- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper addresses embodied reasoning for robotic manipulation, where a vision–language model must infer a physically feasible manipulation plan from a scene observation and an instruction. Existing methods employ outcome-only rewards, which cannot verify whether the reasoning has accounted for the physical constraints behind a feasible plan. The key idea is to use simulator-verified successes and near-failures as process supervision beyond correct-answer rewards. To this end, the paper proposes Contrastive Embodied Reasoning (CER). For each training task, CER builds a fixed library of encoded simulator observations, consisting of a centroid of 32 successful states and 32 individually kept near-failures. A bilinear critic scores every reasoning-token hidden state against this library with an InfoNCE-style log-probability, and the token-averaged score is added to rewards in GRPO training. The evaluation uses Qwen2.5-VL-7B on CRD-Bench, a self-constructed 30-class, four-choice RLBench-derived benchmark which is stratified by physical-constraint count.
Critical strengths
- The problem is important and well motivated. In embodied reasoning, outcome-only rewards cannot verify whether a model has checked the physical constraints behind a feasible plan. They are also extremely sparse on multi-constraint tasks. Using simulator successes and near-failures as process supervision is a reasonable direction.
Critical weaknesses
-
The contrastive design contributes little beyond generic process supervision. The paper’s own ablations show that non-contrastive process rewards using the same library already obtain most of the gain (41.1–42.9% vs. 45.8% for CER). The contrastive-specific gain is only 2.53 points, and it vanishes in executed planning (35.3% vs. 34.3%).
-
The rationale of the method is not adequately analyzed or validated. The paper does not justify why matching reasoning-token hidden states to fixed visual embeddings could capture physically grounded reasoning. It never validates the score itself. Appendix C acknowledges that token-level scores are not reported, and the probing and intervention analyses examine the trained model rather than the reward. It therefore remains unclear whether the gain stems from the claimed mechanism.
Questions for authors
-
Do the non-contrastive baselines in Table 1 differ from CER only in joint normalization, and were they tuned comparably? Since the contrastive gain vanishes in execution, what shows that near-failure contrast drives the improvement?
-
How is the critic trained, and does the process score correlate with answer correctness or rationale validity? Can the authors provide a theoretical analysis of why matching reasoning-token hidden states to a fixed library of visual embeddings should reward physically grounded reasoning
-
Can the authors report results on newer and larger backbones beyond Qwen2.5-VL-7B? Do the gains persist when stronger base models already face less reward sparsity?
-
Do the gains hold on an established embodied benchmark not built from the same constraint taxonomy? For EmbodiedBench and SpatialRGPT, which are currently reported in a single sentence, can they provide full settings and uncertainty, and compare against the strengthened terminal baseline and the process-supervision baselines?
-
How are near-failures constructed and mined, and how is exposure matched in Section 4.1? How many seeds support Table 1 and Figure 3?