AGENTIC · 2026-10-21
Agentic — Paper Daily, 2026-10-21
Community search over short UGC posts, questions, and comments; optimize search relevance (nDCG) and CTR. Focus on semantic matching, query understanding, retrieval, ranking, and LLM-assisted search.
4.70
LEAD STORY ·
RUC /
NUSRUCNUS OpenSaC lets research agents write programs that combine search, reading, reranking, and local processing, returning selected evidence to the model. On BrowseComp, it cuts model-token use by 37.8%.
Key insight: Process intermediate evidence in code and preserve source links for reuse. The library reduces repeated model reading, but requires an execution environment. WideSearch row F1 falls slightly, so efficiency gains do not guarantee better answers.
4.53
RUCAAnt Group EG-GRPO rewards the first search step that finds support for a required claim, then compares actions from similar evidence states. The 4B agent reaches 45.9% accuracy on GAIA.
Key insight: Reward newly acquired evidence rather than repeated citations. Conditioning comparisons on what an agent already knows reduces unfair comparisons between early and nearly finished trajectories; fallback grouping handles sparse state matches.
4.53
ClueFlow links browser actions to research-state updates and the final answer. It uses those attribution chains for supervised learning and turn-level RL. In a controlled comparison, selected supervision matches all-turn SFT with about 30% of its training tokens.
Key insight: Record how evidence changes the research state. Verified local progress can remain useful even when the final answer is wrong, recovering training signals that whole-trajectory rejection would discard.
4.43
KAIST
MetaRetrieval Program Learning trains a model to compose executable search programs and learn from constrained mutations of unsuccessful programs. Its 4B model averages 51.5% exact-match accuracy across seven QA benchmarks.
Key insight: Separate retrieval breadth from context size: merge, deduplicate, and rerank locally, then expose only selected evidence. This allows broader evidence gathering without forcing the model to read every retrieved item.
4.35
HybridResearcher trains a 4B agent to connect database results with unstructured evidence. It constructs bridge trajectories around entities from validated single-source tasks and rejects examples solvable from either side alone. It improves pass@8 over its base model by 30.0 percentage points.
Key insight: Cross-source training needs genuine dependencies. A bridge entity should come from a seed answer, and removing either source should make the task unsolvable. This offers a concrete recipe for combining structured attributes with post-level evidence.
4.15
OpenVLHarness combines consistent tool interfaces with persistent, addressable multimodal evidence. Programs can reuse images, structured data, and web sources across steps. It improves average scores by 9.7–12.6 points over matched no-tool baselines.
Key insight: Tool observations need readable representations and stable references. Retaining the original visual artifact supports later verification and reuse, addressing missing evidence more directly than increasing the reasoning budget alone.
4.15
This study shows that asynchronous result timing changes what search agents retrieve next. Confluent Evidence Acquisition waits for each subquestion’s evidence and allocates shared search independently of arrival order, improving accuracy over early commitment by up to 12.6 percentage points.
Key insight: Scheduling changes retrieval behavior. Timing effects can persist after all initial evidence arrives, so evaluations should record delivery order, release conditions, and shared search budgets as well as the evidence set.
3.95
AmazonAWS fine-tunes a search agent to choose between lexical and vector search, rewarding nDCG@10 and penalizing turn or token overruns. On BrowseComp-Plus, nDCG@10 improves by 23.7% relative to the base model.
Key insight: Include failures in retrieval evaluation. Assigning failed tasks zero nDCG avoids evaluating only successful searches. A slight FreshStack regression also shows why cross-domain checks matter when selecting training data.
3.85
MemT-1.0 integrates updatable memory into model computation and expands capacity through Mixture-of-Storage. At a 128K-token history length, the Nano model delivers about 1.86× end-to-end inference speedup over a full-context KV-cache baseline.
Key insight: Persistent memory need not retain the full cache. Compact internal state changes what information survives, so evaluate fact recovery and general capabilities alongside latency and storage.
3.55
AdaExploreBench hides environment rules and changes them without announcement, testing whether agents gather evidence and reconsider stale assumptions. Across eight evaluated models, GPT-6 Sol achieves 61.9% calibrated accuracy.
Key insight: Reasoning from supplied evidence differs from discovering when evidence needs refreshing. Evaluate knowledge changes, stale observations, and re-exploration triggers explicitly; static QA tests miss these failures.