Reframe: Post-Training Framework Makes Visual Agents Actually Use What They See
_akhaliq · x · 2026-11-11
Reframe is a new post-training framework for visual agents addressing a core failure mode: evidence returned by tool calls (crops, frame selections) often goes unused, and answer-only rewards let shortcuts skip the returned view.
- Method: Grounds rewards on the tool call itself and repairs observations during distillation, with expert supervision ensuring answers are conditioned on valid views; offers specialized GD-GROPO and OG-MOPD variants.
- Results: Beats prior art on 14 benchmarks without an SFT cold start.
Related event: Reframe: Post-Training Framework Aligns Visual Agents' Observations(2 posts)→
More from Research
- MKA benchmark results: top models hit 28-57% on hard agentic visual tasks — ArtificialAnlys · 2026-11-11
- MKA cost breakdown: Claude Fable 5.2 spends $68 and 3 hours on a task Gemini 4 Carbon does for $8 — ArtificialAnlys · 2026-11-11
- MKA rubric design: top seven models cluster at 95–97% on Regression but span 28–57% on Climb — ArtificialAnlys · 2026-11-11
- MKA by output type: Fable 5.2 leads four file types, DeepSeek V4.1 Flash tops Text Files — ArtificialAnlys · 2026-11-11
- Inside an MKA task: auditing building plans costs Claude Fable 5.2 224 turns and ~$68 — ArtificialAnlys · 2026-11-11
- MKA by vertical: no single model dominates; Fable leads five industries, Gemini takes Healthcare and Science — ArtificialAnlys · 2026-11-11