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

TFRank: Think-Free Reasoning Enables Practical Pointwise LLM Ranking

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in 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 TFRank, a pointwise reasoning reranker trained with multi-task SFT on CoT-distilled data that switches to a no-think mode at inference, matching much larger rerankers on BRIGHT and BEIR.

Problem

Reasoning-intensive LLM rerankers rely on large models and explicit chain-of-thought generation, which makes ranking costly and slow for real-world deployment.

Approach

Train small LLMs with multi-task SFT (pointwise, pairwise, listwise) on DeepSeek-R1 CoT and fine-grained labels, then use a /no think switch at inference to output only a pointwise relevance score fused from yes/no probability and expected 0-4 score, so CoT is used only as training supervision.

Key results

• TFRank-1.7B (Qwen3, SFT) averages 16.7 NDCG@10 on BRIGHT with BM25, comparable to Rank1-7B (18.5), REARANK-7B (17.7) and Rank-R1-7B (16.4).
• TFRank-8B (Qwen3) GRPO reaches 32.0 NDCG@10 on BRIGHT with ReasonIR-8B retrieval, versus 24.2 for REARANK-7B.
• Forcing explicit thinking at inference drops TFRank-0.6B from 15.6 to 10.5 average NDCG@10 on BRIGHT, so no-think outperforms think mode.

Surprising findings

• Think-free inference outperforms explicit reasoning-chain inference, with forced thinking reducing NDCG@10 from 15.6 to 10.5 on BRIGHT.
• Direct GRPO on Qwen3 backbones gives larger gains than SFT or SFT+GRPO in the reported setting.

Limitations

• Full GRPO training is costly (about 200 hours for the 0.6B model).
• Gains from GRPO on smaller models are small or inconsistent in some settings (e.g., SFT+GRPO underperforms SFT on some Qwen3 sizes).
• Training data relies on DeepSeek-R1 annotations and MS MARCO, and evaluation is on BRIGHT, BEIR and R2MED.

Experimental setup

Datasets: MS MARCO BGE-M3 data BRIGHT BEIR R2MED

Benchmarks: BRIGHT BEIR R2MED

Metrics: NDCG@10

Baselines: BM25 RankGPT-4 Rank1-7B Rank1-14B REARANK-7B Rank-R1-3B Rank-R1-7B Rank-R1-14B Deepseek-R1-rank_llm

Builds on: Rank1 REARANK Qwen3 GRPO BGE-M3 fine-grained relevance

Closest papers by meaning

• A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.87)
• Accelerating Large Language Model Reasoning via Speculative Search (ICML 2025 · similarity 0.87)
• GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · similarity 0.87)
• Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization (ICLR 2026 · similarity 0.87)
• RATT: A Thought Structure for Coherent and Correct LLM Reasoning (AAAI 2025 · similarity 0.87)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.87)