Soft Spatial Reasoning replaces hard CoT with adaptive continuous soft states for spatial tasks
Rafi Ibn Sultan · hf · 2026-10-01
LVLMs typically do spatial reasoning via chain-of-thought, committing to a single discrete token at each step—premature discretization that propagates errors.
Soft Spatial Reasoning is a post-training framework:
- Each intermediate step forms a continuous soft state by mixing token embeddings, letting multiple candidate interpretations influence the next step.
- Softness must vary per step: AdaptSoft, a controller driven by hidden state and predictive uncertainty, adapts the degree of softness; a gradient-alignment objective trains it without intermediate supervision.
Across spatial benchmarks it beats hard and fixed-soft CoT baselines on the same backbone plus a range of existing LVLMs. Code is open-sourced.
More from Research
- Custom CUDA SGEMM kernel matches cuBLAS, wins on 16384x16384+ matrices — goyal__pramod · 2026-10-02
- Mathematician Kevin Buzzard: math community is 'grieving' through five stages over AI — littmath · 2026-10-02
- AI runs DeepMind's consciousness calculator on itself: scores swing 0.10–0.72 by theory — izzycognita · 2026-10-02
- MINTEval, accepted at NeurIPS, shows LLM agents fail at tracking evolving contexts — EliasEskin · 2026-10-02
- MazeBench goes free online: agents prefer top-down views and stall in 3D puzzles — patience_cave · 2026-10-02
- Aleph Alpha Details MoE Pre-training Scaling: 30B-A3B on 16-512 B200s at 35% MFU — CatAstro_Piyush · 2026-10-02