The psychophysics of affect in large language models

Joseph Marks, Marius C. VollbergarXiv 2026 · arXiv

TL;DR

Researchers have discovered that Large Language Models (LLMs) exhibit a unique "silicon gain amplification," where they rate the emotional intensity of financial gains higher than equivalent losses—the exact opposite of human loss aversion. However, these affective mappings are highly malleable, linearizing when models can compare outcomes, and surprisingly, these reported "feelings" do not reliably predict the models' actual choices, which tend toward cold expected-value maximization.

What Problem Does This Paper Address?

As LLMs move from chatbots to agentic decision-makers—such as AI agents authorized to trade stocks on platforms like Robinhood—we face a significant alignment challenge. Human decision-making is rarely linear; we experience "diminishing sensitivity" (the difference between $0 and $10 feels larger than the difference between $100 and $110) and "loss aversion" (loses hurt more than gains help).

If an AI agent simply maximizes expected value, it might take risks that a human user finds unacceptable. Conversely, if we want AI to align with human preferences, we need to understand how they map objective outcomes (like money) to subjective values (feelings). This paper investigates whether LLMs possess a structured "value function" similar to human psychophysics and whether they actually use that function when making choices.

Methods

The authors conducted two primary studies:

  1. Study 1 (Abstract Gambling): Models were presented with 50/50 gambles involving monetary gains and losses (£0.20–£12).
  2. Study 2 (Stock Trading): A more naturalistic framing using percentage price moves in a simulated trading app.

To probe the stability of LLM "feelings," they varied the elicitation context across three conditions:

  • Single Evaluation: Each outcome was rated in total isolation.
  • Memory: Outcomes were rated sequentially, with the model’s previous ratings visible in the prompt history.
  • Joint Evaluation: Multiple outcomes were presented and rated together in a single list.
Task design for Study 1, illustrating the different evaluation conditions and the rating vs. choice tasks.

The researchers tested ten LLMs, including frontier models like GPT-5-nano, Kimi-k2.5, Claude-sonnet-4.5, and Llama-4-maverick. They used power-law functions to model affect ratings and logit models (based on Prospect Theory and Expected Utility) to analyze choice behavior.

Datasets

The study did not use a static training dataset but rather generated a behavioral dataset by simulating 30 "LLM-participants" per model/condition. The stimuli were drawn from a fixed 20-point grid of magnitudes. The authors have deposited the resulting per-trial data and quality-cleaned CSVs in Zenodo for reproducibility.

Key Insights and Ideas

  • Silicon Gain Amplification: In isolation, LLMs show diminishing sensitivity to magnitude, mimicking the "curved" value function of humans. However, they consistently rate gains as more intense than losses. This is a complete inversion of human psychology, where "losses loom larger than gains."
  • Valuation is Constructed, Not Retrieved: The models' internal "feelings" are not fixed. When models were allowed to compare outcomes (Joint Evaluation), their ratings shifted from curved to almost perfectly linear. This suggests that LLMs don't have a stable internal utility curve; they "calculate" a mapping based on the immediate context provided in the prompt.
  • The Affect-Choice Dissociation: There is a striking gap between what a model says it "feels" and what it chooses. While their ratings were non-linear and gain-biased, their choices were largely driven by expected value (risk-neutrality). Elicited affect ratings did not explain choices better than simple monetary utility.
  • The Pollyanna Hypothesis: The authors speculate that "silicon gain amplification" might be an inherited bias from human language. Human text is statistically skewed toward positive emotion words, even if our lived experience is more sensitive to the negative. LLMs, trained on this text, may simply be reflecting the "positivity bias" of human discourse.

Key Results

  • Curvature: In single evaluation, all nine post-trained models showed magnitude exponents ($\rho$) significantly less than 1 (ranging from 0.01 to 0.37), indicating extreme diminishing sensitivity.
  • Linearization: In the joint evaluation condition, the exponent rose to $\rho \approx 0.87–0.98$, representing a near-complete linearization of the value function.
Affect ratings for Kimi-k2.5 and GPT-5-nano across the three conditions, showing the transition from curved, asymmetric ratings to linear, symmetric ones.
  • Choice Behavior: Unlike humans (whose loss aversion parameter $\lambda$ is typically $\approx 2$), LLMs hovered around $\lambda \approx 1$. For example, GPT-5-nano’s $\lambda$ was measured between 1.005 and 1.04, indicating almost no loss aversion in its choices.

Limitations and Future Work

The authors acknowledge several bounds to their findings:

  • Model Selection: Comparative conditions were only run for two focal models (GPT-5-nano and Kimi-k2.5).
  • Prompt Engineering: The results are specific to the "first-person" elicitation used. It remains unknown if the mapping changes if the model is asked to report a user's feelings instead of its own.
  • Scale Artefacts: The use of a bounded -50 to +50 scale might mechanically contribute to some of the observed diminishing sensitivity.

Future research could involve interpretability probes to see if these affective mappings have internal neural counterparts within the model's activations, or expanding the study to non-monetary domains like health or social outcomes.

Summary updated

Abstract

Good decisions often require considering how outcomes will feel. Yet large language models (LLMs) increasingly enter decision workflows, despite little being known about how they map outcomes to feelings or use those mappings in choice. We tested how LLMs anticipated gambling or stock-trading outcomes would feel and whether they incorporated these ratings into choice. When outcomes were rated in isolation, ratings exhibited diminishing sensitivity to magnitude, resembling human value functions; but models rated gains more intensely than equivalent losses, an inversion of canonical loss aversion we call silicon gain amplification. When outcomes were rated together, enabling direct comparison, ratings linearised. In choice, the gain-loss asymmetry largely disappeared, and models' own ratings did not reliably explain their choices better than monetary utility did. These findings reveal malleable mappings from outcomes to affect, raising questions about whether and how agentic LLMs should incorporate anticipated feelings into decisions.
BibTeX
@misc{marks2026psychophysics,
  author = {Joseph Marks and Marius C. Vollberg},
  title = {{The psychophysics of affect in large language models}},
  year = {2026},
  eprint = {2610.40985},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40985},
}
Figure