Research map / Large language models / Reasoning and chain of thought

Reward Is Enough: LLMs Are In-Context Reinforcement Learners

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ICRL prompting, a multi-round prompt that feeds an LLM its prior responses and scalar rewards so it improves at inference time without weight updates, and evaluates it on Game of 24, creative writing, ScienceWorld, AIME and HMMT.

Problem

Whether LLMs can perform reinforcement learning purely in context at inference time, improving on tasks using only scalar rewards, without parameter updates or external feedback text.

Approach

A minimal multi-round prompting loop that places prior responses with their scalar rewards (no verbal feedback, no filtering, no replay) into the context, plus exploration/exploitation meta-instructions, so the frozen LLM improves as context grows; the same LLM can serve as its own reward model.

Key results

• On Game of 24 with GPT-4.1, ICRL Preset reaches 90% running-max success after 50 trials versus 49% for Best-of-N, 47% for Self-Refine and 44% for Reflexion.
• On creative writing, ICRL achieves 86.32% LC win rate against Self-Refine and 59.48% against Reflexion (Alpaca-Eval 2).
• On AIME with Qwen3-32B (32k), ICRL reaches 46.66% versus 43.33% for Self-Refine and 33.33% for Reflexion; on HMMT, 33.33% versus 16.66% and 23.33%.

Surprising findings

• Even when the reward is produced by the same LLM through self-evaluation, ICRL still improves performance over episodes.
• Exploration-only prompting without rewards performs much worse in running max than ICRL, suggesting gains are not merely best-of-N selection.

Limitations

• Gains from self-evaluated rewards are expected to have a lower ceiling than external feedback, which the authors hypothesize but do not fully test.
• Experiments mostly rely on closed API models (GPT-4.1) with results on open models limited to a few tasks.
• Several comparisons use single-run settings and small task sets (100 problems, 30 tasks, 30 arXiv papers).

Experimental setup

Datasets: Game of 24 (100 problems) creative writing sentence pool (100 problems) ScienceWorld (30 tasks) AIME HMMT arXiv papers post-cutoff (30)

Benchmarks: Game of 24 Creative Writing (Alpaca-Eval 2 LC win rate) ScienceWorld AIME HMMT unseen arXiv abstract generation

Metrics: success rate length-controlled win rate return running max % solved ROUGE-recall

Baselines: CoT-only Long-CoT Best-of-N Self-Refine Reflexion ReAct

Builds on: Self-Refine Reflexion Best-of-N Long-CoT prompting

Closest papers by meaning

• Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs (ICML 2025 · similarity 0.90)
• Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
• LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL (NeurIPS 2026 · similarity 0.89)
• Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.89)
• TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Are Human-generated Demonstrations Necessary for In-context Learning? (ICLR 2024 · similarity 0.89)