Soft Spatial Reasoning
The paper argues that LVLMs make spatial reasoning mistakes by committing too early to one text token.
Soft Spatial Reasoning keeps intermediate reasoning in a continuous “soft” state by mixing token embeddings instead of choosing a single next token. Its AdaptSoft controller changes how soft that state should be at each step, using the hidden state and predictive uncertainty. The authors train that controller with a gradient-alignment objective that does not require supervised intermediate reasoning. They report gains over hard and fixed-soft chain-of-thought baselines on spatial benchmarks using the same backbone. HF Daily Papers' note
Soft Spatial Reasoning keeps intermediate reasoning in a continuous “soft” state by mixing token embeddings instead of choosing a single next token. Its AdaptSoft controller changes how soft that state should be at each step, using the hidden state and predictive uncertainty. The authors train that controller with a gradient-alignment objective that does not require supervised intermediate reasoning. They report gains over hard and fixed-soft chain-of-thought baselines on spatial benchmarks using the same backbone. HF Daily Papers' note
score 4