acceptodds
Under review as a conference paper at ICLR 2027

AutoFlow: Autoregressive Normalizing Flows for Next Vector Prediction

30 views

Abstract

Next-token prediction gives language models a simple interface for both understanding and generation. Extending this idea to continuous signals such as images and robot actions is challenging because these signals are continuous vectors rather than discrete tokens. We introduce Autoflow, a framework for next-vector prediction in autoregressive foundation models. The key idea is to equip a language model with a conditional normalizing flow that predicts the next continuous vector while leaving the autoregressive backbone unchanged. We find that this simple formulation better connects generation and understanding than diffusion-based alternatives: models trained for image generation transfer more effectively to visual understanding, while also supporting interleaved image–text reasoning and robot action prediction. Autoflow further generates continuous outputs in a single forward evaluation and provides their likelihood directly. Our results suggest that next-vector prediction via Autoflow is a promising interface for extending autoregressive foundation models beyond discrete tokens.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Next-Vector Prediction, Conditional Normalizing Flows, Autoregressive Foundation Models, Multimodal Learning

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_YYb9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AutoFlow, which attaches a conditional normalizing-flow head to an autoregressive language model to predict the next continuous vector. The head is a channel-wise affine coupling conditioned on the backbone state hth_t, trained by exact negative log-likelihood on clean targets alongside next-token cross-entropy (Eq. 5). Sampling a continuous vector takes a single flow evaluation. The authors argue that diffusion-based alternatives, either a MAR-style diffusion head or Transfusion-style denoising inside the backbone, give the shared representation weaker or mismatched learning signals. They test this on a Qwen2.5-1.5B backbone trained on ~10M image–text pairs at 256px, comparing the three heads under matched data and training budgets:

  • After text-to-image-only pretraining and per-task fine-tuning, AutoFlow averages 44.0 over eight understanding benchmarks, vs. 39.6 (Diffusion Head) and 32.0 (Transfusion) (Table 1).
  • On ThinkMorph navigation/Jigsaw it reaches 38.8/46.8 with one sampling step, vs. 30.8/45.5 for a 20-step Diffusion Head (Table 2).
  • It attains FID 31.4 in one step, vs. 27.1 for the 10–20-step baselines (Fig. 4).
  • Gradient diagnostics are reported in Table 3.

On three real-robot tasks with a Qwen3-VL-2B backbone, an AutoFlow action head succeeds in 60/90 rollouts, vs. 36/90 for flow matching and 33/90 for diffusion (Table 4).

Critical strengths

1. A simple interface with promising results under a shared backbone, data and training budget. AutoFlow and the Diffusion Head share the same backbone, data and 200K-step budget, with the continuous-output head as the main architectural difference (Sec. 4.1). Under these conditions AutoFlow transfers better to understanding (Table 1: 44.0 vs. 39.6 average). It achieves higher navigation accuracy with 1 vs. 20 sampling steps (Table 2: 38.8 vs. 30.8), and it captions better after co-training (63.5 vs. 61.6 CIDEr, Fig. 4c). The robot study provides additional evidence: it uses 30 rollouts per task and reports pooled Wilson intervals that do not overlap ([56.4, 75.5] vs. [30.5, 50.3] for flow matching, Table 4), even though AutoFlow uses a smaller head. Together, these results make a credible case that the choice of continuous-output head can affect downstream performance.

2. One-step sampling with exact likelihood at competitive quality. AutoFlow reaches FID 31.4 in one forward pass (1.6 TFLOPs/image, App. C). The baselines reach 27.1 but need 10–20 steps, up to 28.1 TFLOPs for Transfusion. AutoFlow also needs a single step per action chunk. This matters for interleaved reasoning and control loops, where generation is called repeatedly, and the paper demonstrates both settings.

Critical weaknesses

1. Further controls would help clarify the proposed explanation for the gains. The thesis is that exact likelihood on clean targets gives the backbone a better signal than denoising (Sec. 3.1, Sec. 4.5). The current experiments leave some uncertainty about its contribution relative to other factors:

  • (i) A useful complementary comparison would be a one-step, exact-likelihood mixture head on clean targets, such as a GIVT-style head or the GMM head discussed for JetFormer. Such a comparison could help distinguish benefits associated with clean-target likelihood training from those associated with the particular coupling-flow architecture.
  • (ii) AutoFlow's generation pretraining uses noise-perturbed context with a noise-level embedding (s≤0.25s \le 0.25, App. F.2). Sec. 3 does not mention this, and it is unclear whether the Diffusion Head uses the same perturbation.
  • (iii) Clearer reporting of the loss weight λ, method-specific image-head architectures and parameter counts, and the FID protocol would help readers reproduce and interpret the results. Appendix C mentions a 170M-parameter generation head but does not clearly map the head sizes to each method.
  • (iv) Table 3 compares gradient magnitudes and variances across different losses (flow NLL vs. ε-MSE). These quantities can depend on loss scaling, so the reported "20× larger" and "293× lower" ratios are difficult to interpret without precise definitions and normalization. The cosine comparison is less sensitive to uniform loss scaling, but the measurement procedure still needs specification.

The diagnostics in Sec. 4.5 provide useful initial evidence; further controls would help establish how specifically they support the proposed mechanism.

2. Uncertainty estimates and a more task-specific interpretation would strengthen the image-side results. The fine-tuning configurations for Tables 1–2 specify seed 0 (Tables 6, 8), with 2k–6k fine-tuning steps, and no variance is reported. Roughly 86% of the summed gain over the Diffusion Head in Table 1 comes from RefCOCO/RefCOCO+ (+14.1/+16.2). Those two scores come from the same RefCOCO-fine-tuned checkpoint (Table 7). The other six tasks differ by 0–1.8 points. The Jigsaw improvement is modest (46.8 vs. 45.5 on 400 problems), and its statistical reliability is unclear without uncertainty estimates or a paired comparison. Given the concentration of the largest gains in grounding, the broad framing ("better connects generation and understanding", "a new path towards omnimodal generative models") could be more closely aligned with the scope of the current evidence. Additional fine-tuning seeds for grounding and navigation would strengthen the claim.

Questions for authors
  1. [Suggestion] A one-step, exact-likelihood mixture head (e.g., a GIVT-style GMM) trained on clean targets under the same backbone, data and budget would be a useful additional comparison, at least for the T2I-only → Table 1 protocol. It would help distinguish benefits associated with clean-target likelihood training from those of the particular coupling-flow head. This would be especially useful for assessing the proposed mechanism.
  2. [Clarification] App. F.2 mentions "lightly perturbed context with s≤0.25s \le 0.25" and a noise-level embedding during AutoFlow's generation pretraining. Could you clarify what is perturbed, how ss is sampled, and whether the Diffusion Head receives the same perturbation? It would also help to specify λ, the image-head architecture and parameter count for each method, and the FID protocol (sample count, reference set and guidance).
  3. [Clarification] For Tables 3 and 10, could you provide the definitions of intrinsic gradient variance, conditioning-gradient magnitude and gradient cosine, including the batches and layers measured? It would be useful to understand how the different loss scales (flow NLL vs. ε-MSE, and λ) are accounted for when comparing magnitudes and variances.
  4. [Suggestion] Additional fine-tuning seeds for RefCOCO (including its RefCOCO+ evaluation) and spatial navigation would help establish the reliability of the main gains. For Jigsaw, uncertainty estimates or a paired comparison on the existing 400 problems would help interpret the 46.8 vs. 45.5 difference.
  5. [Clarification] Could you provide the number of demonstrations per task in the robot study and describe how the baseline head hyperparameters and sampling budgets were selected? In particular, were the 16-step Diffusion and 5-step Flow Matching operating points chosen using the 10-rollout sweep in Fig. 10?
Reviewer_FEhbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies continuous next-vector prediction by attaching a conditional normalizing-flow head to an autoregressive language model. The head is trained by maximum likelihood and samples each continuous vector with one flow evaluation. Experiments cover generation-to-understanding transfer, interleaved visual reasoning, image generation, and three real-robot manipulation tasks. The strongest transfer gains appear in visual grounding. Gradient diagnostics are used to explain why this objective might provide more useful supervision to the shared backbone. For images, the exact likelihood is defined over VAE latents.

Critical strengths

The strongest result is the generation-to-understanding transfer experiment. Starting from T2I-only checkpoints avoids direct I2T supervision during pretraining, and the methods share the downstream fine-tuning settings. AutoFlow improves over the Diffusion Head by 14.1 points on RefCOCO and 16.2 on RefCOCO+. These gains provide a concrete reason to take the proposed training interface seriously.

The sampling-efficiency claim is supported by both image and robot experiments. On the robot tasks, AutoFlow succeeds in 60/90 rollouts, compared with 36/90 for Flow Matching, while using a smaller action head. Reporting FLOPs and rollout confidence intervals is helpful.

Critical weaknesses

I am not yet convinced by the explanation that the likelihood objective improves transfer through more stable and stronger backbone supervision. Table 3 does not define the variance estimator, the parameter subset being measured, or the loss normalization. Raw gradient magnitudes and unnormalized variances can change simply by rescaling a loss, so larger or smaller values do not by themselves establish more useful supervision. The gradient cosine is more informative, but remains correlational.

The image experiments compare AutoFlow with diffusion-based alternatives, but not with a Gaussian or Gaussian-mixture likelihood head. This control would help separate the benefit of direct next-vector likelihood training from that of the invertible architecture. The robot positioning also needs attention: NinA already uses a normalizing-flow action expert for one-shot VLA sampling, but is neither discussed nor compared here.

Questions for authors

1.How are the intrinsic gradient variance and conditioning-gradient magnitude computed? Please specify the estimator, parameter subset, loss reductions and weights, and number of measurement batches. Does the advantage remain after controlling for loss scale and using the T2I-only checkpoints underlying Table 1?

2.Does a Gaussian-mixture head trained on the same targets recover a substantial part of the transfer gain? Please also explain how the action-head formulation differs from NinA and what additional contribution the robot experiments establish.

3.How much of the improvement remains when intermediate image generation is removed or the generated image is withheld from the final answer stage?

1 comment

Thanks for the detailed response and additional experiments. The GMM likelihood baseline, T2I-only gradient-variance measurements, and generated-image withholding ablation substantially address my main concerns and strengthen the empirical evidence for the proposed training interface. I also appreciate the clarification of the relationship to NinA. I am therefore raising my recommendation from weak rejection to weak accept.

Reviewer_pqQCOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AutoFlow, which employs a conditional normalizing flow head to directly model the likelihood of continuous vectors, aiming to improve the transfer from generation to understanding. AutoFlow learns directly from clean targets without relying on noise-conditioned denoising objectives, while enabling single-step sampling. Experiments demonstrate improvements in downstream visual understanding, interleaved visual reasoning, and robotic action prediction, along with competitive image generation quality using fewer sampling steps. Additional analyses suggest that AutoFlow provides more stable gradients, stronger conditioning signals, and better alignment between generation and understanding objectives.

Critical strengths
  • Extending multimodal autoregressive models to support both generation and understanding is an important and meaningful research direction, with potential benefits for unified multimodal modeling.
  • The proposed method is simple yet effective, integrating a conditional normalizing flow head into existing autoregressive models without requiring substantial modifications to the backbone architecture.
  • The paper provides relatively comprehensive experiments across multiple tasks, demonstrating the effectiveness of the proposed method. Additional analyses offer empirical evidence supporting the authors' hypothesis regarding the advantages of likelihood-based training over diffusion-based approaches.
Critical weaknesses
  • Although the paper compares AutoFlow with diffusion-based approaches, the comparisons with other flow-based generative methods are insufficient. While a baseline is included in the robotic action prediction experiments, it is missing from the image generation and generation-to-understanding transfer evaluations. Including such comparisons would help better establish the advantages of the proposed method.
  • The paper lacks direct comparisons with other representative unified multimodal models, such as those based on discrete visual tokens. Including these baselines would help better establish the competitiveness of AutoFlow in the broader context of unified multimodal generation and understanding.
  • The experiments are conducted with a limited selection of backbone models, without exploring different model architectures or scales. Evaluating AutoFlow with additional backbones would help demonstrate the generalizability and scalability of the proposed method.
Questions for authors

Please refer to the weaknesses above, particularly the first two points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.