CLAMP: Constrained Decoding for Vision-Language Embodied Planning
Tianyi Ma, Parisa Kordjamshidi
cs.AI
Quick verdict
Proposes decode-time syntax/visibility/reachability constraints + HMM lookahead for frozen VLM planners; solid grounding gains, but metrics mixed and TTA weak.
Summary
The paper presents CLAMP, a constrained-decoding framework that injects scene evidence and a supplied symbolic action model into the decoding loop of a frozen vision-language model (VLM) planner. Three mechanisms modify next-token logits: hard DFA masks for action-schema syntax, hard masks restricting entity arguments to observation-supported identifiers, and a state-conditioned executability mask combined with an HMM-based lookahead score that ranks candidates by budget-aware goal reachability. Only the HMM's emission matrix is adapted at test time via label-free Baum-Welch updates (TTDA/OTTA/TTBA variants). On VLABench, the Qwen3-VL-8B macro score rises from 28.7 to 34.1 with hard masks and 37.1 with text-HMM guidance (38.7 with a supervised vision HMM); matched VirtualHome Action Sequencing success goes from 21.3% to 48.7% to 85.3%; SafeAgentBench long-horizon symbolic violations drop from 0.41 to 0.05 with a recovery mechanism. InternVL3.5 grounding gains concentrate in entity-ID validity (1.9 to 100). Key limitations: Table 1 mixes three incompatible score definitions across rows; test-time adaptation gains are confined to 78 Physical-Laws prompts via single-skill collapse (overall 9.6/100); token-only masks hurt BEHAVIOR Goal Interpretation and halve Subgoal-Decomposition success; visibility is static (initial frame only); semantic safety improvements are modest and LLM-judge dependent. The authors attribute most residual failures to perception errors and misaligned constraint specifications.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Observation-conditioned hard constraints (syntax and visibility DFAs) enforced during decoding improve frozen-VLM planning: VLABench Qwen3-VL-8B macro rises from 28.7 to 34.1.
- HMM-based budget-aware goal-reachability lookahead adds gains beyond hard masks: 34.1 to 37.1 with text HMM (38.7 with supervised vision HMM), and matched VirtualHome Action Sequencing success from 48.7% to 85.3%.
- Emission-only test-time HMM adaptation using label-free continuations preserves constraints while recalibrating guidance, raising Physical Laws from 6.4 to 59.0.
- Safety-policy decoding with recovery cuts SafeAgentBench long-horizon symbolic-rule violations from 0.41 to 0.05 without updating the planner.
- Most remaining failures stem from perception errors or misaligned constraint specifications rather than the decoding machinery.
Strengths
Precise, mechanistic formalization
Eq. 1 composes syntax/visibility/executability masks and the lambda-weighted HMM reachability score; Eq. 4-5 define the budget-aware backward DP; Algorithm 1 (A1.4) gives the full per-token pipeline with the failure gate and state updates.
Unusually honest evaluation bookkeeping
A2.5 separates the official VLABench weighted score, the six-dimension macro, and local diagnostic totals; Table 6 discloses which constraint mechanisms were actually evaluated per benchmark; A3.10 audits its own LLM self-judge (byte-identical plans flipping 3/3 True vs 0/3 False under harness metadata); A1.5 shows a Llama-to-Qwen3 emission remap is a prerequisite (480/480 empty outputs without it).
Grounding gains are large, mechanistically coherent, and CI-quantified
Safety result isolates the mechanism
Measured cost reporting
Weaknesses
Three incompatible score definitions share Table 1
Test-time adaptation is far weaker than the abstract implies
Constrained decoding actively hurts under interface misalignment
Thin uncertainty on headline numbers
Static visibility and specification dependence
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CLAMP (arXiv 2609.08602) is a method paper at the intersection of constrained decoding and embodied planning. It claims that decoding-time constraints — a syntax DFA compiled from the task action schema, a visibility mask restricting entity arguments to observation-supported identifiers (Eseen), and a reachability mask plus HMM lookahead over a supplied symbolic transition model — improve frozen-VLM planning without planner updates. Headline numbers: VLABench Qwen3-VL-8B macro 28.7 to 34.1 (hard masks), 37.1 (text HMM), 38.7 (supervised vision HMM) (Table 1); matched EAI VirtualHome Action Sequencing 21.3% to 48.7% to 85.3% (Llama-3.1-8B, Sec 5.1); SafeAgentBench long-horizon symbolic violations 0.41 to 0.05 (Table 2). A test-time adaptation tier updates only HMM emissions from label-free continuations.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CLAMP in one sentence?
CLAMP is a decoding-time framework that constrains a frozen vision-language model's embodied plans with three constraints — action-schema syntax, scene visibility of referenced entities, and state/goal reachability — the last guided by an HMM-based lookahead over a supplied symbolic action model.
How much does CLAMP improve performance?
On VLABench with Qwen3-VL-8B, the macro score rises from 28.7 (unconstrained) to 34.1 with hard masks and 37.1 with text-HMM guidance; matched VirtualHome Action Sequencing success rises from 21.3% to 48.7% to 85.3%; SafeAgentBench long-horizon symbolic violations fall from 0.41 to 0.05 with a recovery mechanism.
Does CLAMP fine-tune the underlying VLM?
No. The planner stays frozen; at test time only the lightweight HMM's emission matrix is updated from label-free continuations sampled from the same frozen model, via a Baum-Welch step.
What are the main limitations of CLAMP?
Its guarantees are conditional on the supplied symbolic action model and scene-support list; visibility uses only the initial frame (occluded objects are rejected); test-time adaptation gains are confined to one benchmark dimension; and token-only masking can hurt when the constraint interface is misaligned with a benchmark, as on BEHAVIOR Goal Interpretation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with several [...sampling gap...] regions (mid-paper experiments and portions of appendices A1.7, A3.4, A4.1); every number cited in this review was verified in the visible text, but content inside the gaps was not seen.
- Figures are available only as extracted values and captions (e.g., Figure 4 bar values); visual-trend assessment is limited to those embedded numbers.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.