AI Paper Digest · October 28, 2026

This digest covers three papers from the October 26 arXiv wave: DecisionRAG, which replaces the retriever with decision models; Auto-Research Arena, a benchmark that both evaluates and advances autonomous research; and FOCUS, active critic learning for credit assignment in long-horizon agents.

1. Retriever-Free Retrieval: Navigating Documents with Decision Models

  • Title: Retriever-Free Retrieval: Navigating Documents with Decision Models
  • Authors: Chao Wan, Paul Jünger, Beixi Du, Bodong Liu, Navaneetha Krishnan, Kilian Q. Weinberger
  • Problem: Language models have grown dramatically more capable, yet retrieval still relies largely on lexical matching or similarity in fixed embedding spaces — creating a capability asymmetry: what the retriever fails to find, even the strongest generator cannot recover. How can retrieval itself acquire the understanding of a language model?
  • Method: DecisionRAG brings pretrained language model capabilities to the retrieval side: decision models jointly evaluate candidate branches in a single forward pass, without generating any tokens. The key insight is that judging candidates in a shared context makes their relevance estimates comparable. Like a person browsing a library, DecisionRAG navigates the corpus’s inherent structure — from broad topics down to individual passages — deciding at each level which branches are worth exploring. No task-specific training and no first-stage retriever required.
  • Results: Without task-specific training or a first-stage retriever, DecisionRAG outperforms dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions. Across four benchmarks, including a corpus of 2.68 million Wikipedia paragraphs, retrieval takes just 0.7–1.8 seconds per query.
  • arXiv: https://arxiv.org/abs/2610.35868

2. Auto-Research Arena: Benchmarking and Advancing Auto-Research in Language Modeling

  • Title: Auto-Research Arena: Benchmarking and Advancing Auto-Research in Language Modeling
  • Authors: Xi Chen, Chongjian Ge, Yiran Xu, Yiqun Mei, Henrik I. Christensen, Hao Tan, Ziwen Chen, Hanwen Jiang
  • Problem: As autonomous research systems become more capable, benchmarks must do more than test — they should support meaningful research progress. How can different research objectives be compared and discovered within a shared experimental environment?
  • Method: Auto-Research Arena, a multi-target benchmark that decouples admissibility from improvement: model quality and resource constraints determine whether a result qualifies, while the selected research target measures progress among admissible results. The authors instantiate Arena in language model pretraining with nine targets spanning model, data, training, and inference efficiency, and evaluate seven research harnesses using two foundation models.
  • Results: Arena clearly differentiates existing harnesses across research objectives — and several sophisticated methods are outperformed by a simple sequential search baseline, indicating substantial headroom in current auto-research systems. Beyond evaluation, Arena reveals a broader frontier of model and system ideas aligned with active research directions; reusing these discoveries in a new quality-driven search yields lower training loss under matched data and FLOPs, and lower held-out BPB at the same model size. Code is open-sourced (github.com/Auto-Research-Arena/auto-research-arena).
  • arXiv: https://arxiv.org/abs/2610.35866

3. FOCUS: Active Critic Learning for Long-Horizon Language Model Agents

  • Title: FOCUS: Active Critic Learning for Long-Horizon Language Model Agents
  • Authors: Zhengping Jiang, William Jurayj, Liaoyaqi Wang, Chunsheng Zuo, Anqi Liu, Benjamin Van Durme
  • Problem: Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. Learned critics provide value estimates for intermediate states, but those estimates can be inaccurate, while estimating intermediate values with repetitive continuation sampling is expensive. Can a critic learn to be accurate exactly where it matters most?
  • Method: FOCUS, an RL method that uses selective branching to help the critic learn its value estimator. The rollout controller enables active learning for the critic by sampling continuations only from states expected to most reduce the critic’s estimation error; the refined value estimates then score the segment between consecutive states with a TD(λ) advantage, while the sampled evidence supplies training targets for the critic. FOCUS also supports long-horizon agentic RL through a modular context-management interface that decouples context management from rollout allocation.
  • Results: With matched per-step generation budgets, FOCUS outperforms GRPO and VinePPO on legal reasoning, deep search, and code localization benchmarks: at Qwen3.5-9B scale it improves held-out accuracy by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs 48.2) and by 4.1 points on DeonticBench (56.5 vs 52.4).
  • arXiv: https://arxiv.org/abs/2610.35801