ACL2025 Wu: RankCoT refining knowledge for retrieval augmented generation through ranking CoT
Key insight: generally a bunch of chain of thoughts, including on irrelevant documents, re-rank using self reflection, and then DPO
ACL2025 Trienes: behavioral analysis of information salience
Key insight: you can ask models for summaries at shorter lengths, which distill what the models think is salient information
ACL2025 Abbes: small encoders can rival large encoders in detecting groundedness
Key insight: apparently groundedness classification doesn’t require that many parameters
ACL2025 Luo: rethinking diverse human preference learning
Key insight: learn a Bradley terry reward model, PCA it’s embeddings, turns out those are interpretable and matches human reward differences
ACL2025 Chouayfati: GenDLN
Key insight: Used genetic algos for Prompt mixtures to iteratively improve prompt quality
ACL2025 (SRW): Sakunkoo
Key insight: data for defective verbs and trained tagger
ACL2025 Long: reinforcing compositional retrieval
Key insight: auto progressive embedding retriever to be able to get sequential decision tasks for RAG
ACL2025 Shelton: PQB-EQA
Key insight: benchmark for vision grounding context attribution using Minecraft
ACL2025 Green: BabelEdits
Key insight: a cross linguistic model editing benchmark
ACL2025 Ichihara: theoretical guarantees for minimum Bayes risk
Key insight: easy Monte Carlo minimum Bayes risk decoding we can achieve closer to theoretical optimality for LM decoding