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.
Yanyu Chen, Yue Li, Yongyi Cui et al.· arXiv.org· 1 citation
This work proposes eXtreme Gradient Boosting LoRA (XGBLoRA), a novel framework grounded in gradient boosting theory that provides theoretical analysis establishing convergence guarantees and expressiveness bounds, which formally justify why weaker (lower-rank) adapters, when properly combined, can match or exceed the performance of stronger ones.
Yifei Zhang, Hao Zhu, Haoran Shi et al.· Proceedings of the 32nd ACM...· 0 citations
Large language models (LLMs) increasingly handle in-context learning (ICL) tasks where a long, novel context defines the rules, knowledge, and output schema for a series of questions. On benchmarks that grade against every detail of the context, even strong open-weights models pass only 12-16% of tasks: a single overlooked rule fails the whole response. We argue this brittleness is structural: the dominant"read-and-reason"paradigm asks the model to extract, plan, generate, and self-verify in one forward pass. We therefore ask whether explicit context compilation can fix it, how it compares to existing long-context strategies (gist retrieval, multi-agent self-play), and where the resulting harness benefit holds across task structure and model scale. We propose the Context Compilation Architecture (CCA), whose central novelty is a typed intermediate representation (IR) with fixed slots (rules.{must_do, must_not, conditional}, output_spec, available_tools, data_profile) into which any prose context is compiled once; executable verifiers and a violation-gated correction loop follow as downstream consequences. On CL-bench (1,899 tasks across 4 open base models), CCA outperforms vanilla prompting and two long-context baselines (ReadAgent-P, Ctx2Skill) on every base model, lifting Kimi K2.5 from 15.4% to 21.4% with gains concentrated on rule-dense sub-categories. Code and cached completions are available at https://github.com/TonyQJH/cca-emnlp2026.
Jin-Hu Qi, Minda Hu, Wen-Tao Zhang et al.· 0 citations
Fine-tuning Large Language Models (LLMs) has become a crucial technique for adapting pre-trained models to downstream tasks. However, the enormous size of LLMs poses significant challenges in terms of computational complexity and resource requirements. Low-Rank Adaptation (LoRA) has emerged as a promising solution, yet a gap remains between the practical performance of low-rank adaptations and their theoretical optimum. While recent works have explored iteratively merging LoRA adapters, they lack a principled theoretical framework to guide adapter design. In this work, we propose eXtreme Gradient Boosting LoRA (XGBLoRA), a novel framework grounded in gradient boosting theory. Our key insight is that the adapter must adhere to the weak learner principle-each individual adapter should have limited expressiveness---to ensure that the iterative ensemble can effectively raise the model's performance ceiling without overfitting. We provide theoretical analysis establishing convergence guarantees and expressiveness bounds, which formally justify why weaker (lower-rank) adapters, when properly combined, can match or exceed the performance of stronger (higher-rank) ones. Extensive experiments on natural language processing tasks demonstrate that XGBLoRA with rank-1 updates consistently outperforms standard LoRA with significantly fewer trainable parameters.
Yifei Zhang, Hao Zhu, Haoran Shi et al.· Proceedings of the 32nd ACM...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
