Self-speculative decoding that reformulates draft model selection as a knapsack problem to maximize tokens-per-time throughput, accounting for context-dependent attention overhead - ___[ICML 2026](http://icml.cc/Conferences/2026)___
Proactive pitfall avoidance in plan-and-execute frameworks for long-context LLM reasoning by preventing logical pitfalls and false assumptions before plan generation - ___[ACL 2026](https://2026.aclweb.org)___
Membership inference attack for LLMs via an expectation-maximization algorithm that handles the inherent ambiguity of membership in pretraining data - ___[EACL 2026](https://2026.eacl.org)___
Adaptive listwise reranking that dynamically adjusts both the amount and target of computation via iterative Bayesian TrueSkill relevance estimation - ___[NeurIPS 2025](https://neurips.cc)___
Bilingual benchmark for evaluating LLMs on deep expert knowledge and complex academic problem-solving across abstraction, comprehension, and reasoning tasks - ___[Findings of EMNLP 2025](https://2025.emnlp.org/)___
Hierarchical RAG framework that compresses and partitions documents at multiple scales to address imprecise retrieval and fragmented context in long-context LLMs