Search papers, labs, and topics across Lattice.
This paper introduces SelectBench, a benchmark designed to enhance selective evidence adoption in retrieval-augmented large language models (LLMs) by addressing the challenge of distinguishing useful information from misleading content. By post-training the Qwen3.5-4B model using DAPO with deterministic rule rewards and a frozen semantic judge, the authors achieved modest improvements in strict success rates on the SelectBench-v2 test set, increasing from 22.46% to 26.46%. Despite these gains, the results highlight the need for stronger reward shaping and additional training to achieve more substantial improvements in evidence adoption and injection resistance.
SelectBench reveals that while selective evidence adoption in LLMs can be improved, the gains are modest and highlight significant challenges that remain in ensuring robust performance.
Retrieval-augmented large language models frequently face contexts that interleave useful evidence with misleading statements or instruction-like content. Blanket refusal discards valid evidence, whereas uncritical adoption yields incorrect or unsafe answers. The ability to selectively adopt relevant information while rejecting deceptive or harmful content is therefore critical for reliable deployment in real-world retrieval settings. We introduce SelectBench, a controlled benchmark and training set for selective evidence adoption, and post-train Qwen3.5-4B directly with DAPO using either deterministic rule rewards or a frozen semantic judge. On the corrected 325-example SelectBench-v2 test set, strict success rises from 22.46% for the original checkpoint to 25.54% with DAPO-Rule and 26.46% with DAPO-DeepSeek. Both trained policies reduce forbidden-content adoption and produce shorter, more focused responses, yet prompt-injection following does not improve. The paired gains are modest and fail to survive Holm correction, suggesting that stronger reward shaping or additional training iterations may be needed for more robust gains. DAPO-DeepSeek exhibits no material degradation on MMLU or clean HotpotQA, indicating that the post-training procedure preserves general capabilities. These results demonstrate a directional improvement in selective evidence use, while identifying injection resistance and statistical robustness as important remaining challenges for future work.
