- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes a novel multimodal latent reasoning framework, HCLR, which replaces explicit multimodal reasoning (text interleaved with intermediate images) with a short sequence of latent vectors. Instead of aligning latent vectors to visual features, it trains them so that the model's downstream outputs match those produced with the explicit reasoning ("functional equivalence"). After an SFT warm-up on the interleaved reasoning dataset Monet-SFT-125K, this is done in two rounds of distillation: first the intermediate images are replaced by latent vectors; then, the remaining textual reasoning is compressed into 8 latent vectors, so that at inference the model reasons from only the image and question.
HCLR is compared with vanilla SFT and visual latent reasoning methods (LVR, SkiLa, Monet) on V*, HRBench4K/8K, and MME-RealWorld-Lite, and on the out-of-distribution VisualPuzzles. It achieves the best overall scores among the compared latent reasoning methods on V*, HRBench4K, and MME-RealWorld-Lite, and reports results comparable to tool-based models (DeepEyes, PixelReasoner).
Critical strengths
-
The proposed approach, HCLR, is novel. Rather than aligning latent states to a predefined target representation, HCLR supervises them through the downstream behavior they induce, and progressively removes both intermediate visual observations and textual reasoning through a two-level distillation. This is a distinct direction compared to previous works.
-
HCLR shows competitive results against the compared baselines. It outperforms the compared latent reasoning baselines on three of four benchmarks and performs comparably to tool-based models
Critical weaknesses
W1. The core motivation is not validated. The paper claims that aligning latent states to representations does not guarantee they preserve the reasoning they replace, but never shows this. There is no example of alignment-based latents failing in this way, and no measurement of whether HCLR's latents preserve downstream behavior better (e.g., output agreement with the explicit-reasoning model). Fig. 4 shows only that the latents matter. HCLR itself also relies on alignment: removing the alignment loss hurts more than removing behavioral KD (79.06 vs. 81.68 on V*, Table 3). Only one baseline per alignment family is compared (LVR, Monet; see Q1). So the gains cannot be attributed to functional equivalence.
W2. The paper is hard to follow. The core idea is simple but buried under overlapping terms ("functional equivalence," "local/global functional equivalence," "functional induction," "functional latent learning"). Table 3 uses names never defined in the method ("Behavioral KD," "Latent Alignment," "CE Only"), so the ablation is hard to map to Eqs. 1-4. The same motivation is repeated across the Abstract and Sec. 1-3.
W3. Performance may be bounded by the teacher. Each stage distills from a teacher trained on Monet-SFT-125K, but no teacher accuracy is reported, so it is unclear how much each stage loses. See Q3.
Questions for authors
-
The paper contrasts HCLR with perceptual-alignment and privileged-representation-alignment methods but compares against only one of each (LVR, Monet). Could the authors include more recent methods they cite that use the same Qwen2.5-VL-7B base, such as VaLR (Jeon et al., 2026), CoVT (Qin et al., 2025), and UniVLR (Jiang et al., 2026), under the same protocol? These report higher numbers on overlapping benchmarks (e.g., VaLR 86.4 on V*; UniVLR 73.3 / 68.8 on HRBench4K/8K).
-
Could the authors report (1) training cost (e.g., GPU hours per stage and in total) compared with LVR, SkiLa, and Monet, and (2) inference cost (latency, generated tokens, or FLOPs per sample) compared with both latent (LVR, Monet) and tool-based (DeepEyes, PixelReasoner) models?
-
What is the accuracy of the explicit-reasoning model (M_warm) and of each intermediate teacher? This would show how much each distillation stage loses and whether HCLR is bounded by its teacher.
-
Table 2 includes only LVR among the latent baselines and no tool-based models. How do Monet, SkiLa, DeepEyes, and PixelReasoner perform on VisualPuzzles?