Iris Pushes Search Agents Forward with SFT-RL Climbing

Introduction

A capable search agent must do more than retrieve a relevant page. It needs to decompose a question, follow evidence across several entities and links, decide when another search is necessary, and preserve useful observations without exceeding its context budget. The paper Iris: Climbing to the Search Frontier presents Iris-mini and Iris-pro, together with a training pipeline designed specifically for these long-horizon behaviors.

How Iris is trained

• Tasks are reconstructed from the web graph. The researchers start with a seed page and its outgoing links, then build an entity graph and author multi-hop chains. Non-answer entities are rewritten as descriptive references, making it difficult to solve a question through direct name matching.
• Closed-book solvability is used as a filter. A task is admitted only when a reference model fails without evidence but succeeds after the supporting information is supplied. This selection rule is intended to separate genuine search problems from questions that a language model can answer from memory.
• Search trajectories are filtered at two levels. Generated browsing traces are screened as complete trajectories and again at the individual-turn level before supervised fine-tuning. The resulting policy is then optimized with reinforcement learning against live search. The reward judge and observation summarizer run inside the training cluster.
• SFT and RL alternate in a climbing loop. Hard, successful, and efficient rollouts from each reinforcement-learning round are returned to the next supervised-training pass. The process is meant to preserve useful search routines while gradually expanding the difficulty of behaviors the policy can learn.
• Context management is part of the agent design. When a rollout becomes too long, it is interrupted at the request level and resumed from its committed prefix. The authors evaluate each benchmark with and without this management while keeping the tools, context limit, and judge fixed.

Reported results and implications

With context management enabled, the paper reports scores of 82.2, 84.8, 86.9, and 52.3 for Iris-mini on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE, respectively. Iris-pro reports 88.6, 85.1, 92.9, and 56.4 on the same benchmarks. The results come from a single ReAct agent; the system does not use sub-agents or test-time verification.

The main contribution is not simply a larger search model. Iris treats task construction, trajectory quality, policy optimization, and context handling as one training problem. That framing is significant because a long search chain can fail even when the model has adequate factual knowledge: an irrelevant observation, a poorly chosen query, or a lost intermediate result may break the entire chain.

The reported comparisons also suggest that inference-time state management deserves more attention in search-agent evaluations. Still, the supplied material does not establish how the models perform under changing search tools, different latency budgets, or deployment-scale costs. Reproducibility will depend on the promised release of model weights and the complete data, training, and evaluation recipe.

Source: Hugging Face Daily Papers