Qwen model learns to filter contaminated retrieval results with selective adoption training
Researchers have identified a critical vulnerability in retrieval-augmented LLMs: the inability to distinguish trustworthy evidence from contaminated or adversarial content in search results. This work introduces SelectBench, a benchmark designed to train models to selectively adopt relevant information while rejecting deceptive material. Using reinforcement learning with DAPO on Qwen3.5-4B, the team improved strict success rates from 22.46% to 26.46%, addressing a fundamental reliability gap in production RAG systems. The advancement matters because real-world retrieval often surfaces mixed-quality sources, and current models either discard valid data or absorb harmful content uncritically.
Modelwire context
ExplainerThe paper frames evidence rejection as a learnable skill via reinforcement learning rather than a filtering problem solved at retrieval time. The 4-point improvement is modest, but the framing matters: models can be trained to distrust bad sources instead of relying on retrieval ranking or post-hoc fact-checking.
This work sits alongside two parallel robustness vulnerabilities we've covered. The linguistic realization paper from July 22 showed models conflate surface form with meaning, and ENTRAP-VL demonstrated how auxiliary signals inappropriately influence outputs regardless of relevance. SelectBench addresses a related failure: models absorbing contaminated evidence without skepticism. All three expose how current architectures lack principled mechanisms to weight or reject information based on trustworthiness rather than salience. Where those papers diagnosed the fragility, this one offers a training-time intervention.
If SelectBench performance holds when evaluated on out-of-distribution contamination types (adversarial rewrites, domain shifts) that weren't in the training set, the approach generalizes. If the gains collapse on held-out contamination patterns, the model has memorized SelectBench artifacts rather than learning robust rejection. Watch whether Alibaba or other groups report reproduction results on proprietary RAG pipelines within six months.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsQwen3.5-4B · SelectBench · DAPO · Alibaba
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Reinforcement Learning for Large Language Model Selective Evidence Adoption from Contaminated Retrieval Results”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.