# Agentic Search

Video anomaly understanding has long used fixed observation: a model receives a bundle of sampled frames or clustered events and then judges the anomaly, with no way to go back for missing evidence once reasoning starts. SEEK-VAU recasts this as budget-bounded agentic search and verification, letting the policy gather evidence across the precursor, trigger and aftermath stages of an event chain and check sufficiency before closing the case. On its SEEK-Bench, built on a Qwen3-VL-8B backbone, it reaches 55.6 Temporal mIoU, 45.5 QA Accuracy and 43.7 Evidence F1, against 44.2, 37.7 and 28.1 for the fixed-observation Qwen3-VL baseline.

Video anomaly understanding has long used fixed observation: a model receives a bundle of sampled frames or clustered events and then judges the anomaly, with no way to go back for missing evidence once reasoning starts. SEEK-VAU recasts this as budget-bounded agentic search and verification, letting the policy gather evidence across the precursor, trigger and aftermath stages of an event chain and check sufficiency before closing the case. On its SEEK-Bench, built on a Qwen3-VL-8B backbone, it reaches 55.6 Temporal mIoU, 45.5 QA Accuracy and 43.7 Evidence F1, against 44.2, 37.7 and 28.1 for the fixed-observation Qwen3-VL baseline.

Search agents for multi-hop information seeking are trained on step-level judgments of each search action instead of only on final-answer correctness. RAG-Gym casts search as a Markov decision process and compares architecture, parameter optimization and action evaluation, combining a reasoning-reflection architecture, preference optimization and critic-guided inference into Re2Search++. Trained on HotpotQA, it beats outcome-supervised baselines on the unseen Bamboogle benchmark with higher retrieval recall at comparable or fewer search steps, and a critic trained on Llama-3.1-8B raises GPT-4o-mini MedQA accuracy.

Search agents for multi-hop information seeking are trained on step-level judgments of each search action instead of only on final-answer correctness. RAG-Gym casts search as a Markov decision process and compares architecture, parameter optimization and action evaluation, combining a reasoning-reflection architecture, preference optimization and critic-guided inference into Re2Search++. Trained on HotpotQA, it beats outcome-supervised baselines on the unseen Bamboogle benchmark with higher retrieval recall at comparable or fewer search steps, and a critic trained on Llama-3.1-8B raises GPT-4o-mini MedQA accuracy.

Project Greenhouse models LLM training as a directed property hypergraph and uses it to define what makes a model fully open and sovereign; as a first milestone, the authors pre-train a 3.29B-parameter causal language model from scratch on ClimbMix using only publicly available datasets, then fine-tune it on RLHN-250K with localized contrastive estimation to obtain the pointwise reranker Gaggle (base reranker). Its mean nDCG@10 exceeds all compared fine-tuned pointwise and listwise rerankers on TREC DL 2019–2023 and seven BEIR collections, with only GPT-6.1 Sol scoring higher on TREC DL.

Project Greenhouse models LLM training as a directed property hypergraph and uses it to define what makes a model fully open and sovereign; as a first milestone, the authors pre-train a 3.29B-parameter causal language model from scratch on ClimbMix using only publicly available datasets, then fine-tune it on RLHN-250K with localized contrastive estimation to obtain the pointwise reranker Gaggle (base reranker). Its mean nDCG@10 exceeds all compared fine-tuned pointwise and listwise rerankers on TREC DL 2019–2023 and seven BEIR collections, with only GPT-6.1 Sol scoring higher on TREC DL.

The study builds corpus-grounded QA data with intermediate evidence annotations and compares three retrieval signals (evidence coverage Cov, coverage with dependencies Cov-Dep, answer match AM) under scalar reward bonuses versus action-aligned local credit, training Qwen3-4B on seven QA benchmarks and finding local credit outperforms scalar, with Cov-local reaching 54.34 average F1 versus 51.25 for outcome-only training.

The study builds corpus-grounded QA data with intermediate evidence annotations and compares three retrieval signals (evidence coverage Cov, coverage with dependencies Cov-Dep, answer match AM) under scalar reward bonuses versus action-aligned local credit, training Qwen3-4B on seven QA benchmarks and finding local credit outperforms scalar, with Cov-local reaching 54.34 average F1 versus 51.25 for outcome-only training.

The authors introduce a two-stage training framework around colgrep, a local semantic search tool, combining supervised fine-tuning with turn-level credit based on retrieval outcomes and reinforcement learning on localization quality, and train three 1–1.7B-parameter models for file localization; on SWE-bench Lite and Multi-SWE-bench Flash, colgrep-equipped agents outperform matched grep agents at every training stage, with Qwen3-1.7B reaching 67.29 and 54.00 file F1 after weighted SFT plus RL, while using 29.1% fewer tokens and reducing end-to-end CPU latency by 44.1% and transferring better to seven languages unseen during fine-tuning.

The authors introduce a two-stage training framework around colgrep, a local semantic search tool, combining supervised fine-tuning with turn-level credit based on retrieval outcomes and reinforcement learning on localization quality, and train three 1–1.7B-parameter models for file localization; on SWE-bench Lite and Multi-SWE-bench Flash, colgrep-equipped agents outperform matched grep agents at every training stage, with Qwen3-1.7B reaching 67.29 and 54.00 file F1 after weighted SFT plus RL, while using 29.1% fewer tokens and reducing end-to-end CPU latency by 44.1% and transferring better to seven languages unseen during fine-tuning.

The work proposes LexiHorizon, a long-horizon deep-search training framework that expands the trajectory context budget to 128K, manages accumulated retrieval content with a context window that keeps only recent tool observations in full while preserving the reasoning history, and adds an outcome-gated search-effort reward that applies only to trajectories with nonzero answer reward, training Qwen3.5-9B online with GSPO; the resulting LexiHorizon-9B outperforms both its base model and MiroThinker-1.7-mini on XBench, WebWalkerQA, and BrowseComp-ZH, with maximum absolute gains of 8.7 and 23.8 percentage points, and ablations show that a 64K budget lowers the valid answer rate from 99.88% to 66.97% while removing the window reduces accuracy on all three benchmarks.

The work proposes LexiHorizon, a long-horizon deep-search training framework that expands the trajectory context budget to 128K, manages accumulated retrieval content with a context window that keeps only recent tool observations in full while preserving the reasoning history, and adds an outcome-gated search-effort reward that applies only to trajectories with nonzero answer reward, training Qwen3.5-9B online with GSPO; the resulting LexiHorizon-9B outperforms both its base model and MiroThinker-1.7-mini on XBench, WebWalkerQA, and BrowseComp-ZH, with maximum absolute gains of 8.7 and 23.8 percentage points, and ablations show that a 64K budget lowers the valid answer rate from 99.88% to 66.97% while removing the window reduces accuracy on all three benchmarks.

This work scales online reinforcement learning for multimodal deep-research agents to 128k context and 75+ tool-interaction turns and introduces Long-MDR, a three-component recipe of On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue, to address slow early reward growth and later entropy collapse in long-horizon training; at a 50-turn tool-interaction evaluation budget, the trained Long-MDR-9B ranks first on five of six benchmarks among the compared 7B–9B agents.

This work scales online reinforcement learning for multimodal deep-research agents to 128k context and 75+ tool-interaction turns and introduces Long-MDR, a three-component recipe of On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue, to address slow early reward growth and later entropy collapse in long-horizon training; at a 50-turn tool-interaction evaluation budget, the trained Long-MDR-9B ranks first on five of six benchmarks among the compared 7B–9B agents.