal/joint action outcome belongs to a set of pure or SPNE profiles, but subgame-perfect equilibrium is a strategy profile specifying actions at every information/history node, not merely a terminal action sequence. Checking terminal-outcome membership cannot establish that an agent followed a subgame-perfect strategy. There is also a conceptual problem with the MNE metric: the authors average the theoretical mixed strategies of the two player roles into a single probability vector and compare it with pooled action frequencies. This loses the player-specific strategic structure of the equilibrium. A model could match the pooled marginal frequency while using the wrong strategy for each player. Consequently, the reported compliance and degradation scores may not measure equilibrium adherence in the sense the paper claims.
  2. The emotional manipulation does not isolate emotion from semantic and contextual changes. The central causal claim is that emotion changes strategic reasoning, but the emotional prompts differ from the controls in much more than emotional intensity: length, syntax, narrative structure, renamed action labels, moral valence, and specific content. For example, the control Prisoner’s Dilemma is a generic game description, whereas the emotional version is a detailed scenario about a tenured professor facing scandal, expulsion, betrayal, and tenure, with panic and agony. Observed shifts may therefore stem from role-play, relabeled actions (e.g., “FRAME_RIVAL” instead of “DEFECT”), or distraction rather than emotion. Without neutral controls matched on length, syntax, and action-label structure, the causal attribution to emotion is not supported.
Questions for authors
  1. The evaluation uses five relatively small and older open-weight models. Do the results replicate with contemporary LLMs?
  2. Why is there no discussion of closely related work, e.g., GTBench, EAI, B3Emo, and Economics Arena for LLMs?
  3. Can the authors rerun the experiments with neutral paraphrases that match the emotional prompts in length, syntax, and action-label structure but contain no high-arousal language? Would the NE degradation persist?
  4. For asymmetric games, can the authors report MNE compliance per player role rather than aggregating across roles? Would the conclusions change?
2 comments

We thank the reviewer for recognizing the value and simplicity of our work and recognizing the potential of growing this work into a useful benchmark on contextual robustness of LLMs.

Please refer to the general comment regarding the concerns of keeping the prompt isolated to emotional changes only and on larger models.

Below contains answers to your other questions.

  1. See general comment on new results on Qwen 3.8 27B
  2. We have added discussions on GTBench, EAI, B3Emo, and Economics Arena in a new section.
  3. Please see the general comment on more isolated experiments.
  4. As suggested, for sequential games, matching a short action tuple is insufficient to verify SPNE because it does not assess contingent behavior in every subgame. We will remove claims of verified SPNE adherence from the current analysis. We believe that there are multiple ways to target this and each target is a research paper on its own. We want to have our work motivate future works on LLM’s adherence to multi-turn mixed Nash equilibria.

Dear authors, thanks for your efforts. I see a strange addition of embodied agent interface, which is irrelevant to your paper... I believe, it should be clear from context that I meant the paper "EAI: Emotional Decision-Making of LLMs in Strategic Games and Ethical Dilemmas"

Reviewer_g13HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks how framing emotion into game descriptions influences LLM decision-making. The authors compare canonical descriptions with variations that are loaded with emotions but preserve the numerical payoffs across eight different games on five different LLMs. This evaluation framework generates such variations and also considers both adherence to pure Nash Equilibrium and the closeness of action frequencies to mixed-equilibrium targets. The results show behavioral changes (both improvements and declines depending on the game and the model). The authors interpret these changes through emotional framing and propose post-training alignment as a possible explanation.

Critical strengths
  1. Overall, the paper is well written and organized. The motivation, prompt-generation pipeline, figures, and flow from evaluation to discussion are easy to follow. I appreciated the discussion of "underlying mechanisms," although they remain hypotheses and have not been empirically established.
  2. Several model families and a variety of games (parallel and sequential) are covered in the evaluation work, and the observed variation across these settings provides a basis for studying how different presentations of a task change strategic decisions.
Critical weaknesses
  1. The paper is about the effect of emotion, but experiments do not necessarily isolate emotion-specific effects, and the interpretation is stronger than the evidence supports.

    Emotional rewrites modify wording, length, the context of the narrative, and sometimes action labels. Just keeping the numerical payoffs does not control for other changes. To give examples, in "Prisoner's Dilemma," rewrite replaces COOPERATE/DEFECT with MAINTAIN_TRUST/FRAME_RIVAL and defines an academic scandal case. I could not find matched neutral rewrites that would separate the emotional framing from regular prompt sensitivity or changes in the task interpretation. Prior work, such as Sclar et al. (ICLR 2024) and Hejabi et al. (ACL 2026), makes this an important alternative explanation for the change in behavior. Also, merely changing the wording does not establish that the underlying game or its solutions cannot be recognized or recalled from training.

    The interpretation of the results should also consider the substantial positive results more carefully. In Table 1, there is an improvement from 34% to 92% for Qwen on Centipede and from 2% to 54% on Trust. These improvements might be consistent with the overall aggregate decline, but the explanation of "accidental alignment" is not empirically established. Based on the current evidence, support for "sensitivity to the rewrites" seems stronger than for emotion-specific defects in the model's behavior.

  2. The mixed-equilibrium evaluation contains inconsistencies between the stated targets and the displayed payoff matrices, and also, limitations in how compliance is interpreted.

    In Appendix C, two targets conflict with the payoff matrices. In Chicken, let's say q is the opponent's probability of Swerving. Then, the expected payoffs for Swerve and Straight are 3q+(1−q)3q + (1-q) and 4q4q, respectively. Equating them gives q=0.5q=0.5, but the paper stated 0.90.9. In Coordination, the corresponding expected payoffs are 2q2q and 1−q1-q, which gives q=13q=\frac{1}{3} for action A, rather than 0.50.5. It is not clear whether these matrices, reported targets, or implementation are incorrect, so the impact on the reported analysis needs clarification.

    Also, in Section 3.5.2, the statement:

    In simultaneous games with multiple pure equilibria, models must randomize choices to satisfy expected-utility indifference conditions.

    is not generally valid. For instance, in the Coordination game, always selecting A by both players is already a pure equilibrium. Therefore, deviating from a selected mixed equilibrium by itself does not show impaired strategic reasoning.

    Also, pooling action frequencies across player roles might lose some information. Matching the pooled target is not sufficient to establish whether each player follows the correct equilibrium strategy. So pure-equilibrium adherence and closeness to a particular mixed target should be treated as separate measures and not alternative indicators of being rational.


Full references:

  • Sclar et al., "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting," ICLR 2024.
  • Hejabi et al., "Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs," ACL 2026.
Questions for authors
  1. Can you compare multiple emotional rewrites with matched neutral rewrites, controlling for wording, length, scenario, and action labels as closely as possible? Does an additional effect remain that can reasonably be attributed to emotional framing? Can you also explain and clarify how semantic equivalence is checked beyond preserving numerical payoffs, and in particular for the Centipede example's reference to the final turn?
  2. Can you resolve the discrepancies for the Chicken and Coordination targets with Tables 5 and 6? Which matrices and targets were used in the implementation, and how would correcting any discrepancies affect the presented results? Can you please also justify the requirement to match a mixed equilibrium when valid pure equilibria exist, and, along with the pooled action frequencies, provide role-specific frequencies?
  3. What evidence(s) separate the proposed "accidental alignment" explanation for improvements from alternate explanations? Can you discuss positive and negative effects together and clarify which conclusions remain supported without identifying the "underlying mechanism"?
  4. Please clarify these reproducibility details:
    • How do fractional percentages such as 7.2% and 57.1% arise from the stated 100-round binary scoring?
    • Are there independent repeated runs and uncertainty estimates?
    • How fallback rewrites were handled, including the reported emotional score increases of 0.284 and 0.298 that are below the mentioned threshold of 0.30.
2 comments

We thank the reviewer for recognizing the value and organization of our work and the breadth of our evaluation.

Please refer to the general comment regarding the concerns of keeping the prompt isolated to emotional changes only.

Below contains answers to your other questions.

  1. Please refer to the general comment

  2. Thank you for pointing out the error for the Chicken and Coordination game. We found that the appendix C has a typo in the payout matrices and it is an old version of the experiment. For our reported experiment, we have used the following payouts with the equilibrium probabilities as stated.

    • For chicken game, the experimented payoffs are
      • "Both swerve: Each gets 0 points"
      • "One swerves, one goes straight: Straight gets 1 point, Swerve gets -1 point"
      • "Both go straight (crash): Each gets -10 points"
    • For coordination game, the experimented payoffs are
      • "Both choose A: Each gets 1 point"
      • "Both choose B: Each gets 1 point"
      • "Different choices: Each gets 0 points" We also agree that pooling actions to adhere to mixed-nash equilibrium may cause information loss. We believe that there are multiple ways to target this and each target is a research paper on its own. We want to have our work motivate future works on LLM’s adherence to multi-turn mixed Nash equilibria.
  3. We agree that the accidental alignment argument is not supported by the experiment and have removed it as an explanation.

  4. Reproducibility details

    • Fractional percentages are also incorrect from our 500 runs that we have done in previous experiments. We have corrected its values in the general comment above.
    • Every run is independent of one another.
    • Fallback rewrites were performed by running another query with varying temperature to ensure that a prompt can be generated. Once a prompt is generated, it was kept static and was reused across each run.

We hope these changes will help mitigate all the remaining the critical weaknesses and steer it towards acceptance.

Thank you

Thank you for your response and the additional experiments. I appreciate that you kept the action labels and key instructions fixed and closely matched the prompt lengths. The corrected payoffs also resolve my question about the targets for Chicken and Coordination.

I will keep my original score. My main concern is that the experiment still uses one fixed rewrite per game. Without several matched neutral and emotional rewrites of the same scenario, it is difficult to tell how much of the change comes from emotion and how much comes from the particular wording. I also remain concerned about how the mixed-equilibrium results are interpreted. For example, if both players always choose A in Coordination, they are playing a valid equilibrium even though neither follows the 50–50 target.

I appreciate that you withdrew the "accidental alignment" explanation. One result I still find hard to interpret is Qwen‑8B on Trust. Adherence increased from 2% to 54% in the paper but fell from 95% to 0% in the rebuttal. It remains unclear how much of that reversal comes from the new prompts and how much comes from correcting the earlier results.

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

The paper introduces EMPAThE, which generates emotionally reframed variants of eight canonical games while holding payoff structures constant, and evaluates five open-weight LLMs on Nash Equilibrium adherence (Abstract, p.1; Sections 3.1–3.3). It claims emotional prompting "systematically" degrades game-theoretic rationality across architectures (Sections 4.2–4.3).

Critical strengths
  1. A controlled-isolation pipeline for affective prompting: emotional variants are gated by a regex payoff-isomorphism verifier and EmoPoint scoring, with payoff matrices held fixed, yielding a reusable, fully artifacted evaluation toolkit.

  2. Empirical breadth: five models × eight games × two equilibrium paradigms, 100 rounds each, with model-specific profiles such as Gemma-2B Trust Δ −54.0 and Qwen-8B Truel Δ −78.0.

Critical weaknesses
  1. The emotional variants change more than the tone: the generator rewrites the canonical description and remaps the action labels, while the control prompts are described as testing "direct recall of memorized NE". Without a neutral rewrite that uses the same relabeled actions, the drops in adherence cannot be separated from the loss of the familiar canonical wording, and adding this one control condition would settle the attribution to emotion.

  2. The Conclusion states that this vulnerability "persists regardless of parameter scale", but the five evaluated models are compact open-weight models. Results on one or two substantially larger models would be needed to support this claim about scale.

Questions for authors

See Critical Weaknesses

1 comment

We thank the reviewer for recognizing the value of our work and the breadth of our evaluation.

Please refer to the general comment regarding the concerns of keeping the prompt isolated to emotional changes only.

To address your concerns on the large open-weight models, we have added new results for Qwen 3.8 27B in the general comments as well. We hope these changes will help mitigate all the remaining the critical weaknesses and steer it towards acceptance.

Thank you

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment to All Reviewers

We thank the reviewers for recognizing the value of our work. We want to emphasize the works’ contribution to the community on the potential of growing into a useful benchmark on contextual robustness of LLMs. Below, we highlight answers to commonly shared questions.

  • Isolation of emotion specific effects: We have ran additional experiments with the following setup. 1. Keep the length of the prompt constant, 2. Keep the action labels constant, 3. Keep the direction prompts the same, and 4. Change the emotion levels of remaining part of the prompt. With that, we obtain the following results.

Table 1: Reported NE adherence (%) across models, games, and prompt types.

ModelTypeBoSChkCoordPDStagCentTrust
Gemma-2BControl74.010.0100.084.062.025.088.0
Emotion66.00.0100.00.096.00.00.0
Δ-8.0-10.00.0-84.0+34.0-25.0-88.0
Gemma-4BControl84.036.0100.0100.072.00.00.0
Emotion55.027.098.098.051.00.00.0
Δ-29.0-9.0-2.0-2.0-21.00.00.0
LLaMAControl76.08.0100.00.097.00.00.0
Emotion6.08.00.092.048.0100.00.0
Δ-70.00.0-100.0+92.0-49.0+100.00.0
NemotronControl50.050.0100.018.0100.020.032.0
Emotion100.04.0100.099.0100.00.00.0
Δ+50.0-46.00.0+81.00.0-20.0-32.0
Qwen-8BControl100.02.0100.0100.050.020.095.0
Emotion100.00.0100.0100.010.00.00.0
Δ0.0-2.00.00.0-40.0-20.0-95.0
Qwen-27BControl0.077.0100.075.0100.05.00.0
Emotion100.00.0100.0100.099.045.00.0
Δ+100.0-77.00.0+25.0-1.0+40.00.0

Parallel games: BoS = Battle of the Sexes; Chk = Chicken; Coord = Coordination; PD = Prisoner’s Dilemma; Stag = Stag Hunt.

Sequential games: Cent = Centipede; Trust = Trust Game.

Δ = Emotion − Control, in percentage points. — = not provided. Values reproduce the summary tables in the supplied files.

  • Observing the trend for larger model: We have also run the experiment on Qwen 3.8 27B to see the trends for the larger model in the above table.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.