acceptodds
Under review as a conference paper at ICLR 2027

MONAD: One Shared Diffusion Vision-Language Model for Continuous Control

Abstract

Vision-language-action policies begin with a vision-language backbone that emits discrete tokens and must end in continuous control. Every current answer pays a different price. A second module holds action-specific parameters, and its gradients degrade the pretrained backbone; the common remedy blocks them at the cost of fixing the separation. A symbol vocabulary commits the trajectory to bins whose errors carry no fixed physical size. One that keeps the backbone and swaps its objective writes a second target into the representation that language modeling trained. We observe that none of these prices is necessary. A diffusion vision-language model already reconstructs masked spans inside its own hidden space, and a complete action vector is a span like any other. We introduce a method that represents each control timestep as one continuous latent token and reconstructs temporally ordered action blocks with the same masked-token procedure the backbone applies to language. Two position-wise projections carry the coordinate mapping and hold 0.45% of model parameters. Action experts in π₀ and π₀.₅ instead allocate 16.5% and 19.2%. Layer probes place the cross-time response inside the shared backbone. The method adds no action expert, tokenizer, or codec, and replaces no pretrained objective. On standard benchmarks, it reaches 98.15% on regular LIBERO, 85.9% on the 10,030-task LIBERO-Plus aggregate, and 4.142 average completed chain length on CALVIN ABC–D.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robot Learning, Masked Diffusion Language Models

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

MONAD represents each continuous action vector as one latent token and reconstructs action blocks using a shared diffusion vision-language backbone with lightweight action interfaces. The approach is evaluated on LIBERO, LIBERO-Plus, and CALVIN, with additional ablations and temporal-response diagnostics.

Critical strengths

The unified architecture generates continuous actions without a separate action expert or pretrained action tokenizer. Competitive results across three simulation benchmarks, particularly LIBERO-Plus, demonstrate the practical promise of this design.

Critical weaknesses
  • The benefit of the proposed reconstruction formulation is not sufficiently isolated. The paper lacks comparisons against direct regression and flow matching under the same backbone, data, and training budget, while the auxiliary-coordinate experiment in Section 4.4 does not directly establish the advantage of retaining the pretrained reconstruction formulation.
  • The headline parameter accounting and temporal-response diagnostics come from legacy non-BARD models rather than the main benchmark checkpoints. Reporting the corresponding measurements for the evaluated BARD models would clarify how these architectural claims relate to the reported performance.
Questions for authors

Please address the concerns raised in the Weaknesses section.

3 comments

[W1, Q1] Isolating masked reconstruction against direct regression and flow matching under the same shared backbone, data, and training budget.

Thank you for this valuable suggestion. We conducted a strictly controlled closed-loop and offline benchmark on the 10 long-horizon multi-stage tasks of LIBERO-10 using the exact same 4.46B shared BARD-VL-4B Pretrain backbone, identical LIBERO-10 training stream, identical trainable parameter count (1.23B [1,234,432,533] trainable parameters: top 12 shared VLM layers + identical 4-layer point-wise I/O bridge and risk head), and identical 2,000-step budget (effective_batch_size = 32, 64,000 trajectory chunks, 31.63 GiB gradient-checkpointed peak VRAM, FP32 master-weight AdamW, lr = 2.5e-5). Across 100 paired closed-loop episodes (10 tasks ×\times 10 paired seeds per arm) and 64 held-out trajectory windows, we compared MONAD (Masked Reconstruction) (self-conditioned draft-and-remask reconstruction over [MASK] action slots, evaluated under both single-pass D1/R0 and risk-guided D4/R2), Shared + Direct Regression (single-pass L1L_1/BCE regression from 100%-masked query slots), and Shared + Flow Matching (conditional optimal-transport velocity matching v=x1−z0v = x_1 - z_0 from z0∼N(0,I)z_0 \sim \mathcal{N}(0, I), evaluated with 6-step Euler integration):

Training Objective (4.46B Shared Backbone, 1.23B Trainable)Decoding ScheduleAction MAE ↓\downarrowGripper Acc ↑\uparrow50% Trajectory Completion MAE ↓\downarrowLIBERO-10 Closed-Loop Success (100 Paired Eps) ↑\uparrowPaired Δ\Delta vs. MONAD (D4/R2)
MONAD (Masked Reconstruction)D4/R2 (Risk-Guided)0.33992.8%0.055 (98.1% grip)91.0% (91/100)—
MONAD (Masked Reconstruction)D1/R0 (1-Pass Draft)0.33992.3%0.055 (98.1% grip)89.0% (89/100)-2.0 pp
Shared + Direct Regression1-Pass Regression0.31693.4%0.326 (98.1% grip)77.0% (77/100)-14.0 pp (95% CI: [-23.0, -5.0])
Shared + Flow Matching6-Step Euler ODE0.447 (1-step: 0.784)80.0%0.195 (88.4% grip)21.0% (21/100)-70.0 pp (95% CI: [-80.0, -59.0])

First, comparing MONAD against Flow Matching directly establishes the advantage of retaining the pretrained masked-reconstruction formulation. Adapting a discrete [MASK]-pretrained VLM backbone with continuous Gaussian interpolants zt=(1−t)z0+tx1z_t = (1-t)z_0 + t x_1 and velocity targets x1−z0x_1 - z_0 creates a severe representation mismatch during downstream adaptation—degrading action MAE (0.447 at 6 Euler steps; 0.784 at 1 step), gripper accuracy (80.0% vs. 92.8%), conditional 50% completion MAE (0.195 vs. 0.055), and paired closed-loop success on LIBERO-10 by 70.0 percentage points (21.0% vs. 91.0%, paired 95% bootstrap CI [+59.0%, +80.0%]). Retaining discrete masked reconstruction preserves seamless alignment with masked multimodal-action pretraining.

Second, comparing MONAD against Direct Regression highlights why closed-loop physical execution and conditional trajectory coupling matter far more than unconditional open-loop L1L_1 error alone. Because Direct Regression dedicates 100% of its gradient updates to the fully masked query state (t=1t = 1), it attains slightly lower unconditional open-loop MAE (0.316 vs. 0.339); however, supervising only t=1t = 1 sacrifices intra-chunk conditional trajectory coupling—eroding 50% trajectory completion MAE from 0.056 at initialization to 0.326 (6.0× higher error than MONAD's 0.055)—and restricts the policy to a deterministic single-pass average across multi-modal contact transitions. By contrast, MONAD's masked-reconstruction objective jointly trains full-mask drafting and partial-mask conditional denoising, regularizing the shared backbone to capture coherent cross-timestep dependencies: even at a single forward pass (D1/R0), MONAD outperforms Direct Regression by +12.0 percentage points in paired LIBERO-10 closed-loop rollouts (89.0% vs. 77.0%), and risk-guided iterative remasking (D4/R2) further lifts closed-loop success to 91.0% (+14.0 percentage points, paired 95% CI [+5.0%, +23.0%]) (92.0% with D4/R4 on our full multi-suite checkpoint in Reviewer tnFY [W2]). In the revised manuscript, we have added this parameter- and budget-matched benchmark against Direct Regression and Flow Matching—together with the analysis of pretraining paradigm alignment and conditional trajectory coupling—to Section 4.5 (Table 2).

[W2, Q1] Reporting exact parameter accounting and finite-difference temporal-response (NLR) diagnostics on the evaluated BARD-VL-4B checkpoints.

Thank you for catching this mismatch. Below we report exact parameter accounting and float32 finite-difference non-local temporal response (NLR) diagnostics measured directly on all four evaluated BARD-VL-4B (MONAD) checkpoints (Pretrain, LIBERO, CALVIN, and bimanual 20D RoboTwin). In the revised manuscript, we have updated the Abstract, Section 1, Section 4.6 (formerly Section 4.5), Section 5, and Appendix B.9 (Tables 14–15) by replacing the earlier 8.35B / 0.45% preliminary figures and NLR table with these BARD-VL-4B measurements so that all architectural and mechanistic statistics align directly with the main benchmark checkpoints.

BARD-VL-4B (MONAD) ModuleParameters% of Total (4.46B)Architectural Specification
Shared VLM Backbone4,437.82M99.48%24-layer ViT (415.35M) + 36-layer LM (4,022.47M, d=2560d=2560, tied language head)
Separate Action Transformer00.00%None (vs. 14.3%–51.5% in π0\pi_0, π0.5\pi_{0.5}, X-VLA, GR00T-N1.7)
Action/State Bridge & Risk Modules23.26M (10.15M active)0.52% (0.23% active)Point-wise 4-layer state/action MLPs (9.49M) + 2-layer token-risk head (0.66M)
Total Parameters4,461.08M100.00%Unified BARD-VL-4B parameter count across all evaluated checkpoints

Following Eq. (5) of the paper, we also evaluated central finite differences (ε=0.02\varepsilon = 0.02 in exact float32, averaged over 11 deterministic (t0,d0)(t_0, d_0) probes) across all four BARD-VL-4B checkpoints under both minimal (lang + state) and real dataset observations:

BARD-VL-4B CheckpointContext & Mask (HH, DvalidD_{\text{valid}})Point-Wise Bridge NLRBackbone Layer NLR≥1\text{NLR}_{\ge 1} (L0 -> L9 -> L18 -> L27 -> L35)Output NLR≥1\text{NLR}_{\ge 1}Output NLR≥3\text{NLR}_{\ge 3}Output NLR≥6\text{NLR}_{\ge 6}
PretrainMinimal, Bidirectional (20, 20D)0.00%2.53 -> 3.58 -> 15.46 -> 29.11 -> 34.79%36.12%10.06%4.23%
PretrainMatched, Bidirectional (50, 20D)0.00%1.79 -> 2.32 -> 9.95 -> 24.65 -> 28.55%30.68%9.13%7.13%
LIBEROReal Obs, Bidirectional (50, 10D)0.00%0.34 -> 0.87 -> 9.00 -> 23.46 -> 29.99%23.52%2.55%0.84%
LIBEROReal Obs, Block-Causal K=10K=10 (50, 10D)0.00%0.37 -> 0.96 -> 9.43 -> 24.77 -> 30.34%22.49%2.02%0.63%
CALVINReal Obs, Block-Causal K=10K=10 (25, 10D)0.00%2.66 -> 3.01 -> 11.77 -> 28.12 -> 44.74%37.44%5.04%1.75%
RoboTwinMinimal, Block-Causal K=10K=10 (50, 20D)0.00%2.29 -> 2.66 -> 10.97 -> 29.37 -> 36.36%34.79%4.18%1.34%

These measurements confirm that our architectural and mechanistic claims hold directly on the evaluated BARD-VL-4B models. Across all 19 float32 configurations (and step-size sweeps ε∈{0.01,0.02,0.04}\varepsilon \in \{0.01, 0.02, 0.04\}, where NLR≥1\text{NLR}_{\ge 1} and NLR≥6\text{NLR}_{\ge 6} vary by < 0.01 percentage points with zero repeat-forward jitter), the point-wise 4-layer MLP bridge exhibits strictly 0.00% cross-timestep coupling, whereas a single pass through the shared VLM backbone routes 20.24%–38.05% (NLR≥1\text{NLR}_{\ge 1}) and 0.51%–7.13% (NLR≥6\text{NLR}_{\ge 6}, matching the 1.44%–7.14% scale reported for our preliminary 8B models) of perturbation energy across distinct action timesteps, growing progressively from shallow (0.34%–4.91% at Layer 0) to deep shared layers (21.09%–46.96% at Layers 27–35).

Thank you for the detailed rebuttal and the revised manuscript; the matched-budget comparison against direct regression and flow matching and the BARD-VL-4B parameter and NLR measurements have fully addressed my concerns, so I will maintain my original score.

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

MONAD uses a shared diffusion vision-language backbone to predict continuous robot actions without a separate action expert. Each action vector is represented by one token. During training, selected action vectors are masked or corrupted, and the backbone learns to reconstruct them from images, language instructions, robot states, and visible actions. Experiments on LIBERO, LIBERO-Plus, and CALVIN show competitive performance.

Critical strengths
  1. The design is simple and promising. MONAD generates continuous actions through the shared backbone without a separate action Transformer or a separately trained action encoder–decoder. This offers a compact alternative to common VLA designs.
  2. Competitive results across several benchmarks. The method is evaluated on LIBERO, LIBERO-Plus, and CALVIN and compared with a wide range of baselines. The results are competitive overall, particularly on LIBERO-Plus.
Critical weaknesses
  1. The benefit of sharing the backbone is still unclear. The paper motivates its design by discussing the costs of separate action experts, but does not directly compare the two designs under the same backbone, data, and training budget. Differences in training data and recipes could explain some of the performance gains. Latency, compute, and memory usage are also not measured, so fewer action-specific parameters do not establish better practical efficiency.
  2. Some key design choices are not fully validated. The paper emphasizes reusing a pretrained language backbone, but provides little direct analysis of how language guides action selection. Testing different instruction goals in the same scene would help clarify this point. The evidence for the decoding choices is also limited: the comparison covers only 20 episodes, with four-round prediction reaching 100% both with and without refinement. This leaves the added benefit of refinement unclear. Risk-based selection is not separately tested, and the runtime cost of extra rounds is not measured. Further checks on the final model would help explain these choices and their trade-offs.
Questions for authors
  1. Beyond reducing action-specific parameters, what practical benefits does sharing the backbone offer? Any available results on inference speed or memory use would help clarify this point.
  2. Does MONAD retain the backbone’s language reasoning abilities after action training? Does it also support explaining its actions in language? Any existing tests or examples would help clarify the model’s language capabilities.
  3. Minor question about typos: Several words in the Conclusion (Section 5) contain unusual internal spaces, such as "vec tor", "wit hin", "bl ock" and "op en". What causes these gaps, and could the authors fix them and check for similar issues elsewhere?
4 comments

[W1, Q1] Controlled comparison of shared backbone vs. separate action expert under identical backbone, data, and budget; practical inference speed and memory usage.

Thank you for pointing out the need for a strictly controlled comparison. We evaluated Shared Backbone 12L (MONAD) and Separate 12L Action Expert on the 10 long-horizon multi-stage tasks of LIBERO-10 under the exact same BARD-VL-4B Pretrain backbone, identical LIBERO-10 training stream, identical trainable parameter count (1.23B [1,234,432,533] trainable parameters), and identical 2,000-step budget (effective_batch_size = 32, 64,000 trajectory chunks, FP32 master-weight AdamW, lr = 2.5e-5), testing both held-out offline accuracy and 100 paired closed-loop rollouts (10 tasks ×\times 10 paired seeds). Shared Backbone adapts the top 12 shared VLM layers (layers 24–35) plus the lightweight 4-layer point-wise I/O bridge and token-risk head, whereas Separate Action Expert freezes all 36 VLM layers to extract multimodal prefix features and attaches a newly initialized 12-layer action Transformer of identical width (d=2560d = 2560), attention heads, and FFN dimensions, trained with the same masked-reconstruction objective and evaluated with D4/R2.

Architecture (1.23B Trainable, 64k Chunks on LIBERO-10)Total Params (Non-VLM)Peak Train VRAMAction MAE ↓\downarrowGripper Acc ↑\uparrow50% Trajectory Completion MAE ↓\downarrowLIBERO-10 Closed-Loop Success (100 Paired Eps) ↑\uparrow
Shared Backbone 12L (MONAD)4.46B (0.5%)31.63 GiB0.33992.8%0.055 (98.1% grip)91.0% (91/100) (D1/R0: 89.0%)
Separate 12L Action Expert5.67B (21.8%)33.91 GiB0.45655.9%0.416 (53.8% grip)0.0% (0/100)

Sharing the backbone delivers decisive practical gains in parameter/memory footprint, sample efficiency, and closed-loop task completion. At matched trainable capacity (1.23B), attaching a separate 12-layer action expert inflates total resident parameters by +1.21B (+27.2%) and raises gradient-checkpointed peak training memory (batch_size = 32) by 2.28 GiB. More importantly, adapting the upper shared VLM layers directly inherits pretrained multimodal and cross-timestep representations, reaching 91.0% (91/100) paired closed-loop success on long-horizon LIBERO-10 within 2,000 steps (89.0% even in a single D1/R0 pass, paired 95% bootstrap CI [+85.0%, +96.0%] over Separate Action Expert), alongside +36.9 percentage points higher gripper accuracy (92.8% vs. 55.9%) and 7.6× lower conditional 50% trajectory completion MAE (0.055 vs. 0.416). By contrast, a newly attached 12-layer expert must learn cross-timestep action attention and contact transitions from scratch, failing to complete long-horizon multi-stage rollouts (0/100) under the same training budget. Consistent gains also hold when starting directly from the general vision-language checkpoint (BARD-VL-4B-Instruct, with zero prior action data) under matched 1.23B trainable parameters, where Shared Backbone reduces block-causal MAE by 35.1% (0.316 vs. 0.487) and conditional 50% completion MAE by 6.5× (0.059 vs. 0.385).

For inference speed and memory (full cross-model table in our response to Reviewer 5HNJ [W2, Q2]), profiling on an uncontended NVIDIA L20X GPU (bfloat16, batch_size = 1) confirms that zero-copy read-only prefix KV caching (21.5 MiB) and short-block decoding (K=10K=10 tokens/read) keep MONAD's peak inference VRAM completely flat at 8.44 GiB across 1 to 7 reads—below both π0.5\pi_{0.5} (8.95 GiB) and π0\pi_0 (8.46 GiB). Each K=10K=10 read adds only 25.1 ms and 73.8 GFLOPs, allowing MONAD to run in 68.0 ms (D1/R0), 94.6 ms (D4/R0), and 148.4 ms (D4/R2)—all faster and 2.7×–3.5× less compute-intensive than default 10-step π0.5\pi_{0.5} (184.8 ms, 5,066 GFLOPs) and π0\pi_0 (195.5 ms, 4,420 GFLOPs). Finally, keeping vision, language, and actions in one shared backbone allows MONAD to retain natural-language task planning and action explanations within the same checkpoint (detailed in [Q2] below). In the revised manuscript, we have incorporated this parameter- and budget-matched shared-backbone vs. separate-expert benchmark and the L20X inference speed and memory breakdown into Section 4.5 (Tables 2 and 3).

[W2] Validating same-scene instruction sensitivity, larger-scale closed-loop decoding refinement, isolated risk-based selection, and marginal runtime cost of extra rounds.

Thank you for these constructive suggestions. We conducted targeted evaluations on our BARD-VL-4B LIBERO policy to examine language conditioning, multi-round decoding, risk-based selection, and per-round runtime cost.

First, to directly test how language guides action selection among multiple valid goals within the exact same visual scene, we evaluated MONAD on LIBERO-Goal, where all 10 tasks share the identical tabletop layout and objects (wine bottle, cream cheese, bowl, plate, stove, wooden cabinet, and wine rack) and differ solely in the natural-language goal instruction. Holding the dual-view camera images and robot state 100% fixed across 50 initial LIBERO-Goal scenes and sweeping all 10×9=45010 \times 9 = 450 same-scene counterfactual goal switches, replacing the matched instruction (Block-0 MAE = 0.127) with an alternate tabletop goal induces an immediate 0.952 L1 trajectory shift across continuous action dimensions (0.851 on end-effector xyz, raising MAE relative to the original goal trajectory to 0.975, vs. 0.508 for an empty instruction ""). Crucially, in 94.0% (423/450) of same-scene goal switches, the newly generated trajectory moves directly toward the empirical ground-truth trajectory prototype of the newly instructed goal (halving prototype distance from 1.282 to 0.585). Paired closed-loop rollouts across all 10 LIBERO-Goal tasks (150 episodes total) confirm the same behavior in physical execution: matched instructions achieve 86.0% (43/50) closed-loop success, whereas empty instructions drop to 8.0% (4/50) and counterfactual same-scene instructions achieve 0.0% (0/50) on the original goal as the policy redirects the arm toward the newly specified tabletop target instead.

Second, to overcome the 20-episode ceiling in the original Appendix Table 6 and jointly measure refinement gains, risk-guided selection, and runtime cost, we evaluated all 10 long-horizon tasks of LIBERO-10 with 10 paired seeds per task (100 episodes per schedule, 500 closed-loop episodes total) using a single multi-suite unified LIBERO checkpoint so tasks remain below saturation:

Decoding Schedule (LIBERO-10, 100 Paired Eps/Arm)Remasking & Selection RuleBackbone ReadsL20X Chunk LatencyMarginal Read CostClosed-Loop Success
D4/R2 (Uniform Random Control)Uniform random selection (no risk guidance)4~148 ms+25.1 ms/read85.0% (85/100)
D1/R0Single-pass parallel draft168.0 ms(39.4 ms prefill)87.0% (87/100)
D4/R0Risk-ordered draft commitment, no remasking294.6 ms+26.6 ms88.0% (88/100)
D4/R2 (Learned Token-Risk)Lowest-confidence remasking + risk-gated accept4148.4 ms+26.9 ms/round90.0% (90/100)
D4/R4 (Learned Token-Risk)Up to 4 risk-gated remasking rounds6~201 ms+26.7 ms/round92.0% (92/100)

Closed-loop success rises steadily with refinement budget from 87.0% (D1/R0) →\to 88.0% (D4/R0) →\to 90.0% (D4/R2) →\to 92.0% (D4/R4), while each additional K=10K=10 backbone read adds a predictable 25.1 ms P50 latency (73.8 GFLOPs) and 0 GiB additional VRAM. At the exact same D4/R2 compute budget, replacing MONAD's learned token-risk signal with uniform random remasking lowers closed-loop success by 5.0 percentage points (90.0% -> 85.0%), below even unrefined D4/R0 (88.0%). We also verified slot selection strictly in isolation across 64 windows by injecting a +0.45 impulse corruption into 3 of 10 ground-truth Block-0 slots while holding all other components identical: uniform random selection slightly worsens full-block MAE (0.135 -> 0.142) by overwriting clean slots, whereas MONAD's learned token-risk head raises corrupted-slot Hit@3 from 28.6% to 41.7% and reduces 2-pass full-block MAE by 33.3% (0.135 -> 0.090; corrupted-slot MAE 0.450 -> 0.168). In the revised manuscript, we have added the same-scene LIBERO-Goal multi-goal counterfactual analysis and the isolated token-risk slot-selection results to Section 4.5, and updated Appendix B.4 (Table 8, replacing the former 20-episode Table 6) with this 500-episode LIBERO-10 closed-loop decoding and runtime trade-off study.

[Q2] Whether MONAD retains language capabilities after action training and supports explaining its actions in natural language.

Thank you for this insightful question. When fine-tuned exclusively on continuous actions (0% text supervision), the shared top layers specialize toward continuous trajectory decoding without cross-entropy anchoring to the tied language vocabulary head. Because MONAD preserves the entire 36-layer VLM backbone and vocabulary projection intact without architectural modification, however, lightweight action–language co-training—jointly optimizing continuous action batches alongside multimodal visual planning and action-explanation pairs distilled from the base VLM (language loss weight = 0.25) for 350 steps on our native K=10 LIBERO policy, with no post-hoc weight interpolation—cleanly preserves both capabilities within a single unified checkpoint.

To test genuine generalization rather than template memorization, we strictly excluded all 20 evaluation episodes from the language training pool and completely held out 5 LIBERO task instructions from any language supervision, evaluating the exact same checkpoint on both 200 paired closed-loop episodes across all 4 LIBERO suites (50 episodes per suite) and greedy natural-language generation:

Evaluation Metric (Evaluated on the Exact Same Unified Checkpoint)Action-Only MONADAction–Language Co-Trained MONAD
4-Suite Closed-Loop Control (200 paired episodes)93.0% (186/200)91.5% (183/200) (Spatial: 100.0%, Object: 90.0% vs. 80.0%)
Step-by-Step Action Explanation (20 held-out scenes: Non-Rep / KW Recall / Token-F1)0.0% / 0.0% / 0.00280.0% (16/20) / 83.6% / 0.566
└─ On 5 Completely Unseen Task Instructions (Non-Rep / KW Recall / Token-F1)0.0% / 0.0% / 0.000100.0% (5/5) / 82.4% / 0.538
Scene Grounding & Task Plan (20 held-out scenes: Non-Rep / KW Recall / Token-F1)0.0% / 0.0% / 0.00280.0% (16/20) / 73.5% / 0.574

Co-training retains virtually identical 4-suite closed-loop manipulation (91.5% vs. 93.0%, a difference of only 3 episodes across 200 rollouts, achieving 100.0% [50/50] on LIBERO-Spatial and improving object-centric manipulation on LIBERO-Object by +10.0 percentage points from 80.0% to 90.0%) while simultaneously generating coherent visually grounded task plans and step-by-step action explanations—including 100.0% (5/5) fluent non-repetitive explanations (82.4% keyword recall) on completely unseen task instructions. Below are two verbatim greedy outputs from this single unified policy:

  • Unseen Task Instruction — Action Explanation (Instruction: "pick up the ketchup and place it in the basket"):

    "Step 1: The robot arm will first locate the ketchup bottle on the table. Step 2: The robot arm will then move towards the ketchup bottle and pick it up. Step 3: The robot arm will then move towards the basket and place the ketchup bottle inside it."

  • Held-Out Long-Horizon Scene — Visual Grounding & Multi-Step Plan (Instruction: "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate"):

    "The objects on the table include a plate, a bowl, a box of cereal, and a small container. The robot should first open the top drawer of the wooden cabinet, then pick up the black bowl from inside the drawer, and finally place it on the plate on the table."

In the revised manuscript, we have included this unified action–language co-training evaluation and verbatim qualitative examples in Section 3.3, Section 4.5, and Appendix B.10 (Table 16), and we will further expand the discussion of top-layer representation specialization in the final version.


[Q3] Cause and fix for unusual internal character gaps ("vec tor", "wit hin", "bl ock", "op en") in Section 5.

Thank you for catching this! We inspected our LaTeX manuscript source for Section 5 and confirmed that the underlying phrases ("action vector", "within-family", "action block", and "The open questions") contain no stray spaces or macros; the visual gaps arise from pdfTeX font-expansion kerning interactions between the times and microtype packages on tightly justified lines. In the revised manuscript, we have adjusted our microtype font-expansion settings (\microtypesetup{expansion=false}) to eliminate these internal character gaps in Section 5 and verified the compiled PDF via automated text extraction and visual inspection; we will also perform a full-paper typographic check in the final version.

Thank you for your detailed response, which address my main concerns. I will therefore raise soundness from 2 to 3, and raise my rating from 2 (weak reject) to 3 (weak accept).

Reviewer_5HNJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes MONAD, a VLA framework that enables continuous robotic control using a single shared diffusion vision-language backbone without introducing a separate action expert, action tokenizer, or additional generative objective. The key idea is to represent each continuous action vector as a latent token and leverage the pre-trained diffusion backbone's masked-token reconstruction mechanism to generate continuous action sequences. The experiments demonstrate competitive performance on several VLA benchmarks.

Critical strengths
  • The proposed new architecture that unifies vision-language understanding and continuous action generation within a single diffusion backbone has some novelty. Although it combines techniques from others, it proposes new changes on top of them.
  • The experimental results across multiple benchmarks and VLA baselines are competitive and promising.
Critical weaknesses
  • It is highly recommended that the authors compare their method with the GR00T series.
  • It is highly recommended that the authors report the compute and memory costs of their model and compare it with others. For example, what are the actual inference latency, and GPU memory requirements of MONAD compared with representative VLAs such as π and GR00T? Although MONAD uses fewer action-specific parameters, does repeatedly executing the full diffusion backbone increase the computational and memory cost of action generation?
Questions for authors
  • Could the authors add the results of the GR00T series?
  • Could the authors report the compute and memory costs of their model and compare it with others?
1 comment

[W1, Q1] Comparison with the NVIDIA GR00T series.

Thank you for this suggestion. Below we compare MONAD with NVIDIA's flagship GR00T-N1.7-3B model alongside the representative VLA baselines from our paper on both architectural parameter allocation and standard LIBERO closed-loop success rates (using NVIDIA's official Isaac-GR00T LIBERO benchmark report for GR00T-N1.7-3B).

ModelTotal ParamsShared VLM BackboneNon-VLM Action ModuleSpatialObjectGoalLong (10)Average
π0\pi_04.03B3.45B (85.7%)0.58B (14.3%)96.898.895.885.294.2
π0.5\pi_{0.5}4.14B3.45B (83.3%)0.69B (16.7%)97.299.097.896.097.5
X-VLA0.88B0.57B (65.0%)0.31B (35.0%)98.298.697.897.698.1
GR00T-N1.7-3B3.14B1.52B (48.5%)1.62B (51.5%)97.798.597.594.497.0
MONAD (Ours)4.46B4.44B (99.5%)0.02B (0.5%)99.0100.098.894.898.15

Architecturally, the GR00T series truncates its VLM backbone at an intermediate layer (Cosmos-Reason2-2B truncated to 16 layers in GR00T-N1.7-3B; Eagle-2.5 truncated to 12 layers in GR00T-N1.5-3B) and delegates trajectory generation to a large separate Diffusion Transformer and embodiment-specific adapters—accounting for 51.5% (1.62B parameters, including a 32-layer DiT) of GR00T-N1.7-3B and 44.3% (1.07B of 2.41B) of GR00T-N1.5-3B. By contrast, MONAD retains 99.5% of its parameters inside the full 36-layer shared VLM backbone with zero separate action Transformer layers and only 0.5% (23.3M total; 10.15M [0.23%] active) lightweight state/action projections and risk head. Despite eliminating the dedicated action expert, MONAD outperforms GR00T-N1.7-3B across all four LIBERO suites (98.15% vs. 97.0% average) and reaches 85.9% across the 10,030 perturbed episodes of LIBERO-Plus. In the revised manuscript, we have added GR00T-N1.7-3B to Table 1 (Section 4.2) and contrasted GR00T's truncated-VLM plus separate DiT design against MONAD's unified shared-backbone formulation in Sections 1, 2, and Appendix B.9 (Table 14).

Comments on the paper

1 comment

[W2, Q2] Inference latency, GPU memory, and compute costs compared with π\pi and GR00T; impact of repeatedly running the full shared backbone.

Thank you for raising this practical question. We profiled MONAD (BARD-VL-4B), π0\pi_0, π0.5\pi_{0.5}, and GR00T-N1.5-3B head-to-head on the same uncontended NVIDIA L20X GPU (bfloat16, batch_size = 1, CUDA-event synchronized timing after warmup).

Model & ScheduleTotal ParamsPeak VRAM (GiB)Prefix KV (MiB)Prefix / Step TokensDenoising Reads / StepsChunk Latency P50 (ms)Total Compute (GFLOPs)
4B-Scale VLAs
MONAD (D1/R0)4.46B8.4421.5153 / 101 read68.01,444
MONAD (D4/R0)4.46B8.4421.5153 / 102 reads94.61,518
MONAD (D4/R2)4.46B8.4421.5153 / 104 reads148.41,666
MONAD (D4/R5)4.46B8.4421.5153 / 107 reads228.01,887
π0.5\pi_{0.5} (10-step default)4.14B8.9517.0968 / 5010 steps184.85,066
π0\pi_0 (10-step default)4.03B8.4614.3816 / 5110 steps195.54,420
2B-Scale Reference VLA
GR00T-N1.5-3B (4 / 10 steps)2.41B4.73—531 / 494 / 10 steps43.2 / 79.91,845 / 2,507

Repeatedly executing the 36-layer shared backbone does not inflate memory or compute because MONAD prefills the 153 multimodal prefix tokens only once per action chunk (39.4 ms, 1,371 GFLOPs, 21.5 MiB read-only KV cache). Every subsequent denoising pass attends to a zero-copy read-only view of that cached prefix while processing only K=10K = 10 action tokens (6.1% of sequence length) without appending to the cache. As a result, each additional backbone read costs just 73.8 GFLOPs (19.6× less compute than recomputing the full sequence) and 25.1 ms, while peak allocated GPU memory stays completely flat at 8.44 GiB from 1 to 7 reads—lower than π0.5\pi_{0.5} (8.95 GiB) and π0\pi_0 (8.46 GiB). Even at 7 reads (D4/R5), MONAD's total compute (1,887 GFLOPs) is 2.7× lower than 10-step π0.5\pi_{0.5} (5,066 GFLOPs) and lower than 10-step GR00T-N1.5-3B (2,507 GFLOPs; note that the newer 3.14B GR00T-N1.7-3B doubles the DiT depth from 16 to 32 layers).

In wall-clock latency, 4.46B MONAD runs substantially faster than comparable 4B VLAs—outpacing default 10-step π0.5\pi_{0.5} (184.8 ms) and π0\pi_0 (195.5 ms) across D1/R0 (68.0 ms), D4/R0 (94.6 ms, 2 reads with dynamic early commitment), and D4/R2 (148.4 ms), whereas GR00T-N1.5-3B's shorter per-step time simply reflects its smaller 2.41B 12-layer backbone. Moreover, MONAD already reaches top closed-loop accuracy with a single read: in a paired 50-episode LIBERO-Spatial evaluation, D1/R0 (68.0 ms) achieves 98.0% (49/50), matching 5-step π0.5\pi_{0.5} (98.0% at 122.7 ms) with a 1.8× latency speedup. In the revised manuscript, we have added this dedicated runtime efficiency table and analysis in Section 4.5 (Table 3), reporting peak GPU memory, prefix KV-cache footprint, P50 wall-clock latency, and GFLOPs across decoding schedules alongside π0\pi_0, π0.5\pi_{0.5}, and GR00T.


[Presentation] Improving notation clarity, structural organization, and typography.

We also appreciate your feedback on presentation clarity. In the revised manuscript, we have unified our inline Dp/Rq (K=10K=10, Hexec=10H_{\text{exec}}=10) schedule notation in Section 3.4, incorporated the GR00T comparison (Table 1) and L20X runtime efficiency benchmarks (Section 4.5, Table 3), updated the Abstract, Section 1, Section 4.6 (formerly Section 4.5), Section 5, and Appendix B.9 (Tables 14–15) so all parameter and temporal-response statistics reflect the evaluated BARD-VL-4B models, and adjusted our pdfTeX microtype settings to eliminate the character-spacing artifacts in Section 5. In the final version, we will further add a dedicated notation summary table and an Algorithm 1 pseudocode box in Section 3 and complete a full-paper camera-ready typographic polish.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.