Pith. sign in

REVIEW 12 cited by

ReMA: Learning to Meta-think for LLMs with Multi-Agent Reinforcement Learning

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.09501 v3 pith:5TOP3QL7 submitted 2025-03-12 cs.AI cs.CLcs.LGcs.MA

classification cs.AIcs.CLcs.LGcs.MA
keywords reasoningmeta-thinkingremalearningllmsagentagentsfurther
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Recent research on Reasoning of Large Language Models (LLMs) has sought to further enhance their performance by integrating meta-thinking -- enabling models to monitor, evaluate, and control their reasoning processes for more adaptive and effective problem-solving. However, current single-agent work lacks a specialized design for acquiring meta-thinking, resulting in low efficacy. To address this challenge, we introduce Reinforced Meta-thinking Agents (ReMA), a novel framework that leverages Multi-Agent Reinforcement Learning (MARL) to elicit meta-thinking behaviors, encouraging LLMs to think about thinking. ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions. Through iterative reinforcement learning with aligned objectives, these agents explore and learn collaboration, leading to improved generalization and robustness. Empirical results from single-turn experiments demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks, including competitive-level mathematical benchmarks and LLM-as-a-Judge benchmarks. Additionally, we further extend ReMA to multi-turn interaction settings, leveraging turn-level ratio and parameter sharing to improve efficiency. Comprehensive ablation studies further illustrate the evolving dynamics of each distinct agent, providing valuable insights into how the meta-thinking reasoning process enhances the reasoning capabilities of LLMs. Our code can be found in https://github.com/ziyuwan/ReMA-public

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 12 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Training Language Models to Cooperate with Inference-Time Controllers

    cs.AI 2026-07 conditional novelty 6.5 of 10

    Multi-controller post-training with turn-level GRPO over reusable reasoning modules improves LLM transfer to held-out and shifted inference-time controllers versus single-controller RL.

  2. MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

    cs.CL 2026-07 conditional novelty 6.5 of 10

    MET-D self-distills theory-selected moral grounds into native-language reasoning, lifting macro-F1 by ~3.7–4.2 points on MCLASH and MMoralExceptQA while raising native-language chains by ~62 points.

  3. Coordinating Search-Informed Reasoning and Reasoning-Guided Search in Claim Verification

    cs.AI 2025-06 conditional novelty 6.0 of 10

    A pair of reinforcement-learning-trained agents, one for reasoning and one for search, improves accuracy on multi-hop claim verification benchmarks.

  4. Generalizable LLM Learning of Graph Synthetic Data with Post-training Alignment

    cs.LG 2025-06 conditional novelty 6.0 of 10

    Post-training alignment on synthetic graph data yields statistically significant gains on some real-world implicit-graph tasks, but gains are inconsistent and multi-step reasoning remains fragile.

  5. Enhancing Social Intelligence in LLMs with Hierarchical Reasoning and Utterance-Level Goal Rewarding

    cs.CL 2026-08 conditional novelty 5.0 of 10

    LHRL-VGR, a two-stage RL method with variance-based reward routing, lets a Qwen2.5-7B agent beat GPT-4o by about 7% on SOTOPIA goal completion.

  6. Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation

    cs.HC 2026-04 conditional novelty 5.0 of 10

    A design-then-code multi-agent workflow, dataset, and small trained models produce template-free slides that humans prefer over strong baselines.

  7. R$^2$PO: Decoupling Rollout and Inference Policies for LLM Reasoning

    cs.LG 2026-01 conditional novelty 5.0 of 10

    R²PO trains LLM reasoning policies with a separate residual rollout head, reporting 3.4% higher MATH-500 and 1.3% higher APPS accuracy than GRPO.

  8. How to Train a Leader: Hierarchical Reasoning in Multi-Agent LLMs

    cs.MA 2025-07 conditional novelty 5.0 of 10

    A leader LLM trained with a GRPO variant that conditions on frozen agent responses improves both collaborative and zero-shot accuracy on BBH, MATH, and MMLU.

  9. The Avengers: A Simple Recipe for Uniting Smaller Language Models to Challenge Proprietary Giants

    cs.CL 2025-05 reject novelty 5.0 of 10

    Clustering-based routing plus self-consistency voting among ten 7B open models reportedly outranks GPT-4.1 and GPT-4.5 on average over 15 diverse benchmarks.

  10. Beyond Single Models: Enhancing LLM Detection of Ambiguity in Requests through Debate

    cs.CL 2025-07 conditional novelty 4.0 of 10

    A leader-follower multi-agent debate protocol improves ambiguity detection for two of three tested LLMs, but the reported results lack error bars, a clear success metric, and contain internal numerical inconsistencies.

  11. CogniSQL-R1-Zero: Lightweight Reinforced Reasoning for Efficient SQL Generation

    cs.AI 2025-07 reject novelty 3.0 of 10

    A 7B model trained with GRPO and a sparse execution-correctness reward reaches 59.97% execution accuracy on BIRD dev, though the evaluation protocol and baseline numbers contain inconsistencies.

  12. Mathematical methods of reinforcement learning

    math.OC 2026-07 accept

    A survey unifying the operator-theoretic, probabilistic, and optimization-based mathematical structures underlying modern reinforcement learning algorithms.

Pith tools