acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning

Abstract

Multimodal Large Language Models (MLLMs) are often constrained by a language-space bottleneck, forcing complex visual reasoning into discrete tokens which can lose perceptual nuance. A promising alternative is continuous latent reasoning, where the goal is to discover implicit reasoning pathways that bridge the multimodal query and the final answer. However, this introduces a severe train-inference mismatch: a training-time posterior, conditioned on the ground-truth answer, can exploit answer-dependent shortcuts. Standard variational training then forces the inference-time prior to mimic a posterior that has access to information unavailable at test time, leading to poor performance. To address this, we propose Asymmetric Mutual Variational Learning (AMVL), a framework that resolves this mismatch via a bidirectional calibration objective. A forward KL divergence trains the target-agnostic prior to match the posterior, while a novel reverse KL divergence simultaneously regularizes the posterior, preventing it from collapsing into inference-incompatible regions and mitigating this “answer leakage”. We provide theoretical analysis formalizing this leakage as prior contamination and prove that our dual-KL objective reduces it. We instantiate AMVL in a latent-integrated MLLM and show that it consistently outperforms strong discrete and latent-reasoning baselines, improving the average score on the complex BLINK benchmark by +10.83 and achieving gains of up to +32.00 on individual reasoning tasks, with analyses confirming improved latent-space stability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Models, Multimodal Reasoning, Continuous Latent Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5Kz1OpenReview ↗
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
  1. 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.
  2. 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
  1. How does AMVL compare with standard variational training (without blocked gradients) and with DreamerV2-style KL balancing, under identical data and schedule?
  2. Can you report multiple seeds (or CIs) for Tables 1–3, and full BLINK results (all 14 subtasks) with an SFT baseline?
7 comments
AuthorsOpenReview ↗

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):

MethodVariational KL ObjectiveV∗V^{\ast}HRBench4KHRBench8KAvg.
Standard variational training (Unblocked ELBO)β DKL(qϕ(Z∣x,y)∥pθ(Z∣x))\beta\,\mathcal{D}_{\mathrm{KL}}(q_\phi(Z \mid x,y) \Vert p_\theta(Z \mid x)) (unblocked gradients)82.2066.7564.8871.28
DreamerV2-style KL balancingα DKL(sg[qϕ]∥pθ)+(1−α) DKL(qϕ∥sg[pθ])\alpha\,\mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[q_\phi] \Vert p_\theta) + (1-\alpha)\,\mathcal{D}_{\mathrm{KL}}(q_\phi \Vert \mathrm{sg}[p_\theta])79.0968.6364.8870.87
AMVL (Ours)β DKL(sg[qϕ]∥pθ)+γ(t) DKL(sg[pθ]∥qϕ)\beta\,\mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[q_\phi] \Vert p_\theta) + \gamma(t)\,\mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[p_\theta] \Vert q_\phi)84.2972.1268.5074.97

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)q_\phi(Z \mid x, y) and pθ(Z∣x)p_\theta(Z \mid x) are optimized simultaneously under unblocked DKL(qϕ(Z∣x,y)∥pθ(Z∣x))\mathcal{D}_{\mathrm{KL}}(q_\phi(Z \mid x,y) \parallel p_\theta(Z \mid x)), the model faces an asymmetric shortcut: because qϕ(Z∣x,y)q_\phi(Z \mid x,y) has direct access to the ground-truth answer yy, it easily encodes answer-leaked shortcuts into ZZ to minimize the decoder's next-token prediction loss LNTP\mathcal{L}_{\mathrm{NTP}}. Crucially, to reduce DKL(qϕ∥pθ)\mathcal{D}_{\mathrm{KL}}(q_\phi \parallel p_\theta), the prior pθ(Z∣x)p_\theta(Z \mid x) is pulled backward to fit these answer-leaked posterior samples Z∼qϕ(Z∣x,y)Z \sim q_\phi(Z \mid x,y)—which is effectively analogous to doing SFT on pθ(Z∣x)p_\theta(Z \mid x) using (x,Z)(x, Z) pairs where ZZ contains leaked answer information. During training, the KL divergence still decreases as pθp_\theta memorizes dataset-level correlations toward the answer-leaked ZZ (as proven in Lemma 1, Appendix C.1, where μθ∗(x)=μϕ(x)(x)+δˉ(x)\mu_\theta^{\ast}(x) = \mu_\phi^{(x)}(x) + \bar{\delta}(x) absorbs the expected answer shift). However, at inference time, pθ(Z∣x)p_\theta(Z \mid x) has no access to yy and cannot reproduce the answer-leaked latent ZZ 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)]\mathcal{D}_{\mathrm{KL}}(q_\phi \parallel p_\theta) = \mathbb{E}_{q_\phi}[\log(q_\phi/p_\theta)]. In its posterior regularizer DKL(qϕ∥sg[pθ])\mathcal{D}_{\mathrm{KL}}(q_\phi \parallel \mathrm{sg}[p_\theta]), optimization is over the first argument qϕq_\phi (mode-seeking), which allows qϕ(Z∣x,y)q_\phi(Z \mid x,y) to concentrate on a narrow, answer-specific mode without covering the full support of the input-only prior pθ(Z∣x)p_\theta(Z \mid x).

  • AMVL reverses the second term into Lrev=DKL(sg[pθ]∥qϕ)=Epθ[log⁡(pθ/qϕ)]\mathcal{L}_{\mathrm{rev}} = \mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[p_\theta] \parallel q_\phi) = \mathbb{E}_{p_\theta}[\log(p_\theta/q_\phi)]: Because optimization occurs on the second argument in both Lfwd\mathcal{L}_{\mathrm{fwd}} and Lrev\mathcal{L}_{\mathrm{rev}} (Appendix B.3, Figure 2), Lrev\mathcal{L}_{\mathrm{rev}} takes expectations over the target-agnostic prior pθ(Z∣x)p_\theta(Z \mid x) and heavily penalizes the posterior whenever it assigns low density to prior-reachable regions or shrinks σϕ2→0\sigma_\phi^2 \to 0 around an answer shortcut (Proposition 1), yielding a +4.10 Avg. improvement over DreamerV2-style balancing.

AuthorsOpenReview ↗

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 (mean±std\mathrm{mean} \pm \mathrm{std}) for Table 1 (V∗V^{\ast}, HRBench4K, HRBench8K) and Table 2 (expanded to all 14 BLINK subtasks, including Baseline-SFT):

1. Multi-Seed Results on V∗V^{\ast}, HRBench4K, and HRBench8K (Table 1): Across multiple seeds, AMVL (Ours-7B) achieves 74.9 ± 0.2% average accuracy (83.8 ± 0.7% on V∗V^{\ast}, 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., >7σ>7\sigma 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:

MethodV* (Overall)V* (Attr / Spat)HRBench4K (Overall)HRBench4K (FSP / FCP)HRBench8K (Overall)HRBench8K (FSP / FCP)Avg.
Baseline-SFT81.4 ± 0.482.6±0.0 / 79.6±0.969.8 ± 1.586.5±1.4 / 53.1±1.661.4 ± 0.677.4±1.2 / 45.5±0.070.9 ± 0.5
LVR79.3 ± 1.978.3±2.5 / 80.9±0.969.8 ± 1.083.1±1.6 / 56.5±0.463.6 ± 1.275.6±1.9 / 51.5±0.470.9 ± 1.3
Mull-Tokens80.1 ± 0.077.4±0.0 / 84.2±0.068.8 ± 0.185.9±0.5 / 51.8±0.464.8 ± 0.080.6±0.9 / 48.9±0.971.2 ± 0.0
Monet81.2 ± 2.282.2±1.8 / 79.6±2.870.2 ± 1.285.8±0.4 / 54.6±1.965.1 ± 0.678.0±0.4 / 52.1±0.972.1 ± 1.3
AMVL (Ours-7B)83.8 ± 0.784.4±0.0 / 82.9±1.972.3 ± 0.488.0±0.4 / 56.5±0.468.7 ± 0.383.8±0.0 / 53.6±0.574.9 ± 0.2
AuthorsOpenReview ↗

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 56.8±0.3%56.8 \pm 0.3\%, 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 (mean±std\text{mean} \pm \text{std}, %) across all 14 BLINK subtasks with Baseline-SFT (Appendix K.2).

MethodIQ TestJigsawObj. Loc.Art StyleM-ViewSpatial Rel.Vis. Sim.CountingRel. DepthRel. Refl.Func. Corr.Sem. Corr.Vis. Corr.Forensic Det.Avg.
Baseline-SFT22.7±0.022.7\pm0.052.7±0.052.7\pm0.052.2±4.152.2\pm4.165.2±4.065.2\pm4.044.4±0.044.4\pm0.083.9±0.083.9\pm0.054.6±0.454.6\pm0.471.7±0.871.7\pm0.871.0±2.171.0\pm2.132.6±3.532.6\pm3.522.6±0.922.6\pm0.931.7±0.731.7\pm0.744.2±2.144.2\pm2.125.8±0.825.8\pm0.848.2±0.548.2\pm0.5
Vision-R119.3±1.219.3\pm1.252.9±1.452.9\pm1.449.5±1.349.5\pm1.364.4±1.064.4\pm1.042.9±0.842.9\pm0.879.5±0.479.5\pm0.471.9±0.771.9\pm0.766.7±1.766.7\pm1.772.3±2.072.3\pm2.038.8±3.438.8\pm3.430.3±1.930.3\pm1.933.6±2.233.6\pm2.247.1±2.547.1\pm2.549.5±1.949.5\pm1.951.3±0.051.3\pm0.0
LVR28.9±2.528.9\pm2.551.6±0.451.6\pm0.452.7±0.552.7\pm0.570.4±1.070.4\pm1.042.4±3.842.4\pm3.888.3±0.488.3\pm0.485.7±0.485.7\pm0.468.3±1.768.3\pm1.778.8±2.878.8\pm2.840.1±1.740.1\pm1.726.9±1.326.9\pm1.339.6±0.739.6\pm0.755.6±1.255.6\pm1.234.6±1.634.6\pm1.654.6±0.654.6\pm0.6
Mull-Tokens31.3±1.831.3\pm1.868.2±2.068.2\pm2.046.7±0.846.7\pm0.863.5±0.563.5\pm0.558.4±0.458.4\pm0.484.9±0.484.9\pm0.481.2±0.481.2\pm0.468.9±1.368.9\pm1.382.5±1.782.5\pm1.736.3±2.636.3\pm2.625.9±1.625.9\pm1.627.1±1.527.1\pm1.551.9±2.051.9\pm2.057.1±0.957.1\pm0.956.0±0.356.0\pm0.3
Monet22.0±2.722.0\pm2.745.3±0.045.3\pm0.049.7±2.549.7\pm2.563.3±3.163.3\pm3.151.6±8.751.6\pm8.774.8±2.874.8\pm2.883.0±1.383.0\pm1.359.7±7.059.7\pm7.054.8±4.354.8\pm4.330.1±5.030.1\pm5.026.2±4.626.2\pm4.633.8±2.533.8\pm2.544.2±4.544.2\pm4.534.1±6.234.1\pm6.248.0±1.148.0\pm1.1
AMVL (Ours-7B)31.1±0.831.1\pm0.875.3±1.275.3\pm1.256.6±0.056.6\pm0.068.1±0.568.1\pm0.555.6±0.055.6\pm0.088.3±0.488.3\pm0.487.7±1.187.7\pm1.167.8±1.067.8\pm1.077.4±0.877.4\pm0.844.0±2.244.0\pm2.224.9±0.924.9\pm0.928.5±0.428.5\pm0.446.5±0.646.5\pm0.643.7±0.943.7\pm0.956.8±0.356.8\pm0.3

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 Z=μθ(x)Z = \mu_\theta(x) (τ=0\tau = 0) 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), or reverse-first warmup (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 Lrev\mathcal{L}_{\mathrm{rev}} over forward-only alignment (w/o Rev-KL) not only improves clean accuracy (74.20% →\to 74.97% Avg., +3.8σ\sigma), but also slashes counterfactual answer sensitivity (Dcf:7.77→1.82D_{\mathrm{cf}}: 7.77 \to 1.82) and boosts Counterfactual Latent-Replacement Accuracy on V∗V^{\ast} by +9.5 points (21.0% →\to 30.5%, Appendix K.4), confirming that our asymmetric objective is essential beyond the latent architecture and training data.
Reviewer_V17GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In current multimodal large-language models for visual reasoning, intermediate thought processes must be converted into discrete text tokens. However, spatial perception and fine-grained visual details are high dimensional continuous information. Translating such information into text leads to loss of subtle perceptual details, which further causes hallucinations and reasoning drift. To address this issue, the paper designs two KL losses with decoupled gradients and distinct functional roles. Combined with stop gradient operators, these losses perform bidirectional calibration between the prior and posterior distributions so as to mitigate answer leakage. The forward KL pulls the prior toward informative posterior states, while the reverse KL constrains the posterior within the reachable latent space of the prior and prevents the posterior from exploiting test time inaccessible shortcuts. Significant performance improvements are observed across multiple datasets.

Critical strengths
  1. This paper identifies the structural mismatch between the target‑aware posterior and target‑agnostic prior, and formalizes the performance degradation in continuous multimodal reasoning as answer leakage / prior contamination. Its motivation offers greater explanatory power than simply adding latent tokens.

  2. The forward KL and reverse KL act on the prior and posterior respectively, and stop‑gradient operators preserve the distinct roles of the two distributions. During inference, the mean of the prior can be directly utilized, making it engineering‑friendly for integration into existing MLLMs.

Critical weaknesses
  1. AMVL is jointly trained on multiple visual‑reasoning datasets, whereas the compared baselines may adopt different training data or additional supervision from their original papers. Without unified backbones, training data, training steps and inference budgets, it is difficult to fully attribute performance gains to the bidirectional KL objective.

  2. The discussion on interpretability and computational cost of continuous latent variables is insufficient. Although the method eliminates explicit textual reasoning, it introduces additional posterior/prior variational heads as well as sampling overhead during training. The paper does not systematically report training GPU memory consumption, training wall‑clock time, inference latency, nor quantitative comparisons of token/computation overhead against discrete CoT baselines.

Questions for authors
  1. Please supplement sensitivity curves for β, γ, reverse‑KL delay scheduling and optimization order, and clarify whether posterior collapse or training divergence is observed.

  2. Please report the training‑ / inference‑side FLOPs, GPU memory footprint, and latency of AMVL alongside all compared baselines, to illustrate the practical efficiency gains brought by continuous‑latent reasoning.

  3. It is recommended to add transfer experiments across different backbones or datasets. Furthermore, counter‑factual diagnostics such as answer perturbation and input occlusion should be adopted to verify that the latent representations indeed reduce answer dependency, rather than merely benefiting from stronger supervisory signals.

5 comments
AuthorsOpenReview ↗

Response to Reviewer V17G (1/3).

Thank you for your positive feedback and insightful comments. Our detailed responses are provided below, with corresponding revisions highlighted in blue in the revised paper. Throughout this rebuttal, all references to sections, equations, pages, and citations correspond to the revised manuscript unless noted otherwise.

W1: Strictly Controlled Comparisons Under Unified Backbone, Data, Schedule, and Inference Budget

All core comparisons are conducted under a strictly unified setup using the exact same backbone (Qwen2.5-VL-7B), identical training dataset (Appendix E), identical training steps, and identical inference budget (k=8k=8 latent tokens):

  1. Existing Controlled Baselines (Sections 4.2 & 4.3): AMVL is trained on the exact same corpus as SFT and the strongest continuous latent baseline Monet. Under identical data and backbone, AMVL (74.97% Avg.) outperforms SFT (70.56%, +4.41) and Monet (74.08%, +0.89, despite Monet using an extra RL stage), as well as our controlled objective ablations NTP only (73.01%, +1.96) and w/o Rev-KL (NTP + Fwd-KL) (74.20%, +0.77) under the same k=8k=8 latent budget.

  2. New Controlled Comparisons Against Alternative Variational Objectives (Appendix I): We additionally trained Standard Variational Training (unblocked ELBO) and DreamerV2-style KL Balancing under the exact same backbone, data, schedule, and k=8k=8 budget:

Training StrategyV*HRBench4KHRBench8KAvg.
Standard variational training (unblocked ELBO)82.2066.7564.8871.28
DreamerV2-style KL balancing79.0968.6364.8870.87
AMVL (Ours)84.2972.1268.5074.97

Holding architecture, data, schedule, and inference budget strictly fixed, AMVL outperforms Standard Variational Training by +3.69 Avg. and DreamerV2-style KL Balancing by +4.10 Avg., isolating the gain to our asymmetric bidirectional KL objective.

W2 & Q2: Computational Efficiency and Latent Interpretability

1. Training & Inference Efficiency (Appendix E & Appendix K.7): During training, AMVL adds only lightweight variational heads ($\sim$15M parameters, <0.2%<0.2\% of the 7B backbone) and trains in $\sim$20 hours on 16 A100 GPUs. For inference, we added Appendix K.7 (Table 23) benchmarking FLOPs, peak GPU memory, and throughput across the base model, discrete CoT/tool baselines (Pixel-Reasoner), and continuous latent baselines (LVR, Mull-Tokens, Monet):

MethodParadigmFLOPs (TFLOPs/sample) ↓\downarrowGPU Memory (GB) ↓\downarrowThroughput (Samples/s) ↑\uparrowV* (%) ↑\uparrow
Qwen2.5-VL-7BBase Model98.9817.580.6876.44
Pixel-ReasonerDiscrete CoT / Tool152.2737.440.2081.15
LVRContinuous Latent116.3517.690.4681.15
Mull-TokensContinuous Latent104.3218.710.6479.95
MonetContinuous Latent99.6237.350.4683.25
AMVL (Ours)Continuous Latent100.7917.740.6384.29
  • Negligible Overhead over Base Model: With k=8k=8 latent slots, AMVL adds only +0.16 GB memory and +1.8% FLOPs (100.79100.79 vs. 98.9898.98 TFLOPs) over vanilla Qwen2.5-VL-7B, maintaining near-identical throughput (0.630.63 vs. 0.680.68 samples/s) while improving V∗V^* by +7.85%.

  • Substantial Speedup over Baselines: Compared to Pixel-Reasoner, AMVL cuts FLOPs by 33.8%, reduces GPU memory by 52.6%, and achieves a 3.15$\times$ throughput speedup. Compared to continuous baselines, AMVL is 1.37$\times$ faster than LVR and uses 52.5% less GPU memory than Monet (17.7417.74 vs. 37.3537.35 GB) while achieving the highest accuracy.

2. Interpretability of Continuous Latents (Appendix G & Appendix K.6): Our probing analyses confirm that AMVL's continuous states remain highly interpretable and visually grounded: (i) occlusion heatmaps (Figure 3) show individual latent slots progressively localizing onto queried objects rather than diffuse background; (ii) cosine similarity and permutation tests (Figure 4) show a smooth, non-redundant trajectory across slots L1→L8L_1 \to L_8 that is significantly more sensitive to visual perturbations than text perturbations; and (iii) inference-time interventions (Appendix K.6, Tables 21–22) verify that removing, permuting, or masking visual access to the latent tokens degrades accuracy across all benchmarks.

AuthorsOpenReview ↗

Response to Reviewer V17G (2/3).

Q1: Sensitivity to β,γ\beta, \gamma, Reverse-KL Delay, Optimization Order, and Clarifying "Posterior Collapse" vs. "Answer Leakage"

1. Sensitivity Analyses (Section 4.3 & Appendix J): Across all tested (λNTP,β,γ)(\lambda_{\text{NTP}}, \beta, \gamma) ratios, annealing delays, optimization orders, and stop-gradient ablations, AMVL converges smoothly without training divergence:

Configuration / Optimization OrderV*HRBench4KHRBench8KAvg.
AMVL w/o annealing (delay=0\text{delay}=0, simultaneous)82.2071.5067.5073.73
AMVL, delay=500\text{delay}=50080.1072.3868.2573.58
AMVL, delay=1000\text{delay}=1000 (Default forward-first)84.2972.1268.5074.97
AMVL, delay=2000\text{delay}=200082.2072.7568.3874.44
Reverse-first optimization order (Table 3)80.6372.3868.2573.75
w/o stop-grad in Prior Alignment (Appendix J)83.2572.1267.6274.33
w/o stop-grad in Posterior Reg. (Appendix J)80.1072.2566.7573.03
Loss weights (λNTP,β,γ)=(1,0.5,1.0)(\lambda_{\text{NTP}}, \beta, \gamma) = (1, 0.5, 1.0)81.6871.6267.7573.68
Loss weights (λNTP,β,γ)=(2,1.0,1.0)(\lambda_{\text{NTP}}, \beta, \gamma) = (2, 1.0, 1.0)80.6373.0068.8874.17
Loss weights (λNTP,β,γ)=(2,0.5,1.0)(\lambda_{\text{NTP}}, \beta, \gamma) = (2, 0.5, 1.0)81.6872.0069.0074.23
Loss weights (λNTP,β,γ)=(2,1.0,0.5)(\lambda_{\text{NTP}}, \beta, \gamma) = (2, 1.0, 0.5)81.6870.8867.8873.48

2. Clarifying Terminology: Classic "Posterior Collapse" vs. "Answer Leakage (Shortcut Collapse)": We thank the reviewer for raising this point and have clarified the terminology in the revised text (Section 4.2 & Appendix C.5). There are two distinct failure modes at play:

  • Why Classic Posterior Collapse (qϕ(Z∣x,y)→pθ(Z∣x)q_\phi(Z \mid x,y) \to p_\theta(Z \mid x) with ZZ ignored by the decoder) Does Not Occur: If classic posterior collapse occurred, the decoder would ignore ZZ and performance would fall back to the SFT baseline (70.56% Avg.). Instead, all hyperparameter settings maintain 73.03%–74.97% Avg., latent spread remains expressive (Sp=0.6786,Sq=0.8257S_p=0.6786, S_q=0.8257, Appendix H), and removing or permuting ZZ at test time causes immediate accuracy drops (Appendix K.6), proving the decoder actively relies on ZZ.

  • What Our Paper Refers to When Discussing Posterior Concentration — Answer Leakage (Answer-Shortcut Collapse): In several places in the paper (e.g., Abstract and Section 4.2), "collapsing into inference-incompatible regions" refers specifically to answer leakage, where an unconstrained posterior qϕ(Z∣x,y)q_\phi(Z \mid x,y) over-concentrates (σϕ2→0\sigma_\phi^2 \to 0) onto narrow, answer-leaked modes dictated by the ground-truth answer yy rather than visual evidence xx. Because the test-time prior pθ(Z∣x)p_\theta(Z \mid x) cannot access yy, this answer leakage causes a severe train-inference mismatch. Our reverse-KL regularizer Lrev\mathcal{L}_{\mathrm{rev}} specifically prevents this answer-shortcut concentration, cutting counterfactual answer sensitivity DcfD_{\mathrm{cf}} by 76.6% (Appendix K.4).

AuthorsOpenReview ↗

Response to Reviewer V17G (3/3).

Q3: Transfer Experiments and Counterfactual Diagnostics (Answer Perturbation & Visual Deprivation)

1. Cross-Backbone, Cross-Scale, and Cross-Dataset Transfer (Appendix K.2): Using this exact configuration of our method, we verified AMVL's scalability across diverse model families, parameter scales (up to $\sim$34B), and training corpora:

  • LLaVA-v1.5 (7B & 13B): Full-parameter tuning under our default KL schedule.
  • Qwen2.5-VL-32B ($\sim$33.5B): Parameter-efficient adaptation freezing the vision encoder and 32B backbone, applying LoRA (rank=16,α=32,dropout=0.05\mathrm{rank}=16, \alpha=32, \mathrm{dropout}=0.05) while jointly training the lightweight variational heads and k=8k=8 continuous latent tokens (lr=1×10−5\mathrm{lr}=1\times 10^{-5}).

Table: Cross-architecture and cross-scale transfer on LLaVA-v1.5 (7B and 13B).

MethodV∗V^{\ast}HRBench4KHRBench8KAvg.
LLaVA-v1.5-7B31.9429.2527.6229.60
+ AMVL (Ours)33.51 (+1.57)33.12 (+3.87)32.75 (+5.13)33.13 (+3.52)
LLaVA-v1.5-13B36.6538.7536.2537.22
+ AMVL (Ours)38.22 (+1.57)40.00 (+1.25)38.62 (+2.37)38.95 (+1.73)

Table: Scaling AMVL to Qwen2.5-VL-32B with LoRA adaptation.

MethodV∗V^{\ast}HRBench4KHRBench8KMMMUVisualPuzzlesAvg.
Qwen2.5-VL-32B82.7272.8769.8758.7836.9064.23
+ AMVL (Ours)83.25 (+0.53)74.63 (+1.76)70.75 (+0.88)59.56 (+0.78)38.44 (+1.54)65.33 (+1.10)

AMVL delivers consistent gains across scales—from 7B (+3.52 Avg.) and 13B (+1.73 Avg.) to 32B (+1.10 Avg.)—on both fine-grained perception and complex reasoning. The more modest gain on Qwen2.5-VL-32B (+1.10 vs. +5.57 on 7B) reflects both the tighter headroom of the 32B baseline and the use of LoRA (adopted for compute efficiency), which limits the frozen backbone's plasticity in adapting its internal attention to the continuous latent space compared to full fine-tuning. Nonetheless, these results confirm that scaling model size alone does not resolve the train-inference latent mismatch, and AMVL remains effective even under lightweight adaptation.

To evaluate cross-dataset generalization, we further trained Qwen2.5-VL-7B with AMVL from scratch on a new large-scale reasoning corpus, MMFineReason-123K, again without any hyperparameter retuning:

Table: Cross-dataset transfer on BLINK and ScienceQA (trained on MMFineReason-123K).

MethodObj. Loc.Art StyleM-ViewScienceQAAvg.
Qwen2.5-VL-7B49.1864.9641.3583.5259.75
+ AMVL (Ours)54.92 (+5.74)70.09 (+5.13)55.64 (+14.29)87.50 (+3.98)67.04 (+7.29)

AMVL yields a +7.29 average improvement on this new corpus—highlighted by +14.29 on Multi-view Reasoning and +3.98 on ScienceQA. All transfer experiments have been integrated into Appendix K.2.

2. Counterfactual Diagnostics: Answer Perturbation & Visual Deprivation (Appendices K.4–K.6): To verify that AMVL genuinely reduces answer dependency and grounds latents in visual evidence, we conducted both diagnostics:

Table: Counterfactual answer perturbation on V∗V^* (Appendix K.4) and visual deprivation across benchmarks (Appendix K.6).

Diagnostic / ConditionMetric / Benchmarkw/o Rev-KLFull AMVLImpact / Finding
Answer Perturbation (y+→y−y^+ \to y^-)Counterfactual KL Dcf↓D_{\text{cf}} \downarrow7.771.82-76.6% posterior sensitivity to changing answer yy
Answer Perturbation (y+→y−y^+ \to y^-)Answer-Align Margin ↓\downarrow2.84e-52.84\text{e-}51.12e-51.12\text{e-}5-60.6% alignment bias toward ground-truth answer tokens
Answer Perturbation (y+→y−y^+ \to y^-)Counterfactual Acc (CLRA) ↑\uparrow21.0%30.5%+9.5% robustness when injected with wrong-answer latent z−z^-
Visual Deprivation (Mask image from ZZ)V* / HR4K / HR8K (Avg.) ↑\uparrow—84.29 / 72.12 / 68.50 (74.97) →\to 78.53 / 70.62 / 66.38 (71.84)-5.76% on V* (-3.13% Avg.) when latents cannot attend to image

Perturbing y+→y−y^+ \to y^- confirms that Lrev\mathcal{L}_{\text{rev}} makes the posterior 4.3×4.3\times less sensitive to the target answer (also visually confirmed by our controlled answer-shift interventions in Appendix K.5, Figure 5), while blocking visual attention to the latent tokens (Appendix K.6, Table 22) causes a sharp accuracy drop (-5.76% on V∗V^*), proving that AMVL's latents extract indispensable visual information rather than acting as generic supervision shortcuts.

Reviewer_yRqBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the answer‑leakage / train‑inference mismatch problem that arises when applying variational latent reasoning to multimodal large language models. In standard ELBO‑based variational setups, the training‑time posterior has access to ground‑truth answer y and may learn answer‑dependent latent shortcuts; the inference‑only prior is then forced to mimic these posterior latents, degrading test‑time performance. The authors propose AMVL (Asymmetric Mutual Variational Learning), a bidirectional‑KL variational framework. It combines a forward KL loss that trains the target‑agnostic prior to match posterior latent states, and a novel reverse‑KL regularizer that constrains the posterior from drifting into answer‑biased, inference‑incompatible latent regions.

Critical strengths
  1. Most existing latent‑reasoning work uses vanilla ELBO without acknowledging this train‑inference hazard. This paper formalizes the prior‑contamination problem, and proposes asymmetric mutual variational learning with forward‑KL prior alignment plus reverse‑KL posterior regularization.

  2. Theoretical propositions prove that this bidirectional objective reduces answer‑dependent mean contamination. This offers a general‑purpose recipe for variational latent‑variable autoregressive MLLMs beyond multimodal visual reasoning.

  3. Beyond standard benchmark numbers, the authors conduct a broad suite of analyses: latent geometry spread statistics, occlusion‑based visual grounding heatmaps, counter‑factual leakage metrics, latent token permutation / removal interventions, inference‑time stochastic‑sampling robustness test.

  4. Ablations confirm the necessity of both KL terms, stop‑gradient operators, and the proposed LLM‑native variational head. Empirical gains are observed across fine‑grained perception, complex spatial reasoning and out‑of‑distribution VisualPuzzles.

Critical weaknesses
  1. Propositions and proofs are built upon diagonal‑Gaussian latent variables and focus primarily on posterior mean shift. It leaves covariance‑based leakage, non‑Gaussian leakage unaddressed

  2. The paper does not validate AMVL on larger‑size MLLMs (e.g. 34B‑70B). It is unknown whether the bidirectional‑KL optimization still brings consistent gains at larger scales, or if larger models inherently mitigate answer leakage.

  3. While counter‑factual diagnostics show reverse‑KL reduces leakage‑related metrics, there is no explicit controlled intervention that isolates reduced leakage as the exact causal source of downstream performance improvements.

Questions for authors
  1. All main experiments run on 7B backbone. What obstacles prevented scaling AMVL to larger‑size MLLMs? Do you have preliminary hints whether AMVL would retain its advantages on bigger multimodal models?

  2. How sensitive is training stability to the delayed annealing schedule for reverse‑KL? What practical guidance would you give for choosing β / γ schedules when porting AMVL to new datasets or different base MLLMs?

  3. At inference, AMVL uses deterministic prior mean for generation. Have you explored using stochastic sampling from prior at test‑time for improving hard reasoning examples?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yRqB (1/4).

We appreciate your supportive review and thoughtful suggestions. We have revised the manuscript in light of your feedback to make our presentation more precise. Below, we provide a point-by-point response and indicate the corresponding changes in blue in the revised paper. Unless stated otherwise, references to pages, equations, sections, and bibliography entries throughout this response are to the revised version.

W1: Theoretical Assumptions: Gaussian Parameterization, Covariance Regularization, and Non-Gaussian Generalization

We focus our closed-form propositions on diagonal-Gaussian latents and mean shift for three principled reasons, while our broader framework naturally addresses covariance and non-Gaussian leakage:

  1. Exact Alignment with Model & Deterministic Inference: Factorized diagonal Gaussians are standard in conditional variational inference (Kingma & Welling, 2013; Sohn et al., 2015) for tractable, closed-form KL optimization. More importantly, this assumption exactly matches our architecture (Eqs. 6–7, Section 3.3) and greedy test-time decoding rule (Z=μθ(x)Z = \mu_\theta(x), Section 3.4). Because inference injects only the prior mean μθ(x)\mu_\theta(x), analyzing mean contamination directly targets the exact representation used at test time.

  2. Covariance Leakage is Prevented by Both Mean-Based Inference and Reverse KL (Lrev\mathcal{L}_{\mathrm{rev}}): First, since test-time decoding uses Z=μθ(x)Z = \mu_\theta(x) without sampling from the prior covariance Σθ(x)\Sigma_\theta(x), any residual answer-conditioned covariance variation cannot contaminate inference unless it shifts the mean. Second, during training, Lrev\mathcal{L}_{\mathrm{rev}} directly regularizes the full posterior covariance. For multivariate Gaussians pθ=N(μθ,Σθ)p_\theta = \mathcal{N}(\mu_\theta, \Sigma_\theta) and qϕ=N(μϕ,Σϕ)q_\phi = \mathcal{N}(\mu_\phi, \Sigma_\phi), setting ∂Lrev/∂Σϕ−1=0\partial \mathcal{L}_{\mathrm{rev}} / \partial \Sigma_\phi^{-1} = 0 yields the exact optimum Σϕ∗(x,y)=Σθ(x)+(μθ−μϕ)(μθ−μϕ)⊤⪰Σθ(x)\Sigma_\phi^{\ast}(x,y) = \Sigma_\theta(x) + (\mu_\theta - \mu_\phi)(\mu_\theta - \mu_\phi)^\top \succeq \Sigma_\theta(x) (reducing to σϕ,j2∗=σθ,j2+(μθ,j−μϕ,j)2\sigma_{\phi,j}^{2\ast} = \sigma_{\theta,j}^2 + (\mu_{\theta,j} - \mu_{\phi,j})^2 in the diagonal case, Appendix B.4). Thus, Lrev\mathcal{L}_{\mathrm{rev}} anchors Σϕ(x,y)\Sigma_\phi(x,y) to the target-agnostic prior covariance Σθ(x)\Sigma_\theta(x) and prevents the posterior from encoding answer shortcuts via degenerate or low-rank covariance directions (Σϕ→0\Sigma_\phi \to 0).

  3. Distribution-Free (Non-Gaussian) Guarantees: Beyond parametric Gaussians, our information-theoretic decomposition in Proposition 3 (Appendix C.5)—linking forward KL to conditional mutual information Iq(Z;Y∣X)I_q(Z; Y \mid X), which captures all higher-order and non-Gaussian answer leakage—holds for arbitrary distributions pθp_\theta and qϕq_\phi. Likewise, the mass-covering property (Appendix B.3) holds generally: Lrev=DKL(sg[pθ]∥qϕ)\mathcal{L}_{\mathrm{rev}} = \mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[p_\theta] \parallel q_\phi) universally penalizes posterior concentration onto narrow, answer-leaked modes unsupported by the prior.

AuthorsOpenReview ↗

Response to Reviewer yRqB (2/4).

W2 & Q1: Scaling AMVL Beyond 7B Backbones (7B →\to 13B →\to 32B)

To verify scalability across model families and sizes up to the $\sim$34B regime, we evaluated AMVL on LLaVA-v1.5 (7B and 13B) and Qwen2.5-VL-32B ($\sim$33.5B total parameters) using our default delayed reverse-KL schedule without hyperparameter retuning (Appendix K.2):

  • LLaVA-v1.5 (7B & 13B): Full-parameter training with the exact same KL schedule as our main experiments.

  • Qwen2.5-VL-32B: Parameter-efficient adaptation freezing the vision encoder and 32B language backbone, applying LoRA (rank=16,α=32,dropout=0.05\text{rank}=16, \alpha=32, \text{dropout}=0.05) while jointly training our lightweight variational heads and k=8k=8 continuous latent tokens (lr=1×10−5\text{lr}=1\times 10^{-5}).

Table: Cross-architecture and cross-scale transfer on LLaVA-v1.5 (7B and 13B, Appendix K.2).

MethodV*HRBench4KHRBench8KAvg.
LLaVA-v1.5-7B31.9429.2527.6229.60
+ AMVL (Ours)33.51 (+1.57)33.12 (+3.87)32.75 (+5.13)33.13 (+3.52)
LLaVA-v1.5-13B36.6538.7536.2537.22
+ AMVL (Ours)38.22 (+1.57)40.00 (+1.25)38.62 (+2.37)38.95 (+1.73)

Table: Scaling AMVL to Qwen2.5-VL-32B with LoRA adaptation (Appendix K.2).

MethodV*HRBench4KHRBench8KMMMUVisualPuzzlesAvg.
Qwen2.5-VL-32B82.7272.8769.8758.7836.9064.23
+ AMVL (Ours)83.25 (+0.53)74.63 (+1.76)70.75 (+0.88)59.56 (+0.78)38.44 (+1.54)65.33 (+1.10)

AMVL delivers consistent improvements from 7B (+3.52 Avg.) and 13B (+1.73 Avg.) up to 32B (+1.10 Avg.) across fine-grained perception and complex reasoning. The smaller absolute gain on Qwen2.5-VL-32B (+1.10 vs. +5.57 on 7B) reflects both the tighter headroom of the 32B baseline and the use of lightweight LoRA (due to compute constraints), which restricts the frozen backbone's plasticity in adapting its internal attention to the continuous latent space compared to full-parameter tuning. Overall, larger scale alone does not eliminate the train-inference latent mismatch, and AMVL scales effectively even under LoRA adaptation.

AuthorsOpenReview ↗

Response to Reviewer yRqB (3/4).

W3: Sources of Downstream Gains and Why Our Latent Manipulation Proves a Causal Effect

AMVL's total improvement over base Qwen2.5-VL-7B (69.40% →\to 74.97% Avg., +5.57%) stems from two complementary sources, which we isolate via step-by-step ablations and controlled interventions:

1. Source 1 — Variational Latent Thinking Provides the Foundational Boost: Formulating continuous reasoning as variational inference already improves over base Qwen2.5-VL-7B (69.40%) and discrete SFT (70.56%). Unblocked Standard Variational Training (ELBO) reaches 71.28% Avg., and decoupling the forward KL (NTP + Fwd-KL, i.e., w/o Rev-KL) reaches 74.20% Avg., confirming that discovering continuous latent states ZZ alleviates the discrete language bottleneck.

MethodV*HRBench4KHRBench8KAvg.
Qwen2.5-VL-7B76.4468.0063.7569.40
+ SFT81.6868.3861.6370.56
Standard variational training (ELBO)82.2066.7564.8871.28
w/o Rev-KL (NTP + Fwd-KL)82.7272.1267.7574.20
AMVL (Ours)84.2972.1268.5074.97

2. Source 2 — Mitigating Answer Leakage via Lrev\mathcal{L}_{\text{rev}} (Why Our Experiment Establishes a Direct Causal Effect): In observational evaluations, a correlation between lower leakage metrics and higher accuracy could be confounded by other training dynamics. To prove a causal effect without requiring formal causal-inference background, we follow the standard scientific principle of a controlled intervention (manipulation): keep the image xx, question prompt, and model weights 100% identical, manually change only the latent variable ZZ along the direction of a wrong answer at test time, and observe whether the model's prediction changes.

  • Why This is a Pure Intervention on Answer Leakage (Appendix K.5, Figure 5): For a fixed input xx, we compute the latent shift within the model's own posterior when fed a wrong distractor answer y−y^- versus the true answer y+y^+: d(x)=μϕ(x,y−)−μϕ(x,y+),zint=μθ(x)+d(x).d(x) = \mu_\phi(x, y^-) - \mu_\phi(x, y^+), \qquad z_{\text{int}} = \mu_\theta(x) + d(x). Because the input xx is identical in both μϕ(x,y−)\mu_\phi(x, y^-) and μϕ(x,y+)\mu_\phi(x, y^+), the image/question representation μϕ(x)(x)\mu_\phi^{(x)}(x) cancels out. Thus, d(x)d(x) isolates the pure latent shift caused solely by changing the answer label. At test time, we feed zintz_{\text{int}} into the decoder while giving it only the original image and question.

  • Establishing Causality against Controls: If a model relies on answer-leaked shortcuts, manually pushing its test-time latent along d(x)d(x) will directly cause its output to flip to the injected wrong answer y−y^-. To rule out generic noise sensitivity, we compare d(x)d(x) against equal-norm random noise and question-paraphrase shifts. In Figure 5(a), both w/o Rev-KL and AMVL initially predict "Tree", but injecting d(x)d(x) toward "Gazebo" causes w/o Rev-KL to flip to "Gazebo" while AMVL retains "Tree". In Figure 5(b), random noise and paraphrase shifts leave both models' predictions unchanged, proving the model is not simply brittle to noise; yet injecting d(x)d(x) toward "4 knobs" specifically hijacks w/o Rev-KL from "5" to "4", whereas AMVL stays visually grounded at "6".

  • Dataset-Wide Counterfactual Replacement (Appendix K.4): Across the entire V∗V^{\ast} benchmark, actively replacing the test-time latent with a wrong-answer posterior sample z−∼qϕ(z∣x,y−)z^- \sim q_\phi(z \mid x, y^-) under fixed input xx shows that Lrev\mathcal{L}_{\mathrm{rev}} slashes posterior answer sensitivity by 76.6% (Dcf:7.77→1.82D_{\mathrm{cf}}: 7.77 \to 1.82) and directly causes a +9.5% recovery in downstream accuracy ($21.0% \to 30.5% $ CLRA).

  • Architectural Gradient Isolation (Section 4.3 & Appendix J): Because Lrev=DKL(sg[pθ]∥qϕ)\mathcal{L}_{\mathrm{rev}} = \mathcal{D}_{\mathrm{KL}}(\mathrm{sg}[p_\theta] \parallel q_\phi) applies a stop-gradient on pθp_\theta, it has zero direct gradient on the inference prior pθp_\theta. Its performance gain over NTP + Fwd-KL can mathematically only propagate by regularizing qϕ(Z∣x,y)q_\phi(Z \mid x, y) first, which then supplies a less answer-contaminated target to pθp_\theta via Lfwd\mathcal{L}_{\mathrm{fwd}}.

Together, these controlled manipulations prove that suppressing answer-induced posterior shortcuts is the direct causal mechanism behind AMVL's advantage over standard ELBO and forward-only variational training.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal, Discussion Outcomes, and Manuscript Revisions

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their constructive feedback and the Area Chairs for their time and service. We are grateful that Reviewer 5Kz1 has raised their score after two rounds of constructive engagement, Reviewer V17G has confirmed that all questions were addressed while maintaining their positive score, and we have also provided a comprehensive 4-part empirical and theoretical response to address the insightful questions from Reviewer yRqB.

Below is a concise summary of our new experiments, analyses, and revisions (highlighted in blue in the updated PDF):

1. Cross-Backbone, Cross-Scale (Up to 32B), and Cross-Dataset Scaling (Appendix K.2, Tables 16–18; for yRqB & V17G): Using AMVL’s default delayed reverse-KL schedule with zero hyperparameter retuning, we verified scalability beyond Qwen2.5-VL-7B:

  • LLaVA-v1.5 (7B & 13B, Table 16): Full-parameter AMVL training consistently improves V∗V^*, HRBench4K, and HRBench8K at both 7B (+3.52 Avg.) and 13B (+1.73 Avg.).
  • Scaling to Qwen2.5-VL-32B (~33.5B total params, Table 17): Using parameter-efficient LoRA adaptation on the 32B backbone alongside our variational heads (k=8k=8), AMVL yields consistent gains across all 5 perception and reasoning benchmarks (+1.10 Avg. on V∗V^*, HRBench4K/8K, MMMU, and VisualPuzzles), confirming that scale alone does not eliminate train–inference latent mismatch.
  • New Training Corpus (MMFineReason-123K, Table 18): Training from scratch on MMFineReason-123K boosts Qwen2.5-VL-7B by +7.29 Avg. across BLINK tasks (+14.29 on Multi-View Reasoning) and ScienceQA (+3.98).

2. Variational Baseline Comparisons & Multi-Seed Statistical Rigor (Appendices I & K.8; for 5Kz1 & V17G):

  • Standard ELBO & DreamerV2-Style KL Balancing (Appendix I, Tables 8–9): Under identical architecture, data, and schedules (βmax⁡=1.0\beta_{\max}=1.0, canonical α=0.8\alpha=0.8), AMVL (74.97 Avg.) substantially outperforms unblocked Standard ELBO (71.28 Avg., +3.69) and DreamerV2-style balancing (70.87 Avg., +4.10). Reconciling Table 9 with Table 3 isolates two distinct mechanisms: (i) stop-gradient decoupling (sg[qϕ]\text{sg}[q_\phi] in NTP + Fwd-KL: 74.20 Avg.) prevents uncalibrated early prior gradients from distorting the posterior; and (ii) reversing the second KL into DKL(sg[pθ]∥qϕ)\mathcal{D}_{\text{KL}}(\text{sg}[p_\theta] \Vert q_\phi) (AMVL: 74.97 Avg.) suppresses answer-shortcut concentration.
  • Multi-Seed Evaluations & Full 14-Subtask BLINK (Appendix K.8, Tables 24–25): Across 3 evaluation seeds (42, 43, 44), AMVL achieves 74.9 ± 0.2% Avg. on V∗V^*/HRBench4K/8K—outperforming Baseline-SFT (70.9 ± 0.5%, >7$\sigma$ separation) and all continuous-latent baselines (LVR: 70.9 ± 1.3%, Mull-Tokens: 71.2 ± 0.0%, Monet: 72.1 ± 1.3%). Across all 14 BLINK subtasks (Table 25), AMVL achieves the highest macro-average of 56.8 ± 0.3% (+8.6 over Baseline-SFT).

3. Causal Latent Interventions, Test-Time Sampling & Efficiency (Appendices J & K.3–K.7; for yRqB & V17G):

  • Controlled Answer-Shift Interventions (Appendix K.5, Figure 5) & Leakage Diagnostics (Appendix K.4, Table 20): Injecting the pure within-posterior answer-shift vector d(x)=μϕ(x,y−)−μϕ(x,y+)d(x) = \mu_\phi(x, y^-) - \mu_\phi(x, y^+) into the test-time prior latent zint=μθ(x)+d(x)z_{\text{int}} = \mu_\theta(x) + d(x) directionally hijacks the w/o Rev-KL model toward the wrong target y−y^-, whereas AMVL resists the shortcut and retains the visually grounded answer (while equal-norm random and paraphrase controls rule out generic noise sensitivity). Dataset-wide, Lrev\mathcal{L}_{\text{rev}} cuts counterfactual posterior sensitivity by 76.6%, lowers answer-alignment margin by 60.6%, and boosts counterfactual robust accuracy by +9.5 points.
  • Test-Time Prior Sampling & Ensembling (Appendix K.3, Table 19): Combining the prior mode μθ(x)\mu_\theta(x) with M=4M=4 stochastic prior draws improves average accuracy by +2.18 points, including a +4.60 point (+20.6% relative) gain on MathVision.
  • Schedule Sensitivity & Inference Efficiency (Appendices J & K.7, Tables 13 & 23): AMVL converges smoothly across all tested reverse-KL delays (delay∈0,500,1000,2000\text{delay} \in \\{0, 500, 1000, 2000\\}) and loss weights. At inference, AMVL adds only +1.8% FLOPs and +0.16 GB memory over base Qwen2.5-VL-7B, while running 3.15$\times$ faster than discrete tool-augmented Pixel-Reasoner and using 52.5% less GPU memory than Monet.

Thank you again for your time and consideration.

Best regards,
The Authors of Submission 36556

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.