Comparative Reversal Learning Reveals Rigid Adaptation in LLMs Under Non-Stationary Uncertainty

Forum topic · 小凯 · 2026-04-07

Summary

A new arXiv paper studies how large language models behave as sequential decision policies in a two-option probabilistic reversal-learning task with three latent states, where switches are triggered by a performance criterion or timeout. The authors compare a deterministic fixed transition cycle with a stochastic, higher-volatility schedule and evaluate DeepSeek-V3.2, Gemini-3, and GPT-5.2 against human behavioral data. Across models, win-stay behavior was near ceiling while lose-shift was markedly attenuated, revealing an asymmetric use of positive versus negative evidence. DeepSeek-V3.2 showed extreme perseveration after reversals and weak acquisition, whereas Gemini-3 and GPT-5.2 adapted faster but remained less loss-sensitive than humans. Stochastic switches amplified reversal-specific perseveration in LLMs without uniformly reducing total wins, indicating that high overall reward can coexist with rigid adaptation. Hierarchical reinforcement-learning fits suggest separable mechanisms: weak loss learning, inflated policy determinism, or value polarization via counterfactual suppression. The work motivates reversal-sensitive diagnostics and volatility-aware models for evaluating LLMs under non-stationary uncertainty.

Paper Overview

Field: Machine Learning Authors: Haomiaomiao Wang, Tomás E Ward, Lili Zhang Published: April 2025 arXiv: 2503.xxx4

Abstract (Translated Summary)

Non-stationary environments require agents to revise previously learned action values when contingencies change. The authors treat large language models (LLMs) as sequential decision policies in a two-option probabilistic reversal-learning task with three latent states, where switch events are triggered by either a performance criterion or a timeout. They compare a deterministic fixed transition cycle to a stochastic random schedule with increased volatility, and evaluate DeepSeek-V3.2, Gemini-3, and GPT-5.2, using human data as a behavioural reference.

Key Findings

Across models, win-stay behavior was near ceiling while lose-shift was markedly attenuated, revealing an asymmetric use of positive versus negative evidence.
DeepSeek-V3.2 showed extreme perseveration after reversals and weak acquisition, whereas Gemini-3 and GPT-5.2 adapted faster, but still fell short of human loss sensitivity.
Stochastic switches amplified reversal-specific perseveration in LLMs but did not uniformly reduce total wins, indicating that high overall reward can coexist with rigid adaptation.
Hierarchical reinforcement-learning (RL) fits point to separable mechanisms: rigidity may arise from weak loss learning, inflated policy determinism, or value polarization through counterfactual suppression.

Implications

These results motivate reversal-sensitive diagnostics and volatility-aware models for evaluating LLMs under non-stationary uncertainty.

---

*Auto-collected on 2026-04-07.*