Soft Spatial Reasoning
Soft Spatial Reasoning lets vision-language models mix token embeddings instead of committing to one spatial token.
Soft Spatial Reasoning is a post-training framework for spatial tasks in large vision-language models that avoids committing to a single discrete token at every chain-of-thought step. At each intermediate step the model forms a continuous soft state by mixing token embeddings so several candidate continuations can influence the next step. AdaptSoft uses the current hidden state and predictive uncertainty to choose how soft each step should be, trained with a gradient-alignment objective that does not require intermediate reasoning supervision. On spatial benchmarks it outperforms hard and fixed-soft chain-of-thought using the same backbone, as well as a range of existing vision-language models.
- Soft states mix token embeddings instead of picking one token.
- AdaptSoft sets softness from the hidden state and uncertainty.
- Training uses gradient alignment without intermediate reasoning labels.
- It beats hard and fixed-soft chain-of-thought on the same backbone.
From the original article
1 min read at huggingface.co
Large Vision-Language Models (LVLMs) commonly perform spatial reasoning through chain-of-thought (CoT), encoding intermediate reasoning as autoregressive sequences of discrete language tokens. Such hard thinking requires committing to a single token at each step, even when the correct spatial interpretation remains uncertain.
By Rafi Ibn Sultan, Md. Sajid Alam Chowdhury, Saleh Zare Zade · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.