AIAny
Icon for item

Iris: Climbing to the Search Frontier

Presents two LLM-based search agents (Iris-mini and Iris-pro) trained by alternating supervised fine-tuning and reinforcement learning against live web search. Key features: web-graph-derived multi-hop tasks with entity abstraction, SFT–RL climbing, inference-time context management, and state-of-the-art open-source benchmark results.

Introduction

Why this matters

Search agents must decide what to search, how to interpret evidence, when to continue, and when to stop — tasks that require long-horizon planning and robust tool use beyond closed-book reasoning. This work shows a practical pipeline that converts web hyperlink structure into hard, verifiable multi-hop questions, removes direct string anchors so agents must reason to disambiguate, and trains policies by alternating supervised fine-tuning and reinforcement learning against live search.

Key Findings
  • A data pipeline that reverse-constructs multi-hop questions from web subgraphs and applies an abstraction operator to rewrite non-answer entities into descriptive references, which prevents shortcut string matching and forces compositional reasoning. So what: training tasks better reflect realistic search challenges and reduce trivial retrieval shortcuts.

  • An iterative SFT–RL “climbing” procedure: RL rollouts against live search are used to discover high-quality trajectories, a filtered subset of which is fed back into supervised fine-tuning; repeating this loop amplifies rare but effective behaviors. So what: the method combines exploration-driven improvement with targeted imitation of successful long-horizon strategies.

  • Strong empirical open-source results: with inference-time context management enabled, Iris-mini (35B-A3B) and Iris-pro (∼397B-A17B) achieve leading scores on complex web benchmarks — e.g., BrowseComp, BrowseComp-ZH, DeepSearchQA and HLE — demonstrating substantial gains from both the training pipeline and context management.

Who it’s for and trade-offs

Great fit if you need a reproducible recipe for training search-capable agents that must compose evidence across pages and manage long contexts — researchers building web-grounded agents, teams developing tool-using LLMs, and benchmark-driven labs.

Look elsewhere if you need a lightweight agent: Iris-pro requires very large MoE-scale compute and long context infrastructure; the full training recipe and model weights are promised for release but reproducing the exact setup demands significant engineering resources. Also, the approach focuses on single-agent ReAct-style behavior and does not rely on multi-agent verification at inference time, which may matter for high-assurance deployments.

Where it fits

The paper sits at the intersection of RL-for-LLMs, retrieval-augmented reasoning, and benchmark engineering: its main contribution is operational — a scalable task-construction + SFT–RL training loop that emphasizes long-horizon search behavior and careful trajectory filtering rather than a novel model architecture alone.

Information

  • Websitearxiv.org
  • OrganizationsAllSpark Research
  • AuthorsZiyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, Xu Chu, Shaowei Chen, Yuantao Gu, Mu Chuan
  • Published date2026/09/03

More Items

Evaluates layer dropout (stochastic depth) in LLM pretraining, finds optimal layer distributions and time schedules that save ~20–25% training FLOPs at comparable or lower validation loss, and enables zero-shot layer skipping and self-speculative decoding for up to ~1.4–1.5× inference speedup.

Studies when and how an assistant should ask clarification questions before converting natural-language problem descriptions into optimization models. Introduces OR-Clarify, a benchmark for pre-formulation clarification, and InterOPT, a two-stage framework that diagnoses formulation-critical gaps and guides targeted questioning and stopping decisions.

Decides when past post-training updates should be reused for autonomous LLM adaptation by introducing Boundary-Calibrated Intervention Transfer (BCIT). BCIT binds effects to source context, checks applicability and hard conflicts, and runs bounded trials to obtain current-state evidence—reducing harmful updates and improving equal-budget final-model quality.