- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies continuous latent reasoning in MLLMs, framed as a latent-variable model: a prior that sees only the question (used at test time) and a posterior that also sees the answer (used in training). It argues that standard variational training lets the posterior take answer-dependent shortcuts that the prior then copies. It proposes AMVL, which adds a second, reverse-direction KL term to keep the posterior close to the prior. On Qwen2.5-VL-7B, it reports gains on V*, HRBench, and a subset of BLINK, plus analyses showing the latent tokens are used.
Critical strengths
- A real problem with a cheap fix. The mismatch between a training-time posterior that sees the answer and a test-time prior that does not is a genuine obstacle for latent reasoning. The fix adds only two small heads and negligible inference cost.
- Evidence that the latent tokens are actually used. Removing, shuffling, or cutting the latent tokens at test time hurts performance, and so does blocking them from seeing the image. This addresses a common doubt about latent-reasoning methods.
Critical weaknesses
1. The main claim that a new reverse-KL term fixes a problem standard training cannot is not supported. The objective is very close to "KL balancing" from DreamerV2 (ICLR 2021), which also trains the prior faster than it constrains the posterior, for the same stated reason. The main difference is the direction of one KL term, and DreamerV2 is not cited. The argument that standard variational training fails is made against a modified version with gradients blocked to the posterior. Standard training already pulls the posterior toward the prior, as the appendix itself acknowledges. No experiment compares against the standard objective or against KL balancing. This is consequential because the paper's contribution is the training objective, and without these comparisons its advantage over known methods is unproven.
2. The measured benefit of the new term is small, single-run, and partly confounded. Adding the reverse term to the forward-only version gains 3 of 191 questions on V*, nothing on HRBench4K, and 6 of 800 on HRBench8K. There are no repeated runs, and the default loss weights are the only setting that reaches the V* score. The BLINK gain uses 7 of 14 subtasks, the +32 on Jigsaw corresponds to 48 questions, and there is no SFT baseline on BLINK. This is consequential because the headline gains may come from the latent-token architecture and training data rather than from the proposed objective.
Questions for authors
- How does AMVL compare with standard variational training (without blocked gradients) and with DreamerV2-style KL balancing, under identical data and schedule?
- Can you report multiple seeds (or CIs) for Tables 1–3, and full BLINK results (all 14 subtasks) with an SFT baseline?
7 comments
Response to Reviewer 5Kz1 (2/3).
W2 & Q2: Multi-Seed Evaluations (Tables 1 & 2), Full 14-Subtask BLINK with SFT Baseline, and Isolating the Objective (Table 3)
To directly address your concerns regarding multi-seed stability, full 14-subtask BLINK coverage, and the SFT baseline on BLINK, we conducted multi-seed evaluations () for Table 1 (, HRBench4K, HRBench8K) and Table 2 (expanded to all 14 BLINK subtasks, including Baseline-SFT):
1. Multi-Seed Results on , HRBench4K, and HRBench8K (Table 1):
Across multiple seeds, AMVL (Ours-7B) achieves 74.9 ± 0.2% average accuracy (83.8 ± 0.7% on , 72.3 ± 0.4% on HRBench4K, and 68.7 ± 0.3% on HRBench8K), outperforming Baseline-SFT (70.9 ± 0.5% Avg., a +4.0% gain, i.e., separation) as well as all continuous-latent baselines (LVR: 70.9 ± 1.3%, Mull-Tokens: 71.2 ± 0.0%, Monet: 72.1 ± 1.3%) under identical 7B backbones:
| Method | V* (Overall) | V* (Attr / Spat) | HRBench4K (Overall) | HRBench4K (FSP / FCP) | HRBench8K (Overall) | HRBench8K (FSP / FCP) | Avg. |
|---|---|---|---|---|---|---|---|
| Baseline-SFT | 81.4 ± 0.4 | 82.6±0.0 / 79.6±0.9 | 69.8 ± 1.5 | 86.5±1.4 / 53.1±1.6 | 61.4 ± 0.6 | 77.4±1.2 / 45.5±0.0 | 70.9 ± 0.5 |
| LVR | 79.3 ± 1.9 | 78.3±2.5 / 80.9±0.9 | 69.8 ± 1.0 | 83.1±1.6 / 56.5±0.4 | 63.6 ± 1.2 | 75.6±1.9 / 51.5±0.4 | 70.9 ± 1.3 |
| Mull-Tokens | 80.1 ± 0.0 | 77.4±0.0 / 84.2±0.0 | 68.8 ± 0.1 | 85.9±0.5 / 51.8±0.4 | 64.8 ± 0.0 | 80.6±0.9 / 48.9±0.9 | 71.2 ± 0.0 |
| Monet | 81.2 ± 2.2 | 82.2±1.8 / 79.6±2.8 | 70.2 ± 1.2 | 85.8±0.4 / 54.6±1.9 | 65.1 ± 0.6 | 78.0±0.4 / 52.1±0.9 | 72.1 ± 1.3 |
| AMVL (Ours-7B) | 83.8 ± 0.7 | 84.4±0.0 / 82.9±1.9 | 72.3 ± 0.4 | 88.0±0.4 / 56.5±0.4 | 68.7 ± 0.3 | 83.8±0.0 / 53.6±0.5 | 74.9 ± 0.2 |
Response to Reviewer 5Kz1 (3/3).
2. Multi-Seed Results Across All 14 BLINK Subtasks with Baseline-SFT (Table 2):
Evaluated across all 14 BLINK subtasks, AMVL achieves the highest overall macro-average of , outperforming Baseline-SFT (48.2 ± 0.5% Avg.) by +8.6 points and surpassing all discrete and continuous reasoning baselines (Vision-R1: 51.3 ± 0.0%, LVR: 54.6 ± 0.6%, Mull-Tokens: 56.0 ± 0.3%, Monet: 48.0 ± 1.1%).
Table: Multi-seed performance (, %) across all 14 BLINK subtasks with Baseline-SFT (Appendix K.2).
| Method | IQ Test | Jigsaw | Obj. Loc. | Art Style | M-View | Spatial Rel. | Vis. Sim. | Counting | Rel. Depth | Rel. Refl. | Func. Corr. | Sem. Corr. | Vis. Corr. | Forensic Det. | Avg. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Baseline-SFT | |||||||||||||||
| Vision-R1 | |||||||||||||||
| LVR | |||||||||||||||
| Mull-Tokens | |||||||||||||||
| Monet | |||||||||||||||
| AMVL (Ours-7B) |
3. Determinism of Table 3 and Isolating the Proposed Objective from the Latent-Token Architecture:
- Why Table 3 is deterministic: For the internal objective ablations in Table 3, all models are trained under a fixed seed and evaluated using deterministic greedy decoding with the deterministic prior mean () to ensure exact reproducibility across objective variants.
- Isolating the proposed objective under identical architecture and data: Holding the 8-slot latent architecture and 10k training dataset strictly identical, replacing our asymmetric decoupled objective (
74.97%Avg.) with standard unblocked ELBO (71.28%Avg., Appendix I), DreamerV2-style KL balancing (70.87%Avg., Appendix I), uncoupled posterior gradients (73.03%Avg., Appendix J), orreverse-firstwarmup (73.75%Avg., Table 3) drops average accuracy by 1.22% to 4.10%—far exceeding the multi-seed standard deviation (±0.2% Avg. in Table 1). Moreover, adding over forward-only alignment (w/o Rev-KL) not only improves clean accuracy (74.20%74.97%Avg., +3.8), but also slashes counterfactual answer sensitivity () and boosts Counterfactual Latent-Replacement Accuracy on by +9.5 points (21.0%30.5%, Appendix K.4), confirming that our asymmetric objective is essential beyond the latent architecture and training data.
Response to Reviewer 5Kz1 (1/3).
Thank you for your time and detailed feedback. Below, we provide a point-by-point response, with all corresponding updates highlighted in blue in the revised paper. Unless noted otherwise, all references to sections, equations, pages, and citations throughout this response refer to the revised manuscript.
W1 & Q1: Conceptual and Empirical Comparison Against Standard Variational Training (Unblocked ELBO) and DreamerV2-Style KL Balancing
We thank the reviewer for pointing out this important connection. In the revised manuscript, we have cited DreamerV2 (Hafner et al., ICLR 2021) in Section 2.3 (Related Work) and added a strictly controlled empirical comparison against both Standard Variational Training (unblocked ELBO) and DreamerV2-style KL Balancing under the exact same Qwen2.5-VL-7B backbone, training data, and schedule in Appendix I (Table 9):
Under identical data and schedules, AMVL outperforms Standard Variational Training by +3.69 Avg. and DreamerV2-style KL Balancing by +4.10 Avg. for two fundamental reasons:
1. Why AMVL Outperforms Standard Variational Training (Unblocked ELBO): When qϕ(Z∣x,y) and pθ(Z∣x) are optimized simultaneously under unblocked DKL(qϕ(Z∣x,y)∥pθ(Z∣x)), the model faces an asymmetric shortcut: because qϕ(Z∣x,y) has direct access to the ground-truth answer y, it easily encodes answer-leaked shortcuts into Z to minimize the decoder's next-token prediction loss LNTP. Crucially, to reduce DKL(qϕ∥pθ), the prior pθ(Z∣x) is pulled backward to fit these answer-leaked posterior samples Z∼qϕ(Z∣x,y)—which is effectively analogous to doing SFT on pθ(Z∣x) using (x,Z) pairs where Z contains leaked answer information. During training, the KL divergence still decreases as pθ memorizes dataset-level correlations toward the answer-leaked Z (as proven in Lemma 1, Appendix C.1, where μθ∗(x)=μϕ(x)(x)+δˉ(x) absorbs the expected answer shift). However, at inference time, pθ(Z∣x) has no access to y and cannot reproduce the answer-leaked latent Z that the decoder was trained to rely on, causing a severe train-inference mismatch and degraded test-time accuracy (especially on HRBench4K: 66.75 vs. 72.12, and HRBench8K: 64.88 vs. 68.50).
2. Why AMVL Outperforms DreamerV2-Style KL Balancing (The Crucial Role of KL Orientation): While our stop-gradient decoupling shares DreamerV2's motivation of training the prior faster than constraining the posterior, reversing the orientation of the second KL term is mathematically and empirically critical (Appendix B.3 & Appendix I):
DreamerV2 uses the same orientation for both terms: DKL(qϕ∥pθ)=Eqϕ[log(qϕ/pθ)]. In its posterior regularizer DKL(qϕ∥sg[pθ]), optimization is over the first argument qϕ (mode-seeking), which allows qϕ(Z∣x,y) to concentrate on a narrow, answer-specific mode without covering the full support of the input-only prior pθ(Z∣x).
AMVL reverses the second term into Lrev=DKL(sg[pθ]∥qϕ)=Epθ[log(pθ/qϕ)]: Because optimization occurs on the second argument in both Lfwd and Lrev (Appendix B.3, Figure 2), Lrev takes expectations over the target-agnostic prior pθ(Z∣x) and heavily penalizes the posterior whenever it assigns low density to prior-reachable regions or shrinks σϕ2→0 around an answer shortcut (Proposition 1), yielding a +4.10 Avg. improvement over DreamerV2-style balancing.