Iris-pro at 397B-A17B posts 88.6 BrowseComp and 92.9 DeepSearchQA under fixed ReAct context limits
Iris agents demonstrate that reverse-constructed multi-hop search trajectories plus alternating SFT-RL produce state-of-the-art open-source performance on BrowseComp and HLE. Context management at inference time contributes more to scores than model scale alone. Planned weight and recipe release enables external audit of faithfulness claims.
Iris-mini (35B-A3B) and Iris-pro (397B-A17B) were trained on multi-hop trajectories reverse-engineered from web hyperlink graphs. Questions were filtered so reference models fail closed-book yet succeed with evidence supplied. Training alternated SFT on filtered rollouts with RL against live search, using an internal reward judge and observation summarizer; over-long trajectories were truncated at the request level and resumed from prefix state.
With tool set, context window and judge held constant, Iris-pro outperforms prior open-source agents on all four benchmarks when context management is active. The gap between managed and unmanaged runs exceeds most inter-system differences cited in earlier work, confirming that inference-time context handling accounts for the largest share of measured gains. All results derive from a single ReAct policy without sub-agents or test-time verification.
The method directly targets answer faithfulness by requiring models to retrieve and integrate evidence rather than rely on parametric recall. Releasing the full data-construction pipeline, training code and weights will allow independent verification of the 397B-A17B checkpoint and replication of the SFT-RL climbing loop on other corpora.
Iris-pro: Surpasses 60 on HLE with context management by March 2027
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.04304)
- [2]Supporting Source(https://arxiv.org/abs/2210.03629)