In-Context Learning as Implicit Policy Gradient

• URL: http://arxiv.org/abs/2607.23153v1
• Date: Sat, 25 Jul 2026 11:13:52 GMT
• Title: In-Context Learning as Implicit Policy Gradient
• Abstract summary: Large language models can iteratively improve their outputs by incorporating generated samples and their corresponding evaluation scores as in-context examples.<n>We show that score-conditioned In-Context Learning (ICL) admits a structural correspondence to policy gradient optimization.
• Score: 47.12608115550359
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent work has shown that large language models (LLMs) can iteratively improve their outputs by incorporating generated samples and their corresponding evaluation scores as in-context examples. Despite these empirical findings, the theoretical foundations underlying this phenomenon remain poorly understood. In this paper, we show that score-conditioned In-Context Learning (ICL) admits a structural correspondence to policy gradient optimization. We first provide a constructive proof that self-attention mechanisms can implement reward-weighted aggregation analogous to the REINFORCE algorithm under specific weight matrix configurations, and discuss the relationship between this construction and the behavior of pretrained transformers. The correspondence is directional in hidden-state space and holds exactly only under the stated simplifying conditions; we quantify its strength empirically. Within our simplified hidden-state model, we furthermore derive an exact upper bound on the distribution shift induced by a bounded attention update, yielding a trust-region-like analogy to KL-constrained policy optimization. We validate our theory through extensive experiments across multiple LLMs, demonstrating that LLMs effectively utilize score information to shift output distributions toward high-scoring exemplars, and that attention weights exhibit a strong correlation with example scores.

Related papers

• SimReg: Achieving Higher Performance in the Pretraining via Embedding Similarity Regularization [55.63565289584336]
  SimReg is an embedding similarity regularization loss that explicitly encourages token representations with the same ground-truth label within each sequence to be more similar.<n>Our analysis reveals that this mechanism introduces gains by enlarging multi-classification margins, thereby enabling more efficient classification.
  arXiv  Detail & Related papers  (2026-05-09T08:59:13Z)
• Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models [38.47027398567909]
  Perception-Grounded Policy Optimization (PGPO) is a novel fine-grained credit assignment framework that dynamically reshapes advantages at the token level.<n>We show that PGPO actively amplifies learning signals for visually-dependent tokens while suppressing gradient noise from linguistic priors.<n>Both theoretical and empirical analyses confirm that PGPO effectively reduces gradient variance, prevents training collapse, and acts as a potent regularizer for robust, perception-grounded multimodal reasoning.
  arXiv  Detail & Related papers  (2026-04-02T09:53:20Z)
• A Comedy of Estimators: On KL Regularization in RL Training of LLMs [81.7906270099878]
  reinforcement learning (RL) can substantially improve the reasoning performance of large language models (LLMs)<n>The RL objective for LLM training involves a regularization term, which is the reverse Kullback-Leibler (KL) divergence between the trained policy and the reference policy.<n>Recent works show that prevailing practices for incorporating KL regularization do not provide correct gradients for stated objectives, creating a discrepancy between the objective and its implementation.<n>We study the gradients of several estimators configurations, revealing how design choices shape gradient bias.
  arXiv  Detail & Related papers  (2025-12-26T04:20:58Z)
• OBLR-PO: A Theoretical Framework for Stable Reinforcement Learning [12.77713716713937]
  We provide a unified theoretical framework that characterizes the statistical properties of commonly used policy-gradient estimators.<n>We derive an adaptive learning-rate schedule governed by the signal-to-noise ratio (SNR) of gradients.<n>We further show that the variance-optimal baseline is a gradient-weighted estimator, offering a new principle for variance reduction.
  arXiv  Detail & Related papers  (2025-11-28T16:09:28Z)
• Revisiting LLM Reasoning via Information Bottleneck [57.519119962528166]
  Large language models (LLMs) have recently demonstrated remarkable progress in reasoning capabilities through reinforcement learning with verifiable rewards (RLVR)<n>We present a theoretical characterization of LLM reasoning grounded in information bottleneck (IB) principle.<n>We propose IB-aware reasoning optimization (IBRO), a framework that encourages reasoning trajectories to be both informative about the final correct answer and generalizable.
  arXiv  Detail & Related papers  (2025-07-24T13:14:25Z)
• Provable Low-Frequency Bias of In-Context Learning of Representations [19.066378730056275]
  In-context learning (ICL) enables large language models (LLMs) to acquire new behaviors from the input sequence alone without any parameter updates.<n>Recent studies have shown that ICL can surpass the original meaning learned in pretraining stage through internalizing the structure the data-generating process (DGP) of the prompt into the hidden representations.<n>We present the first rigorous explanation of such phenomena by introducing a unified framework of double convergence.<n>This double convergence process leads to an implicit bias towards smooth (low-frequency) representations, which we prove analytically and verify empirically.
  arXiv  Detail & Related papers  (2025-07-17T21:19:32Z)
• CTRLS: Chain-of-Thought Reasoning via Latent State-Transition [57.51370433303236]
  Chain-of-thought (CoT) reasoning enables large language models to break down complex problems into interpretable intermediate steps.<n>We introduce groundingS, a framework that formulates CoT reasoning as a Markov decision process (MDP) with latent state transitions.<n>We show improvements in reasoning accuracy, diversity, and exploration efficiency across benchmark reasoning tasks.
  arXiv  Detail & Related papers  (2025-07-10T21:32:18Z)
• Large Language Models as Computable Approximations to Solomonoff Induction [11.811838796672369]
  We establish the first formal connection between large language models (LLMs) and Algorithmic Information Theory (AIT)<n>We leverage AIT to provide a unified theoretical explanation for in-context learning, few-shot learning, and scaling laws.<n>Our framework bridges the gap between theoretical foundations and practical LLM behaviors, providing both explanatory power and actionable insights for future model development.
  arXiv  Detail & Related papers  (2025-05-21T17:35:08Z)
• Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLM [11.181783720439563]
  Large Language Models (LLMs) display sophisticated reasoning abilities via extended Chain-of-Thought (CoT) generation.<n>RLMs often demonstrate counterintuitive and unstable behaviors, such as performance degradation under few-shot prompting.<n>We introduce a unified graph-based analytical framework for better modeling the reasoning processes of RLMs.
  arXiv  Detail & Related papers  (2025-05-20T03:54:57Z)
• The Curse of CoT: On the Limitations of Chain-of-Thought in In-Context Learning [56.574829311863446]
  Chain-of-Thought (CoT) prompting has been widely recognized for its ability to enhance reasoning capabilities in large language models (LLMs)<n>We demonstrate that CoT and its reasoning variants consistently underperform direct answering across varying model scales and benchmark complexities.<n>Our analysis uncovers a fundamental hybrid mechanism of explicit-implicit reasoning driving CoT's performance in pattern-based ICL.
  arXiv  Detail & Related papers  (2025-04-07T13:51:06Z)
• Efficient and Flexible Neural Network Training through Layer-wise Feedback Propagation [49.44309457870649]
  Layer-wise Feedback feedback (LFP) is a novel training principle for neural network-like predictors.<n>LFP decomposes a reward to individual neurons based on their respective contributions.<n>Our method then implements a greedy reinforcing approach helpful parts of the network and weakening harmful ones.
  arXiv  Detail & Related papers  (2023-08-23T10:48:28Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.