ActICL trains MLLM agents to build their own ICL contexts, +11.6 points over retrieval
EliasEskin · x · 2026-10-16
A UT Austin/Rutgers paper introduces ActICL, casting multimodal ICL context construction as an agentic task: the model analyzes its own information gaps, retrieves and curates demonstrations via tool calls, and iterates until the context suffices or the token budget runs out.
Key results:
- Trained with rejection-sampling SFT followed by joint RL over acquisition and answering.
- Across 7 benchmarks and 3 backbones: +4.4 points over the strongest trained baseline and +11.6 points over standard similarity retrieval, using only 3.5 demonstrations on average.
- Contexts transfer: a 9B ActICL agent builds contexts that lift GPT-5.6-Sol's 4-shot ICL by +3.5 points.
The core idea: let the model decide what to look for, what to keep, and when it has enough — instead of one-shot similarity scoring.
More from Models
- Analyst: in a year, models 2-3x better across the board at 10x lower cost, token use up 10-20x — DeryaTR_ · 2026-10-16
- Google revamps Gemma docs, site teases Gemma 4 open models — rseroter · 2026-10-16
- Respan's Span-02 brings anytime reasoning: interrupt it mid-thought and still get the best answer — techNmak · 2026-10-16
- Liquid AI ships d1-ultra decision model, tops Decision Index at 65 across 112 evals — maximelabonne · 2026-10-16
- Meta releases Muse Spark Open: 816B/42B active, 1M context, most capable US open-weight model — AIatMeta · 2026-10-16
- Scale AI's open muse spark model hits seven API platforms, runs on vLLM, SGLang, MLX — alexandr_wang · 2026-10-16