{"id":"578b4cc5-3b2e-4a63-b911-70bc2292943f","arxiv_id":"2607.14688","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Intent-sharing with edge-based arbitration improves simulated highway maneuver efficiency and safety over isolated, FCFS, and MARL baselines.","lead":"This paper proposes a system where connected self-driving cars share their planned maneuvers with roadside edge computers, which decide which car goes first and log the reasoning. A generalist might read it because it aims to make multi-car coordination safer, more efficient, and more auditable than cars acting alone.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"S2/S3 completion-time advantages fall inside 95% CI overlap; without significance tests, the outperformance claim is not yet established.","rationale":"The reader's weakest assumption—that 10 stochastic runs provide a meaningful comparison—captures the same root issue as my concern: the reported confidence intervals overlap substantially in S2 and S3, yet no significance test is reported. If the differences are not statistically significant, the paper's headline efficiency advantage is not supported. I considered the 50 km/h speed cap as an alternative concern, but even if it applies only to MIND-CAVs, it would disadvantage the proposed system, not artificially favor it, so it is not the most load-bearing issue for the comparative claim. The under-specified MARL baseline is also a concern, but it affects only one of the three baselines, whereas the statistical issue affects the entire efficiency comparison. The reader's CONDITIONAL verdict remains appropriate: the architecture and S1 results are promising, but the missing significance testing and run-level data must be addressed before the efficiency claim can be independently confirmed.","tokens_in":11729,"tokens_out":5695,"duration_ms":54782,"concrete_test":"Extract the per-run completion times for S2 and S3 (the paper should release these logs) and compute a non-parametric two-sided test (e.g., Mann-Whitney U or bootstrap 95% CI for the difference) comparing MIND-CAVs to IA, FCFS, and MARL. If the p-value exceeds 0.05 or the difference CI includes zero for S2/S3, the efficiency claim must be restricted to S1. If the data cannot be released, re-run with a pre-registered n=30 and report the same tests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim asserts MIND-CAVs achieves the lowest maneuver completion time across S1–S3. In S1 the advantage is clear (29.10±1.17 vs 38.57±2.70 for IA). But in S2 the reported means are 39.46±4.84 vs 44.69±5.25; the 95% CIs are [36.00,42.93] and [40.93,48.44], which overlap between 40.93 and 42.93. In S3 the means are 39.56±2.49 vs 40.73±4.71; CIs [37.78,41.34] vs [37.36,44.10] overlap almost entirely. The paper never reports a significance test or effect size. With n=10 and independent stochastic VLM sampling, these differences could easily arise from sampling noise. The central 'outperforms' claim therefore rests on one scenario for completion time; the other two scenarios only show numerically lower means without evidence. This is load-bearing because the abstract and conclusions explicitly claim improved maneuver completion time as a headline result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MIND-CAVs, a hierarchical Vehicle-MEC-Cloud architecture in which vehicles generate structured intent–plan pairs, roadside MEC nodes arbitrate conflicting intents through a constrained VLM with deterministic safety validation, and cloud storage maintains audit records for transparency. The system is implemented in a CARLA-based AI-in-the-loop platform and compared against Isolated Autonomy, FCFS arbitration, and MARL in three two-vehicle highway scenarios. The central claims are improved maneuver completion time, reduced gap violations, reduced unnecessary braking, and sub-second bounded-latency negotiation.","tokens_in":11970,"tokens_out":5268,"duration_ms":55417,"significance":"If the claims held, the architecture would be a useful contribution to connected-vehicle coordination by making intent exchange, edge arbitration, and auditable decision records first-class design elements. The accountability-by-design logging and the AI-in-the-loop simulation platform are valuable engineering contributions. However, the empirical evidence as presented is not yet sufficient for the headline claims: statistical support is weak in two of three scenarios, the safety metric overlaps with a constraint the system enforces by construction, the latency claim is extrapolated rather than measured, and the baseline configuration for MARL is underspecified.","major_comments":[{"comment":"The claim that MIND-CAVs achieves the lowest completion time across S1–S3 is not supported by significance testing. In S3, MIND-CAVs (39.56±2.49 s, CI [37.78,41.34]) and IA (40.73±4.71 s, CI [37.36,44.10]) have heavily overlapping 95% CIs; a direct comparison suggests the difference is not significant (t≈0.7). In S2 the difference vs IA is more promising (5.2 s, approximate t≈2.3) but still not formally tested. The abstract and conclusions state 'improved maneuver completion time' without restricting this to S1. Report per-scenario significance tests or effect sizes with bootstrap intervals, and adjust the claims to match the evidence.","section":"§VII and Table I"},{"comment":"The gap-violation metric uses the same threshold d_safe=5 m that the MIND-CAVs deterministic validator enforces as a hard constraint on approved plans. Thus near-zero gap violations for MIND-CAVs are largely a design consequence, not an empirical discovery. This makes the 'reduced unsafe proximity' claim partially circular as a comparative result. Please evaluate safety with at least one metric not directly gated by the system (e.g., time-to-collision, maximum deceleration, or distance below a stricter threshold), or compare against baselines augmented with the same safety filter, or explicitly reframe the result as an architectural guarantee.","section":"§V-B.2 and §VI-D"},{"comment":"The 'sub-second negotiation' and 'bounded-latency' claims are estimates, not measured results. The end-to-end latency of 335–690 ms is obtained by summing on-device inference throughput (Jetson ~270 ms), assumed 5G transmission times (<10 ms each way), and A100 inference throughput (45–400 ms). The paper does not report any end-to-end measurement from the implemented simulation loop, nor does it report latency distribution or worst case. Since this is an advertised contribution, either measure the actual loop latency in the platform or explicitly label the numbers as an analytical model and remove the unqualified 'sub-second negotiation' wording.","section":"§V-C and Abstract"},{"comment":"The fixed 50 km/h upper speed bound is a limitation that affects external validity, and the paper's extrapolation to 'highway speeds (100 km/h)' in §V-C is not justified. Although all scenarios initialize at 40 km/h, the results are only for low-speed driving and the speed cap was introduced for API rate-limit reasons unrelated to the method. The abstract and conclusions imply highway-relevant conclusions; please either remove the 100 km/h extrapolation or run at least one scenario without the cap at realistic highway speeds.","section":"§VIII Limitations and §V-C"},{"comment":"The MARL baseline is described only as 'a multi-agent policy observes shared vehicle states and outputs maneuver decisions optimized to minimize delay and safety violations.' No algorithm, network architecture, training procedure, hyperparameters, or reward function is provided. Since the paper's central comparison is empirical, an underspecified baseline makes the results non-reproducible and risks an unrepresentative comparison. Provide full implementation details or cite a specific public implementation and configuration.","section":"§VI-B, MARL baseline"}],"minor_comments":[{"comment":"The abstract states reduced unsafe proximity and unnecessary braking 'compared with ... FCFS'. In Table I, S2 shows MIND-CAVs unnecessary braking 0.40±0.70 vs FCFS 0.30±0.48, so this comparison is not uniformly improved. Please qualify or correct the claim.","section":"Abstract and Table I"},{"comment":"The 'Optimization Objective 1' formalization is not numbered as an equation, and the footnote '2' appears in the text '∆max 2' rather than after 'bounded arbitration delay.' Clarify the notation and numbering.","section":"Section III"},{"comment":"The stated A100 inference throughput of 'approximately 22 tokens/s' seems unusually slow for the given hardware and model; if this is correct, specify the model size and quantization, and if it is a typo, correct it.","section":"Section V-C"},{"comment":"The right-hand panels use 'T otal' with a space; also the figures would benefit from error bars or box-plot overlays to show the per-run variability referenced in the text.","section":"Figures 3–5"},{"comment":"The title and some headings contain 'MIND-CA Vs' with an unwanted space before 'Vs'. Standardize the name as 'MIND-CAVs'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful architectural idea and a reproducible simulation scaffold (seeds, JSON logs), but the central comparative claims currently exceed the statistical and measurement support. The overlap between the enforced safety constraint and the safety metric is a structural concern that should be addressed by the authors before publication. The underspecified MARL baseline also warrants scrutiny."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know upfront. First, the novel part of this paper is the composition: vehicle-side VLM intent generation, MEC-side VLM arbitration with a deterministic safety validator, and an audit log that records intents, decisions, and rationales. The ingredients come from prior work, and the paper's own related work section says so honestly. Second, the headline claim—\"improved maneuver completion time\"—is only clearly supported in Scenario 1. In S2 and S3 the 95% confidence intervals for completion time overlap with the Isolated Autonomy baseline, and with n=10 and no significance test, those differences are not established. The safety metrics (gap violations, unnecessary braking) are more consistently better, especially in S3, so the paper is not empty—but the efficiency claim needs a more modest framing or stronger evidence.\n\nWhat the paper does well: the architecture is concrete, the protocol details are specified (ACK/PLAN/NACK, lexicographic priority, one-shot revision), and the design decision to put a deterministic spatial-temporal validator after the VLM is exactly the right way to make a stochastic semantic planner safe enough for a simulation study. The limitations section is unusually candid: it admits the fixed 50 km/h speed cap is there to avoid API rate limits, and it acknowledges there are no formal guarantees. That counts for something.\n\nSoft spots, in proportion. The speed cap means the evaluation is not at realistic highway speeds, which matters for a paper about highway maneuvers. The \"sub-second negotiation\" number is an estimate built from measured inference throughput and reported 5G latency, not an end-to-end measurement. That should be labeled as an estimate or measured. Reproducibility is a real gap: no code, no data, no prompts, and the MARL baseline is under-specified to the point where I can't tell what was trained or how. A reader cannot verify the comparison. The paper also doesn't study VLM failure modes, though the validator cushions that limitation.\n\nMy verdict: this is a legitimate systems paper that deserves a serious referee. Send it out, but the authors should be asked for statistical tests or effect sizes, end-to-end latency data, and at least the prompt templates and MARL configuration. The audit-log mechanism is the most citable idea; I'd want to see the comparison repeated independently before I cite the results.","headline":"The architecture and audit-trail idea are worth taking seriously, but the performance comparison only clearly holds in S1; the paper needs significance tests, measured latency, and artifacts before I'd trust the headline.","tokens_in":12514,"tokens_out":4410,"would_cite":false,"duration_ms":43761,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Intent-driven edge arbitration beats isolated, FCFS, and MARL driving in highway tests.","keywords":["connected autonomous vehicles","intent-driven autonomy","maneuver negotiation","V2X communication","multi-access edge computing","arbitration","vision-language model","CARLA simulation"],"falsifier":"Run S1–S3 at 100 km/h with no speed cap and compare completion time, gap violations, and braking against the three baselines; also instrument the full in-simulation negotiation loop end-to-end to check whether latency stays below the claimed ~700 ms bound.","tokens_in":11553,"feed_emoji":"🚗","tokens_out":3949,"duration_ms":39363,"temperature":0.7,"pith_summary":"This paper argues that connected autonomous vehicles become safer and more efficient when they exchange structured maneuver intents instead of only kinematic state, and when a roadside edge server arbitrates conflicting intents with a language-model-based decision agent gated by a deterministic safety validator. The authors describe a three-tier Vehicle-MEC-Cloud architecture in which vehicles send intent-plan pairs, receive ACK/PLAN/NACK decisions, and the cloud stores every arbitration for audit and replay. In CARLA highway simulations, this intent-driven arbitration completes maneuvers in less time than isolated autonomy, first-come-first-served arbitration, or multi-agent reinforcement learning, with fewer gap violations and less unnecessary braking. If the result holds beyond the paper's constrained testbed, it would point to a practical division of labor: vehicles handle perception and intent proposal, while infrastructure handles conflict resolution and accountability.","feed_headline":"Intent-driven edge arbitration beats three driving baselines","feed_subtitle":"Structured maneuver intents negotiated at roadside servers cut completion time by up to 29% and reduce unsafe braking.","key_machinery":"The load-bearing mechanism is the intent abstraction (goal, maneuver class, execution horizon, safety constraints) coupled with a constrained VLM arbitration loop: the MEC node merges incoming intents with active plans, applies lexicographic priority rules, generates at most one revised plan, and a deterministic spatiotemporal validator overrides any unsafe decision to NACK. This turns coordination into a gated, audit-trailed decision problem, with bounded latency achieved by single-pass reasoning and no multi-round negotiation.","core_discovery":"MIND-CAVs proposes that connected autonomous vehicles should exchange structured maneuver intents—goal, maneuver class, execution horizon, safety constraints—rather than only kinematic state, with roadside MEC servers arbitrating conflicting intents via a vision-language model under deterministic safety validation, returning ACK/PLAN/NACK, and logging every decision for audit. In CARLA highway scenarios this reduces maneuver completion time by 24–29% over the three baselines while cutting gap violations and unnecessary braking. The paper's central claim is that externalizing negotiation as an auditable, safety-gated protocol improves both efficiency and stability without sacrificing safety.","pith_inferences":["The audit-log pipeline could double as a dataset generator: large-scale logs of intents, revisions, and outcomes might be used to distill the VLM arbiter into a smaller onboard policy, sidestepping the API rate limits that forced the 50 km/h cap.","The same Vehicle-MEC-Cloud split could apply to unsignalized intersections and ramp merging, but the paper's two-vehicle highway scenarios do not yet demonstrate that extension.","If the speed cap is lifted and the arbitration advantage shrinks, the efficiency result may owe more to the cap than to intent reasoning; that is a testable boundary of the claim.","The architecture externalizes negotiation so authorities can inspect why a maneuver was approved; that could change how regulators audit autonomous fleets, but the paper does not address liability or certification."],"forward_implications":["Intent-aware arbitration can reduce maneuver completion time by about 24–29% relative to isolated autonomy, FCFS, and MARL in multi-lane highway scenarios (S1).","Gap violations and unnecessary braking approach zero in simple lane-change scenarios, indicating fewer unsafe proximity events and less oscillatory driving.","Every arbitration outcome is recorded as a structured audit record, enabling deterministic replay and regulatory inspection—an architectural primitive, not a post-hoc tool.","With at most one plan revision and a fixed priority rule, the negotiation loop can finish below 700 ms (about 20 m of travel at 100 km/h), making sub-second coordination plausible.","Because the deterministic safety validator overrides unsafe suggestions, intent generators need not be perfect; the safety gate is the final authority."],"fun_headline_variants":["Edge-arbitrated intent negotiation beats three CAV baselines","Intent-driven edge arbitration cuts completion time 24–29%","Negotiated intents at edge reduce unsafe braking in CAVs","Auditable edge intent arbitration improves CAV efficiency","Intent-sharing with edge arbitration outperforms isolated driving"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison is made under a fixed 50 km/h speed cap imposed to avoid external API rate limits, which the paper concedes does not reflect realistic highway speeds; if the efficiency gains vanish when the cap is removed, the central advantage is conditional, not general.","fun_headline_variants_meta":{"raw":{"variants":["Edge-arbitrated intent negotiation beats three CAV baselines","Intent-driven edge arbitration cuts completion time 24–29%","Negotiated intents at edge reduce unsafe braking in CAVs","Auditable edge intent arbitration improves CAV efficiency","Intent-sharing with edge arbitration outperforms isolated driving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001225,"raw_usage":{"total_tokens":4875,"prompt_tokens":751,"completion_tokens":4124,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":4056}},"tokens_in":495,"tokens_out":4124,"duration_ms":27783,"temperature":1.0,"reasoning_tokens":4056,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:18:34.365098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run S1–S3 at 100 km/h with no speed cap and compare completion time, gap violations, and braking against the three baselines; also instrument the full in-simulation negotiation loop end-to-end to check whether latency stays below the claimed ~700 ms bound.","supporting_citations":[],"review_version":1}