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

Retro*: Optimizing LLMs for Reasoning-Intensive Document Retrieval

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation 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.

Retro* is a rubric-prompted pointwise LLM reranker that outputs 0-100 relevance scores, integrates multiple sampled trajectories by averaging, and is trained with SFT followed by GRPO using intra- and inter-document composite rewards, reaching SOTA on BRIGHT.

Problem

Reasoning-intensive document retrieval needs fine-grained, implicit relevance judgments; existing LLM rerankers give only relative orderings, lack test-time scaling via multiple reasoning paths, and are listwise/setwise and thus slow to parallelize.

Approach

A rubric-based pointwise relevance scorer that produces an absolute 0-100 score from reasoning, with test-time score averaging over K sampled trajectories and an RL reward combining intra-document consistency and inter-document discrimination.

Key results

• Retro* (7B) with mean-score@16 reaches 38.7 average nDCG@10 on BRIGHT (top-100 from BGE-Reasoner-Embed), above ReasonRank (32B) at 36.6.
• Retro* (32B) with mean-score@16 reaches 40.6 average nDCG@10 on BRIGHT.
• Retro* (7B) reaches 55.8 average nDCG@10 on seven BEIR datasets (56.8 with test-time scaling), above ReasonRank (32B) at 55.4.

Surprising findings

• Test-time scaled 7B model (38.7) exceeds the single-sample 32B model (38.5) on BRIGHT.
• Zero-RL training with the composite reward (no SFT) reaches 35.1 average nDCG@10, close to the full SFT+RL result.

Limitations

• Evaluation restricted to re-ranking top-100 candidates from a single strong first-stage retriever in the main results.
• Score integration uses uniform weights in experiments; likelihood-based weighting is proposed but not evaluated.
• Relies on a very large teacher (Qwen3-235B-A22B) for SFT data.

Experimental setup

Datasets: BGE-Reasoner-Data BRIGHT MS MARCO BEIR

Benchmarks: BRIGHT BEIR

Metrics: nDCG@10

Baselines: RankLLaMA RankZephyr JudgeRank Rank1 Rank-R1 ReasonRank

Builds on: GRPO BRIGHT Rank1 ReasonRank Rank-R1 RankLLaMA

Closest papers by meaning

• Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.87)
• BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval (ICLR 2025 · similarity 0.87)
• Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.86)
• Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.86)
• Scent of Knowledge: Optimizing Search-Enhanced Reasoning with Information Foraging (NeurIPS 2025 · similarity 0.86)
• DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.85)