acceptodds
Under review as a conference paper at ICLR 2027

LogicVLA: Neural Logic Critics for Efficient Vision-Language-Action Fine-Tuning

Abstract

Vision-language-action (VLA) models achieve strong performance in robotic manipulation, yet downstream adaptation remains computationally expensive. A fundamental limitation of standard VLA fine-tuning is its reliance on pointwise behavioral cloning, which penalizes numerical deviations from expert actions but provides no explicit supervision on whether predicted actions are compatible with the task instruction and current state. To solve this challenge, this paper develops LogicVLA, which is a training-time logical regularization framework that complements imitation learning with task-conditioned action consistency. Particularly, a lightweight neural logic consistency critic maps language instructions, visual observations, proprioceptive states, and actions into a shared latent space, composes their interactions through differentiable logic operators, and produces a consistency score that measures the compatibility between the multimodal task context and the predicted action. The critic contains 737K parameters and accounts for only 0.01% of the VLA backbone. It is trained directly from the original demonstrations by contrasting expert tuples with in-batch mismatched tuples, without requiring additional annotations or environment interaction. During fine-tuning, the learned consistency score provides an auxiliary regularization signal for predicted actions, while a dedicated stability mechanism prevents unstable critic gradients from affecting policy optimization. Extensive simulation results on LIBERO-Spatial show that, compared with OpenVLA, LogicVLA improves closed-loop success by 2.6% at 15K training steps and by 26% on the complex tasks, while introducing only 0.0096% additional trainable parameters during fine-tuning and no additional deployment-time computation. Code is available at https://anonymous.4open.science/r/LogicVLA-379E.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action, robotic manipulation, neural logic, consistency regularization, imitation learning

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_ReqsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LogicVLA introduces a task-conditioned neural logic consistency critic (NLC critic) beyond behavior cloning, evaluating whether predicted actions are compatible with language instructions, visual observations, and robot state. Trained on matched and in-batch mismatched samples from the original demonstrations and removed after training, the critic adds no deployment-time parameters or computation. Using OpenVLA-7B and OFT fine-tuning, the authors evaluate on LIBERO-Spatial with three seeds, plus ablations and paired statistical tests. At 15K steps, overall success rises from 94.4% to 97.0%, and success on the fixed hardest (stove) task from 68% to 94%. LIBERO-Long experiments mainly verify co-training stability, without a complete closed-loop comparison.

Critical strengths
  1. Task–action compatibility serves as auxiliary supervision for direct policy optimization. The key contribution is applying neural logic to supervise VLA-generated actions, so the objective accounts for action–context relations and supplements pointwise behavior cloning. This supervision learns directly from existing demonstrations without extra annotations, reward design, or environment interaction, offering a valuable direction for better demonstration-data utilization.

  2. Meaningful early-training gains are achieved without deployment burden. The critic is used only during fine-tuning, with the original policy structure retained after deployment. In Table 1, success at 4K steps rises from 84.6% to 91.6%, with main results covering three seeds, indicating potential to improve performance under limited training budgets in the current setting.

Critical weaknesses
  1. Key controls are missing, so the logical structure cannot yet be confirmed as the main source of improvement. Table 2 mainly compares the full method with removing the entire consistency supervision; it neither compares against an ordinary scoring network of similar parameter count nor separately removes the logic-algebra regularization (Eq. 9). The controls are not fine-grained enough to isolate the contributions of logical composition/operators versus logic-algebra regularization.

  2. The added critic introduces training sensitivity, with insufficient evaluation of its stability and efficiency cost. Table 2 shows success dropping to 1.8% without warmup and to 0% with consistency weight 1.0, indicating the auxiliary supervision may substantially interfere with policy learning. Existing ablations use only a single seed and do not verify each stabilization mechanism separately. Moreover, without reported training time or total computation, fewer training steps do not directly establish a proportional reduction in wall-clock time or total computation.

Questions for authors
  1. With the same inputs, negative-sample construction, warmup, and stabilization mechanisms, how does a plain MLP scoring network with a comparable parameter count perform compared with LogicVLA? How does performance change when the logical-identity penalty in Eq. (9) is removed while retaining the latent-magnitude regularization term?
  2. Why is an expert action chunk from another sample in the same minibatch treated as a valid negative for the current context? Has there been any analysis of “mismatched but still reasonable” actions and their potential false-negative effect on critic learning?
  3. Can ablations of warmup, consistency weight λ\lambda, and each stabilization component be provided across multiple training seeds, to show whether the results depend on a specific random seed or a narrow hyperparameter setting?
  4. Please clarify how the McNemar tests in Table 5 account for the three training seeds: are they computed per seed, on pooled episodes across seeds, or using a single selected training seed? Please also report the discordant counts (b, c) or more complete statistical results.
  5. On the same hardware (4×A800-80GB), what are the wall-clock time/GPU hours required to reach comparable success rates, and what is the per-step training overhead of the auxiliary critic?
2 comments

W1: Separating logical composition from regularization

Removing consistency supervision alone does not isolate logical composition from algebraic regularization, so we added two controls — a plain MLP (structured design versus a generic scorer) and removal of the identity penalty. Both underperform the full model at 15K under the same stabilized setup; Q1 gives the design, table, and reference.

W2: Training sensitivity and critic overhead

The 1.8% (no-warmup) and 0% (λ = 1.0) results in Table 2 used the earlier recipe, before three-layer stabilization; they show that recipe was sensitive to how the critic was introduced, not whether it still collapses once stabilized — with all three mechanisms the full model is repeatable across seeds (Q3). The added cost is ~+23% per step, so a tenfold step reduction is not a tenfold compute reduction (Q5).

Q1: Plain MLP and Eq. (9) ablations

The plain MLP replaces the learned logical composition and truth-vector readout with a feed-forward scoring network over the same four projected modality vectors (736,873 parameters versus 736,833 for the full critic). We held the backbone, input projections, pairs, ranking margin, warmup, schedule, and all three stabilization mechanisms fixed; for the Eq. (9) control we set α = 0 while retaining the latent-magnitude term weighted by β, keeping the architecture unchanged.

CriticSuccess at 15KDifference from fullDiscordant counts (b, c)Paired p
Full neural-logic critic99.2%———
Parameter-matched plain MLP96.6%−2.6 pp(16, 3)0.006
No logical-identity penalty97.2%−2.0 pp(12, 2)0.016

Counts and continuity-corrected McNemar p-values are from 500 paired rollouts for one usable seed. The MLP tests whether generic compatibility scoring recovers the full result under matched conditions; the Eq. (9) control isolates removing the identity penalty. All rows use stabilization, distinct from the pre-stabilization Table 2 results.

Q2: Plausible negatives and false-negative supervision

We treat each permuted expert chunk as a candidate negative because it was not demonstrated for the current context, not because it is known invalid. Reassigning in-batch chunks by ρ(b)=(b+⌊B/2⌋) mod B\rho(b)=(b+\lfloor B/2\rfloor)\bmod B preserves the action distribution and avoids a real-versus-noise shortcut, but can produce plausible negatives when two contexts admit similar actions.

The ranking hinge max⁡(0,m−(s+−s−))\max(0,m-(s^+-s^-)) yields zero loss once the positive exceeds the negative by at least the margin; if a plausible negative scores high and the gap is below the margin the pair is still penalized, so the hinge does not eliminate false-negative supervision. The objective scores relative compatibility, not action validity.

Q3: Stabilization and repeatability across seeds

We evaluated the full three-layer stabilization on two training seeds: at 15K the full model achieves 99.2% and 98.4% for seeds 2 and 42. The warmup and consistency-weight results in Table 2 characterize the sensitivity of the earlier recipe; the two stabilized runs show similar success across those seeds but do not isolate the effects of warmup, λ, or any single mechanism.

Q4: Per-seed McNemar tests and discordant counts

We applied McNemar's test separately to each of the three training seeds in Table 5: for a given seed, the 500 matched rollout conditions for LogicVLA and its baseline form one paired contingency table — not pooled across seeds, not a single selected seed. For the new single-seed critic controls, bb = paired rollouts where the full critic succeeds and the control fails, cc = the opposite (the (b, c) columns of the Q1 table); these address paired evaluation outcomes given the trained checkpoints, not uncertainty across training seeds.

Q5: Four-GPU timing and per-step overhead

On four A800-80GB GPUs at batch size 16 per GPU, the no-critic OFT run averages 2.70 s per step versus ~3.31 s for LogicVLA — about 0.6 s more, or roughly 23% additional training cost per step. Because the logged runs use different GPU counts, we compare GPU-hours, not raw wall-clock.

Run on LIBERO-SpatialGPUsStepsWall-clock timeGPU-hours
LogicVLA, logged run215K27.7 h55.5
Matched no-critic OFT, logged run415K~11.3 h45.1
LogicVLA, four-GPU timing estimate415K~13.8 h~55
Published OFT recipe, extrapolated from the matched run4150K~113.0 h~451

At the four-GPU step rates, 15K LogicVLA ≈ 55 GPU-hours and 150K OFT ≈ 451 (table) — estimates at the respective recipe lengths, not timed runs to an identical threshold. With reported Spatial outcomes of 97.0% and 97.6%, LogicVLA uses roughly one-eighth of the published OFT recipe's GPU-hours at similar success; the critic's cost is training-only.

Dear Reviewer Reqs,

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. We remain committed to addressing any remaining points you may have during the discussion phase.

If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly.

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

LogicVLA adds a neural logic critic on top of behavior cloning for VLA models. The critic takes the language instruction, images, and robot state, and checks whether a predicted action actually fits them, using learned AND, OR, and NOT operations to output a consistency score. It is trained contrastively: expert context-action pairs should score higher than mismatched pairs from the same batch. After a warm-up, the score gets added to the behavior cloning loss. Once training is done, the critic is removed, so inference cost stays the same. They evaluate with OpenVLA-7B on LIBERO-Spatial. LogicVLA beats the baseline both early and later in training, with the largest gain on the hardest stove task. LIBERO-Long is used to testtraining stability.

Critical strengths
  1. The comparison against the baseline is well controlled: same backbone, same hardware, same training setup. The main numbers are averaged over three seeds, and the gains are already there at early training steps, plus on the hardest task.
  2. The method is practical. The critic is trained from the original demonstrations alone, with no extra labels and no environment interaction, and it gets removed after training, so the final policy costs nothing extra at inference.
Critical weaknesses
  1. The main issue is that the paper does not clearly show the gain comes from the neural logic itself. There is no baseline with a simple MLP critic that sees the same inputs and is trained on the same pairs, just without the AND/OR/NOT operations or the logic rules. Since the negatives are just mismatched actions from the batch, a simple context-action matching model would probably learn this too. So the results tell us "adding a critic helps", but not "the logic design specifically helps".
  2. The efficiency claim is not fully supported. The 15K vs. 150K comparison is done in training steps, but we never see training time or total compute, and LogicVLA runs an extra critic update every step. Also, the main closed-loop results use one backbone and LIBERO-Spatial only. The evidence behind the main improvement is limited, and the ablations aren't repeated across training runs, so it's hard to know whether the gains are stable or would generalize.
Questions for authors
  1. How does LogicVLA do against a plain MLP critic trained on the same positive and negative pairs, but without the neural logic operations and rules?
  2. How much wall-clock time or total compute does LogicVLA need to hit its 15K result? How does that stack up against the baseline and OpenVLA-OFT?
  3. How are the mismatched actions picked? Can they come from different tasks? If so, couldn't the critic just learn coarse task/action differences instead of fine-grained action consistency?
2 comments

W1: Contribution of the logical structure

We added two controls to separate the contribution of logical structure: a capacity-matched MLP on the same positive and negative examples (Q1), and a variant keeping the neural operators but removing the logical-identity term in Eq. (9) while retaining the latent-magnitude term. The latter reaches 97.2% at 15K versus 99.2% for the full critic (paired McNemar p = 0.016), giving separate evidence for the identity regularization under the tested protocol.

W2: Training efficiency, repeatability, and evaluation scope

We quantified training cost in GPU-hours as well as steps: at 15K LogicVLA uses 55.5 GPU-hours versus 45.1 for the matched no-critic OFT run; the lower total compute against OFT's published 150K recipe comes from the shorter schedule, not a cheaper update (Q2 separates the measured runs from the extrapolated cost).

We also evaluated the full model across two seeds under the stabilized recipe: with all three mechanisms enabled it reaches 99.2% and 98.4% at 15K for seeds 2 and 42, while the new controls each have one usable seed. The paired evaluation supports a difference for the evaluated checkpoints but does not establish variation in the ablation effect across training runs. Our same-budget UniVLA and π0 runs improve method comparison on LIBERO-Spatial but do not test transplanting the critic to another backbone; closed-loop evidence remains limited to OpenVLA-7B/OFT on LIBERO-Spatial.

Method4K10K15KStove at 15K
LogicVLA91.6%93.8%97.0%94%
UniVLA76.8%93.0%95.2%84%
π080.8%92.0%92.6%92%

The gap is largest at 4K; at 15K the paired difference is statistically distinguishable for π0 (McNemar p = 0.0015) but not for UniVLA (p = 0.16).

Q1: Comparison with a parameter-matched MLP

We evaluated the plain MLP against the full critic on 500 paired rollouts at 15K: 96.6% versus 99.2%, a 2.6-point difference (continuity-corrected McNemar p = 0.006). It receives the same four projected inputs and the same matched and permuted expert pairs with the same ranking objective, replacing the AND/OR/NOT composition and truth-vector readout with a feed-forward score and no logical-identity rules; parameters are 736,873 versus 736,833, with warmup, schedule, and stabilization unchanged. With capacity and examples matched, the gap supports a benefit from the structured critic in this run, while the single usable seed limits the conclusion's strength.

CriticSuccess at 15KDifference from fullDiscordant counts (b, c)Paired p
Full neural-logic critic99.2%———
Parameter-matched plain MLP96.6%−2.6 pp(16, 3)0.006
No logical-identity penalty97.2%−2.0 pp(12, 2)0.016

Q2: Wall-clock time and GPU-hours

We measured 27.7 h of wall-clock for the 15K LogicVLA run on two A800 GPUs (55.5 GPU-hours at effective batch 64) and 45.1 GPU-hours for the matched no-critic OFT run at 15K on four A800 GPUs (~11.3 h wall-clock). Because the logged runs use different GPU counts, raw wall-clock times are not like-for-like, but GPU-hours are: the ratio is 1.23×, and the four-GPU per-step estimates are ~3.31 s for LogicVLA and 2.70 s for OFT.

Run on LIBERO-SpatialGPUsStepsWall-clock timeGPU-hours
LogicVLA, logged run215K27.7 h55.5
Matched no-critic OFT, logged run415K~11.3 h45.1
LogicVLA, four-GPU timing estimate415K~13.8 h~55
Published OFT recipe, extrapolated from the matched run4150K~113.0 h~451

At the published-recipe operating points, LogicVLA reaches 97.0% at 15K and OFT reports 97.6% at 150K; extrapolating the measured OFT step time to 150K gives ~451 GPU-hours versus 55.5 measured for LogicVLA, roughly one-eighth. This is not a directly measured time-to-threshold comparison. The critic is removed at inference, adding no deployment-time parameters or computation.

Q3: In-batch negatives and coarse task cues

We used an in-batch cyclic permutation to construct negatives: for a minibatch of size BB, context bb is paired with the expert chunk at index ρ(b)=(b+⌊B/2⌋) mod B\rho(b)=(b+\lfloor B/2\rfloor)\bmod B. This preserves the batch's marginal distribution of expert chunks while breaking the original context–action pairing, avoiding an easy expert-versus-noise distinction.

The permutation does not condition on task identity, so a reassigned chunk can come from a different task or a different sample of the same task; coarse task cues can therefore help satisfy the ranking objective. The current experiments do not rule out that shortcut or establish fine-grained discrimination conditional on task. We use the score as a soft policy-training signal, not a verifier of action correctness.

Dear Reviewer mSNy,

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. We remain committed to addressing any remaining points you may have during the discussion phase.

If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly.

Reviewer_thi2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper augments behavioral cloning with a lightweight training-time critic that scores the compatibility of language, vision, proprioception, and predicted actions. The critic uses learned neural operators labeled as AND, OR, and NOT, and is trained with matched expert tuples and in-batch mismatched actions. On LIBERO-Spatial, the method improves early-stage fine-tuning performance over an otherwise matched OpenVLA-OFT baseline without adding inference-time cost.

Critical strengths
  • The proposed method only requires very few additional trainable parameters.
Critical weaknesses
  • The current ablation only demonstrates that the complete NLC auxiliary critic improves policy fine-tuning. It does not establish that the improvement is attributable to neural logic. The learned AND, OR, and NOT operators are implemented as generic MLPs, while the critic is trained using a standard matched-versus-mismatched ranking objective. Without parameter-matched non-logical critics and an ablation of the logic-identity regularization, the observed gain could equally arise from generic context-action compatibility learning.

  • The empirical evaluation is rather narrow. The only controlled experiments use a single OpenVLA-7B backbone with the OpenVLA-OFT recipe, mainly on LIBERO-Spatial. Comparisons with other VLA methods are based largely on paper-reported numbers rather than applying LogicVLA to those backbones under matched settings. LIBERO-Long is used only for optimization diagnostics, and no real-robot evaluation is provided. It therefore remains unclear whether the proposed critic generalizes across VLA architectures, action representations, and manipulation domains.

Questions for authors

Please refer to the comments in weaknesses.

2 comments

W1: Neural logic versus generic compatibility scoring

We have added two controls on LIBERO-Spatial to test whether a generic context–action critic can account for the improvement. The first replaces the learned AND/OR/NOT composition and truth-vector cosine readout with a feed-forward MLP over the same four projected modality vectors. The MLP is trained on the same matched and mismatched pairs with the same ranking objective. The two critics have 736,873 and 736,833 trainable parameters, respectively, a difference of 40 parameters (approximately 0.005%). The second control keeps the neural-logic architecture and removes only the logical-identity penalty in Eq. (9) by setting α = 0 while retaining the latent-magnitude term weighted by β. The backbone, warmup, training schedule, and three-layer gradient stabilization are held fixed across all three conditions.

CriticSuccess at 15KDifference from fullDiscordant counts (b, c)Paired p
Full neural-logic critic99.2%———
Parameter-matched plain MLP96.6%−2.6 pp(16, 3)0.006
No logical-identity penalty97.2%−2.0 pp(12, 2)0.016

At 15K steps, the full critic achieves 99.2% success, compared with 96.6% for the plain MLP (−2.6 pp, paired McNemar p = 0.006) and 97.2% for the critic without the identity penalty (−2.0 pp, p = 0.016). The first comparison shows that, in this run, a generic critic with the same supervision and nearly identical capacity does not recover the full model's performance. The second supports a contribution from the identity penalty within the same architecture. These controls provide one usable training seed each. The tests quantify differences across 500 paired evaluation rollouts conditional on that seed, rather than reproducibility across training seeds. We have included both controls and their matched full-model reference in this response.

W2: Benchmark coverage and cross-backbone evidence

We have run UniVLA and π0 on LIBERO-Spatial to supplement the paper-reported comparisons. The runs use a common 15K-step budget, evaluation at 4K, 10K, and 15K, and 500 matched rollout conditions per checkpoint under the seed-7 protocol. The controlled evidence for adding the critic itself remains limited to OpenVLA-7B with OFT fine-tuning on LIBERO-Spatial.

Method4K10K15KStove at 15K
LogicVLA91.6%93.8%97.0%94%
UniVLA76.8%93.0%95.2%84%
π080.8%92.0%92.6%92%

The largest overall gaps occur at 4K, when LogicVLA leads UniVLA and π0 by 14.8 and 10.8 percentage points, respectively. At 15K, the paired difference remains statistically distinguishable for π0 (McNemar p = 0.0015) but not for UniVLA (p = 0.16). This is a method-level comparison under a matched step budget. It does not test whether adding the critic improves either alternative backbone.

The critic takes pooled language and visual representations, proprioception, and an action chunk rather than depending on OpenVLA-specific operators. This design permits adaptation in principle. The LIBERO-Long results are optimization diagnostics, and the closed-loop evidence for the critic remains specific to OpenVLA-7B/OFT on LIBERO-Spatial.

Dear Reviewer thi2,

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. We remain committed to addressing any remaining points you may have during the discussion phase.

If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for the review. The authors have responded to the points you raised, and the response period closes on 18 November (AoE). I would be grateful if you could post a brief reply on whether your concerns are resolved, partly resolved, or still open, so that the authors have a fair chance to address them. If anything still needs clarifying, please ask now while they can still respond.

Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.