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

OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 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 OneSearch-V2, a generative e-commerce search framework that adds keyword-based CoT query understanding, reasoning-internalized self-distillation (shared-weight teacher-student with R-Drop, FGM, focal loss), and TPMA-GRPO behavior-feedback alignment, validated offline on Kuaishou mall search and by online A/B tests.

Problem

Generative search models like OneSearch handle complex and long-tail queries shallowly, fail to infer personalized latent intent without costly explicit CoT, and suffer from reward-model-based reward hacking and historical-log overfitting in e-commerce search.

Approach

Keyword-based CoTs generated offline by an LLM serve as privileged input to a teacher that shares weights with the student; the student is trained to match the teacher's output distribution without keywords (self-distillation), and RL uses direct behavior feedback with a token-position marginal advantage plus a prefix gate instead of a separate reward model.

Key results

• Online A/B test: OneSearch-V2 gives +3.98% item CTR, +1.17% PV CTR, +2.90% PV CVR, +2.07% buyer volume, +2.11% order volume vs OneSearch-V1.
• Offline order HR@10 rises from 0.2046 (OneSearch baseline) to 0.2314 for OneSearch-V2 (MRR@10 0.0985 to 0.1151).
• Manual evaluation: +1.37% page good rate and +1.65% query-item relevance.

Surprising findings

• Explicit textual CoT before SID generation substantially degrades performance (order HR@10 0.0898 vs 0.2046 baseline), while keywords injected at input help.
• After self-distillation, the student without keywords slightly surpasses the keyword-augmented teacher (Self-Distill (S) 0.2163 vs (T) 0.2155 order HR@10).

Limitations

• Explicit CoT degrades SID generation and inference-time keyword CoT adds latency, so gains rely on offline keyword generation pipeline.
• Keyword coverage may limit recommendations to items explicitly covered by keywords.
• Evaluation largely on a single proprietary industrial platform with offline data from one company.

Experimental setup

Datasets: Kuaishou mall search interaction logs

Benchmarks: Kuaishou mall search offline test (7,229 order, 30k click PVs)

Metrics: HR@10 MRR@10 SID rate Item CTR PV CTR PV CVR Buyer volume Order volume page good rate item quality

Baselines: OneSearch (V1) CODI EMA-mode self-distillation joint-mode distillation special-token distillation listwise DPO GRPO

Builds on: OneSearch CODI SDFT GRPO R-Drop OneRec-V2

Closest papers by meaning

• Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers (NeurIPS 2025 · similarity 0.87)
• Scent of Knowledge: Optimizing Search-Enhanced Reasoning with Information Foraging (NeurIPS 2025 · similarity 0.86)
• AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play (NeurIPS 2025 · similarity 0.86)
• DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.86)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.85)
• Learning to Rank in Generative Retrieval (AAAI 2024 · similarity 0.85)