acceptodds
Under review as a conference paper at ICLR 2027

FibVLA: An Efficient Temporal Vision-Language-Action Model with Fibonacci Sampling

Abstract

Vision-language-action models (VLAs), which leverage the cognition of multimodal information to infer physical-world actions, provide a generalized solution for embodied AI applications. Conventional VLAs usually concentrate on current digital cognition. While some efforts are made to enhance VLAs’ reasoning capabilities by capturing temporal information, encoding the longcontext history causes an efficiency-decreasing issue. To reconcile the conflict between capturing temporal information and maintaining inference efficiency in VLAs, this paper introduces FibVLA, an efficient framework featuring temporal perception of long-context history. Specifically, we leverage logarithmic hindsight sampling to both proprioceptive states and visual frames to capture long-term temporal dependencies with minimal redundancy. For the action expert, we introduce the flow matching to produce action distributions, and the Fibonacci recurrent inference strategy to generate long-range planning steps based on realtime closed-loop feedback. Experiments demonstrate that FibVLA significantly improves action smoothness and success rates without retraining large-scale visual encoders. Efficiency analysis demonstrates superior real-time responsiveness compared to video-based baselines in real-world evaluations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models,Temporal Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BwNROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FibVLA, an efficient temporal VLA framework that addresses the tension between capturing long-context temporal information and maintaining real-time inference efficiency. The framework introduces three components: (1) Logarithmic Hindsight Sampling, which constructs non-uniform temporal indices following a dense-near, sparse-far pattern with a Fibonacci sparsity constraint; (2) a Channel-wise Temporal Encoding (CTE) module that compresses multi-frame temporal dynamics into a single RGB image via frame differencing and recursive decay, mapping Near/Mid/Far temporal ranges to R/G/B channels; and (3) a Fibonacci Recurrent Inference strategy that binds the action chunk length to Fibonacci terms, enabling KV cache reuse across inference steps. Evaluated on LIBERO (96.8% avg, 95.2% Long), SimplerEnv-Bridge (67.3%), SimplerEnv-Fractal (72.1%), MIKASA-Robo (46.5%), and 7 real-world tasks (85.7/100 score), FibVLA demonstrates competitive performance with 177ms inference latency.

Critical strengths
  1. This paper proposes FibVLA, an efficient temporal VLA framework that addresses the tension between capturing long-context temporal information and maintaining real-time inference efficiency.

  2. The paper is well-structured and clearly written. Figure 1 provides an intuitive motivation for non-uniform sampling. Figure 2 clearly illustrates the overall architecture. The mathematical formulation in §3.3–3.5 is precise, and the proof in Appendix H is self-contained.

  3. The idea of binding action chunk length to Fibonacci terms to enable KV cache reuse is a creative design.

Critical weaknesses

W1. Ablations fail to isolate the Fibonacci recurrent inference contribution from the sampling pattern. Table 5 removes only entire components (w/o Sampling, w/o CTE) and does not separate the Fibonacci KV-cache reuse mechanism from the logarithmic sampling pattern itself. Table 6 compares Fibonacci vs. Logarithmic vs. Long-Short sampling, but the "w/ Logarithmic" variant (201ms, 94.6%) still uses logarithmic sampling—just without Fibonacci alignment. A clean ablation holding sampling points fixed while toggling KV-cache reuse on/off is missing. Since the paper's core efficiency claim rests on this specific mechanism (§3.5, Eq. 8), the evidence does not yet cleanly demonstrate that the Fibonacci alignment—rather than the sampling density or CTE encoding—is the source of the speedup.

W2. Baseline comparison fairness is unclear regarding training data and fine-tuning protocol. Table 12 (Appendix F) shows FibVLA is fine-tuned on benchmark-specific data (e.g., 25,460 Bridge V2 episodes for SimplerEnv-Bridge; 26,152 Fractal episodes for SimplerEnv-Fractal). However, the paper does not state whether baselines (RT-1-X, OpenVLA, π₀, CogACT) are fine-tuned on the same data or evaluated zero-shot as pre-trained generalists. On SimplerEnv-Bridge (Table 2), FibVLA's 67.3% vs. π₀'s 55.7% is difficult to interpret without this clarification. Additionally, HiF-VLA—a direct temporal VLA competitor scoring 94.4% on LIBERO-Long (Table 9, within 0.8% of FibVLA)—appears only in the efficiency table (Table 6) but is absent from the main comparison table (Table 1).

W3. CTE module lacks hyperparameter specification and robustness analysis. The CTE module (§3.4) depends on a binary motion threshold ξ (Eq. 6), maximum intensity τ, and decay parameter δ (Eq. 7), but none of these values are reported in the paper or appendices. The frame differencing approach (Eq. 5) is inherently sensitive to camera noise, lighting changes, and static clutter, yet no sensitivity or failure-mode analysis is provided. Given that removing CTE causes a 4.0% drop on LIBERO-Long (Table 5), understanding the module's operating envelope is essential for assessing reliability.

W4. No statistical significance reporting across any experiment. All results are point estimates without confidence intervals, standard deviations, or significance tests. On LIBERO-Long, FibVLA's 95.2% vs. HiF-VLA's 94.4% (Table 9) is a 0.8% gap over 250 episodes (10 tasks × 25 episodes, per Table 10), likely within random variation. MIKASA-Robo (Table 4) evaluates only 4 tasks. Real-world experiments use 10 trials per task (Appendix E, Eq. 9). Without uncertainty quantification, the reliability of these margins—particularly the 0.8% LIBERO-Long gap and the 4-task MIKASA comparison—cannot be assessed.

W5. Real-world evaluation uses a non-standard metric with limited scope and unclear task selection. The real-world evaluation (Figure 5) employs a custom step-wise scoring rubric (0–2 points per sub-step, 5 sub-steps, 10 trials; Appendix E) rather than standard binary success rate, introducing subjective judgment (e.g., distinguishing "fluent" from "corrective" execution). Only 7 of 15 available tasks are evaluated, with selection criteria unexplained. All 7 tasks are tabletop pick-and-place or stacking variants; contact-rich, deformable-object, or truly dynamic tasks are absent. It is also unclear whether baselines (OpenVLA, OpenVLA-OFT, π₀) are fine-tuned on the same 700-episode dataset or evaluated zero

Questions for authors
  1. For the SimplerEnv-Bridge and SimplerEnv-Fractal results (Tables 2–3), are the baselines (RT-1-X, OpenVLA, π₀, CogACT, etc.) fine-tuned on the same Bridge V2 / Fractal training data as FibVLA, or are they evaluated zero-shot as pre-trained generalist models? Please specify the training protocol for each baseline row.

  2. Section 3.3 states the growth rate rr can be "flexibly adjusted," but Section 3.5 requires exact Fibonacci alignment (ki=ki−1+ki−2k_i = k_{i-1} + k_{i-2}) for KV cache reuse. In your actual experiments, does the sampling sequence satisfy exact Fibonacci equality at all indices, or only approximately? If approximately, how do you handle the misalignment, and what is the impact on inference accuracy?

  3. What are the specific values of the motion threshold ξ\xi, maximum intensity τ\tau, and decay parameter δ\delta used in the CTE module (Eqs. 6–7)? How sensitive is performance to these values? A brief sweep of ξ\xi on LIBERO-Long would be informative.

  4. HiF-VLA appears in Table 6 (efficiency) and Table 9 (LIBERO-Long, 94.4%) but is absent from the main Table 1. Could you include HiF-VLA results on all four simulation benchmarks? Given the 0.8% gap on LIBERO-Long, understanding the broader comparison is important.

  5. For the real-world experiments (Figure 5), are OpenVLA, OpenVLA-OFT, and π₀ fine-tuned on the same 700-episode dataset as FibVLA? If not, the comparison is between a fine-tuned model and zero-shot/few-shot baselines, which would need to be stated explicitly.

  6. How were the 7 real-world evaluation tasks selected from the 15 available? Were they selected before or after observing model performance?

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

FibVLA adds temporal context to a PaliGemma-based VLA through logarithmic history sampling, channel-wise temporal encoding, and Fibonacci recurrent inference. The method compresses motion history into a compact visual representation and aligns historical sampling points across action chunks to reuse features. Experiments cover LIBERO, SimplerEnv, MIKASA-Robo, and real-world manipulation. The authors report 95.2% success on LIBERO-Long and an inference latency of 177 ms.

Critical strengths
  1. Compressing motion history into a compact visual input is a practical way to incorporate temporal information without processing a long sequence of RGB frames. Nonuniform sampling is also well motivated for tasks that depend on both recent motion and earlier observations.
  2. The evaluation covers several simulation benchmarks and real-world tasks. The component ablations provide useful evidence: in Table 5, LIBERO-Long success drops from 95.2% to 88.4% without the sampling strategy and to 91.2% without CTE.
Critical weaknesses
  1. Appendix H does not justify the step from minimum sequence growth to maximum coincidence count or explain why another admissible sequence could not attain the same overlap. This step is central to the claim that Fibonacci sampling is uniquely optimal.
  2. The validity of feature reuse also needs verification. Equation (8) establishes timestamp alignment, but the representation of an overlapping frame may change if CTE is recomputed as the history window moves. Table 6 does not isolate reuse from changes in sampling and action execution, making it difficult to attribute the reported speedup to the caching mechanism.
  3. The baseline configurations are insufficiently explained. OpenVLA-OFT is absent from the main LIBERO table, while Table 9 reports a LIBERO-Long score of 91.0%, compared with 94.5% in the original paper. The reported HiF-VLA score of 94.4% corresponds to its third-view setting, whereas Table 12 lists wrist and third-person cameras for FibVLA. HiF-VLA reports 96.4% with both views. These scores may reflect different configurations, so they do not provide a clear basis for assessing FibVLA’s relative performance.
  4. Table 10 also reports the better seed for methods with multiple runs. For example, Fixed Fibonacci and Standard Adaptive with the Hunyuan 3D VAE achieve 239/250 and 240/250 successes, respectively. Reporting the better seed without variation across runs makes it difficult to determine whether this one-episode difference reflects a consistent sampling benefit.
Questions for authors
  1. How does Appendix H exclude other admissible sampling sequences from achieving the same overlap as Fibonacci sampling?
  2. What ensures that cached features remain valid when the history window changes? Please compare reuse with full recomputation in terms of task performance and latency.
  3. What explains the OpenVLA-OFT score in Table 9? Does FibVLA retain its advantage over OpenVLA-OFT and HiF-VLA with matched camera inputs and evaluation settings?
  4. Can the authors report the mean and variation across all seeds for Table 10 and clarify whether the sampling gains persist on average?
Reviewer_CqpAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FibVLA combines sparse history sampling, channel-wise temporal encoding, and flow matching for robotic action generation. Fibonacci-aligned updates are intended to reduce repeated encoding.

Critical strengths

This paper addresses the coverage of historical information and the cost of redundant encoding within the same design. The experiments cover both routine operations and tasks in MIKASA-Robo that rely on historical cues, which are more aligned with the motivation of introducing memory compared to testing solely on static visible targets.

Critical weaknesses
  1. Baseline reporting is inconsistent: π0 scores 85.2% in Table 1 but 72.0% in Table 9. HiF-VLA reaches 94.4%, only 0.8 percentage points below FibVLA; Table 10 also selects the better seed.

  2. Equation (8) aligns selected frame timestamps, but CTE features depend on multiple frames and their temporal context. Specify cached tensors and compare inference with and without caching under identical sampling.

  3. Table 10 clearly states: ''For methods with multiple seeds, we report the better official seed result''; each configuration has only 250 evaluations, and all seed results or confidence intervals are not provided.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.