Iris search agents hit open-source SOTA via SFT-RL climbing at 35B and 397B scales
_reachsumit · x · 2026-09-07
Researchers released Iris-mini and Iris-pro, open-source search agents trained at 35B-A3B and 397B-A17B scales, claiming SOTA among open search agents.
Key method:
- Task construction: multi-hop questions reverse-engineered from hyperlink structure — chains authored over an entity graph, non-answer entities rewritten into descriptive references so string matching can't cheat; only questions a reference model fails closed-book but solves with evidence are kept.
- Training: trajectory- and turn-level filtering before SFT, then RL against live search with the reward judge and observation summarizer served in the training cluster.
- SFT-RL climbing: alternating stages where the hardest-solved and most efficient rollouts from each RL round feed the next supervised pass.
Evaluation: the paper stresses that inference-time context management matters more than most reported system differences, so every benchmark is evaluated with and without it; results come from a single ReAct agent with no sub-agents or test-time verification.
Related event: Iris: Open-Source Search Agents Hit New SOTA(2 posts)→
More from Research
- Researcher's blunt advice: cancel ICLR 2027 and restore conference prestige — hyhieu226 · 2026-09-21
- Dev clones Jev by LoRA-tuning Qwen3.5 4B on 25M synthetic tokens for $2 GPU hours — nato_nob · 2026-09-21
- Jitendra Malik: robots ignoring 3D structure are wasting a valuable signal — JitendraMalikCV · 2026-09-21
- Biological Millennium Problems Should Be Specific, Verifiable Datasets, Argues Poster — owl_posting · 2026-09-21
- Gzip, RE-PAIR Grammar Induction and the Case That Deep Learning Is Just the Cerebellum — ryunuck · 2026-09-21
- Encoder-style classification gets hot again: one multimodal BERT solved 1,000+ classification tasks — cwolferesearch · 2026-09-21