Research map / Large language models / Retrieval and knowledge

Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Retrieval-augmented reasoning

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 FREESON, where a single large reasoning model acts as both generator and retriever by traversing the corpus with a retrieval-specialized MCTS (CT-MCTS) constrained by an FM-index, with a learned value network for answer containment.

Problem

Retrieval-augmented reasoning relies on separate dual-encoder retrievers whose embedding space creates a representation bottleneck and adds hardware cost; the paper asks whether an LRM can retrieve from a corpus itself during reasoning.

Approach

Replace embedding similarity with corpus traversal: constrain token-level MCTS to sequences existing in the corpus via a prefix-based FM-index (CorpusTree), using multi-token nodes, stochastic beam-style multi-node expansion driven by LRM probabilities, and a value network trained on on-policy CT-MCTS rollouts to score answer presence.

Key results

• FREESON averages about 14.4% gain in EM and F1 over four multi-step reasoning baselines with separate retrievers.
• FREESON matches Search-R1 overall, surpassing it by about 3% on PopQA and 2WikiMultihopQA (2Wiki EM 0.55 vs 0.54).
• Node granularity G=6 gives EM 0.55/F1 0.63 on 2Wiki versus 0.30/0.33 at G=1; CT-MCTS reaches 0.44 EM on PopQA vs 0.17 for greedy search.

Surprising findings

• Ground-truth labels in QA benchmarks built with embedding retrievers may penalize semantically correct answers that FREESON produces, potentially underestimating its performance.

Limitations

• Best suited to predefined corpora; may underperform web-based retrieval for open-ended QA.
• Inference-time traversal is not optimized for document structure; RL (e.g., PPO) for traversal is future work.
• Evaluation uses 500-example subsets per dataset and ground-truth labels may bias against correct but differently phrased answers.

Experimental setup

Datasets: PopQA TriviaQA HotpotQA 2WikiMultihopQA MuSiQue

Benchmarks: PopQA TriviaQA HotpotQA 2WikiMultihopQA MuSiQue

Metrics: EM F1 latency

Baselines: Qwen2.5-7B R1-Distill-Qwen-7B E5 + Qwen2.5-7B FLARE Self-Ask Search-o1 Search-R1 Greedy Search Beam Search

Builds on: Search-R1 MCTS UCT FM-index PPO-MCTS

Closest papers by meaning

• BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval (ICLR 2025 · similarity 0.89)
• Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
• Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.89)
• From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning (ICML 2025 · similarity 0.88)
• Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)
• SLQ: Bridging Modalities via Shared Latent Queries for Retrieval with Frozen MLLMs (ICML 2026 · similarity 0.88)