Skip to content

Retro*: Optimizing LLMs for Reasoning-Intensive Document Retrieval

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Optimizes LLMs for indirect reasoning-heavy retrieval required in agentic pipelines

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Retro* is a reasoning-intensive document retrieval model that introduces rubric-based pointwise relevance scoring — where an LLM reasons against explicitly defined criteria to produce a 0–100 interpretable relevance score — combined with a two-stage SFT+RL training strategy using composite intra- and inter-document rewards. It achieves state-of-the-art nDCG@10 on the BRIGHT benchmark, outperforming all reasoning-enhanced baselines at both 7B and 32B scale. Test-time scaling via score integration of multiple trajectories further improves accuracy, enabling the 7B model to exceed the standard 32B baseline.

Problem

Existing re-ranking approaches for reasoning-intensive retrieval suffer three compounding limitations: (1) they produce only relative ranking orders rather than interpretable absolute relevance scores needed by downstream RAG applications; (2) they rely on a single long-form reasoning chain without exploiting multiple sampling paths for more reliable estimates; and (3) listwise and setwise methods must process documents sequentially, creating latency bottlenecks as candidate set size grows.

Method

Retro* operates as a pointwise relevance scorer. Given a query \(q\), a candidate document \(d\), and a task-specific relevance rubric \(\Gamma\), the model generates a reasoning trajectory \(y\) and an integer relevance score \(s \in [0, 100]\): \(\text{Retro}(\Gamma, q, d) \to (y, s)\). The rubric defines five labeled score bands (0–20 Irrelevant to 80–100 Highly Relevant) with a user-configurable relevance definition placeholder.

Test-time scaling samples \(K\) trajectories per query-document pair and aggregates them as: $\(\bar{s} \leftarrow \frac{\sum_{i=1}^{K} w_i s_i}{\sum_{i=1}^{K} w_i}\)$ with uniform weights \(w_i = 1/K\) (mean-score@\(k\)) as the default.

Training is two-stage: - SFT warm-up: A teacher model (Qwen3-235B-A22B) generates \(K=8\) trajectories per sample; the one whose score is closest to the integrated mean \(\bar{s}\) is selected for distillation. - RL stage: GRPO optimizes a composite reward \(R = \alpha \cdot R_{\text{intra}} + (1-\alpha) \cdot R_{\text{inter}}\). - \(R_{\text{intra}}\): ternary reward (\(+1\)/\(-1\)/\(0\)) based on whether a trajectory's score is closest/furthest from \(\bar{s}\) across \(N\) rollouts, with a minimum-gap threshold \(\tau\). - \(R_{\text{inter}}\): for a positive-negative document pair, the reward equals the proportion of cross-document comparisons won: \(R_{\text{inter}}(y_i^+, s_i^+) = \frac{1}{N}\sum_j \mathbb{I}(s_i^+ > s_j^-)\).

Key Contributions

  • Rubric-based pointwise relevance scoring that yields interpretable absolute relevance estimates (0–100), enabling direct use in RAG pipelines beyond relative ranking.
  • Score integration strategy for test-time scaling that is compatible with any parallel per-document inference.
  • Composite RL reward combining intra-document scoring stability and inter-document ranking discrimination, trained via GRPO on top of distilled SFT initialization.
  • Demonstration that test-time scaling (mean-score@16) on a 7B model surpasses the standard 32B model on BRIGHT.

Results

  • Retro* (7B): nDCG@10 = 36.6 on BRIGHT; beats ReasonRank (7B) by +3.1 points.
  • Retro* (32B): nDCG@10 = 38.5; beats ReasonRank (32B) by +1.9 points.
  • Retro* (7B) + mean-score@16: nDCG@10 = 38.7, exceeding the standard 32B model (38.5).
  • Retro* (32B) + mean-score@16: nDCG@10 = 40.6.
  • Results hold across multiple first-stage retrievers (BM25: 35.3 → 36.6 for 7B@1 → 32B@1; ReasonIR: 36.8 → 37.4).
  • BGE-Reasoner-Embed retriever baseline (no re-ranking): nDCG@10 = 27.8 average on comparable subsets.

Limitations

  • Training data is entirely synthetic (BGE-Reasoner-Data), potentially limiting generalization to domains not covered by BRIGHT's 12 datasets.
  • Score integration requires multiple inference passes (\(K\) trajectories per document), increasing latency relative to single-pass methods despite parallelism over documents.
  • The rubric threshold \(\tau\) and weighting parameter \(\alpha\) require tuning; sensitivity analysis is deferred to appendices.
  • Out-of-domain evaluation (R2MED biomedical benchmark) is discussed only in the appendix; the main-text evidence for generalization is limited.
  • Only Qwen2.5-Instruct family backbones are used in primary experiments; transferability to other model families is not fully established in the main paper.

Relevance to Agentic AI / LLM Agents

RAG is the standard memory-access mechanism for LLM agents tackling tasks in software engineering, mathematics, and scientific research — precisely the domains where implicit, reasoning-required document relevance is the bottleneck. Retro* directly addresses retrieval quality in these agentic pipelines by replacing brittle semantic-similarity retrieval with explicit rubric-guided reasoning, making it a practical drop-in re-ranker for agent memory modules. Its pointwise, parallel-inference architecture is well-suited to agent loops where latency and modularity matter, unlike listwise methods that must observe the full candidate set. The test-time scaling result — that spending more inference compute improves retrieval rather than generation — extends the "more compute at test time" paradigm from reasoning to retrieval, which is relevant for agentic systems that can budget compute dynamically across retrieval and action steps.