{"id":"15c70a7b-c70d-45e0-9d2b-edb0025ad681","arxiv_id":"2607.19774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Moving multi-agent fusion from perception to planning, via an autoregressive decoder with MoE tokenization, yields 79.72 driving score on V2Xverse vs CoDriving's 77.15.","lead":"This paper proposes an end-to-end collaborative driving system that postpones multi-agent data fusion to the planning stage, using an autoregressive decoder to flexibly weight the ego car's and roadside sensors' information. On the V2Xverse closed-loop simulator it reports a 3.33% higher driving score than the CoDriving baseline, alongside lower infractions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation of the autoregressive decoder conflates fusion timing with planning-head capacity, so the core 'defer-to-plan' claim is not cleanly isolated.","rationale":"The reader identified the missing seeds/error-bar/baseline-reuse as the weakest assumption, and that is a valid measurement concern. However, the most load-bearing issue for the paper's stated central claim ('defer fusion to planning with adaptive weights') is the internal-validity confound in the ablation: the full model vs. the w/o AR Decoder variant differ in both the fusion stage and the planning-head architecture. The ablated variant appears to be both a fixed-fusion model and a different (weaker) trajectory decoder, so the 4.85% ADE gain cannot be uniquely attributed to the adaptive fusion. This is a concrete, addressable flaw in the experimental design that directly tests the causal mechanism. The reader's rationale also noted the fixed-fusion ablation is under-specified, so there is partial agreement. Since the correct remedy is an additional controlled experiment rather than a fundamental invalidation, the existing CONDITIONAL verdict remains appropriate; hence UNCHANGED. If the proposed control reveals no advantage for planning-stage fusion, the verdict would need to be reconsidered toward REJECT or at least re-scoped to a systems contribution rather than a design-principle contribution.","tokens_in":7960,"tokens_out":10863,"duration_ms":103110,"concrete_test":"Run a controlled ablation with three variants under identical training/evaluation settings (same seeds, same 67-route closed-loop protocol): (1) the full model (planning-stage adaptive fusion with AR decoder); (2) a perception-stage fixed-fusion variant that fuses ego and road tokens before the decoder with a learned fixed weighted sum, but uses the exact same autoregressive decoder and tokenizer; (3) the existing 'w/o AR Decoder' variant as in Table V. Compare DS, ADE, and IS across at least 5 seeds with confidence intervals. If variant (2) matches or exceeds variant (1), the improvement is not due to deferring fusion to planning; if variant (1) is substantially better than both (2) and (3), the deferral hypothesis is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that deferring multi-agent fusion to the planning stage, with adaptive weights learned by an autoregressive decoder, improves planning over fixed perception-stage fusion. The key evidence is Table V: 'w/o Autoregressive Decoder' yields ADE 0.628 vs. 0.598 for the full model, and the text attributes the largest gain to this component (4.85% ADE). However, this comparison changes two variables at once: the fusion mechanism (adaptive planning-stage attention vs. fixed perception-stage fusion) and the planning head (a LLaMA-style multi-layer transformer decoder vs. the presumably simpler head used in the fixed-fusion variant). The improvement credited to 'deferring fusion to planning' could instead be due to the higher capacity of the autoregressive decoder as a trajectory generator. The paper does not include the necessary control: the same autoregressive decoder fed with perception-stage fused features. Additionally, the w/o AR variant (0.628 ADE) is worse than the CoDriving baseline (0.619 ADE), suggesting that the fixed-fusion implementation used in the ablation may not be a strong baseline, further muddying the attribution. Without ruling out decoder-capacity effects, the manuscript's core design principle is not established, independent of the measurement-reproducibility concerns raised by the reader.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end V2X collaborative driving system that shifts multi-agent feature fusion from the perception stage to the planning stage. The architecture has dual-stream feature processing (ego and roadside), channel-wise purification of roadside features, MoE-enhanced tokenization, and a LLaMA-style autoregressive decoder that learns scenario-dependent fusion weights while generating waypoints. On the V2Xverse benchmark, the authors report a closed-loop driving score of 79.72 versus 77.15 for CoDriving, lower ADE/FDE, and better infraction score, with robustness to pose noise and communication latency.","tokens_in":8391,"tokens_out":3242,"duration_ms":33962,"significance":"If the central claim holds, the 'defer to plan' design principle would be a useful contribution: adaptive, planning-aware fusion at the trajectory-generation stage instead of fixed perception-stage fusion. The paper's strengths are its end-to-end formulation, explicit robustness experiments under noise and latency, and open-loop gains over several baselines. The work is also not circular: the main comparisons are against an external method on an external benchmark. However, the key evidence is currently undermined by a confounded ablation and an under-specified closed-loop evaluation protocol, so the significance of the claimed 3.33% DS improvement is not yet established.","major_comments":[{"comment":"The ablation 'w/o Autoregressive Decoder' changes both the fusion stage (planning-stage adaptive fusion vs. perception-stage fixed fusion) and the planning head (a multi-layer transformer decoder vs. a presumably simpler head). The 4.85% ADE gain credited to deferring fusion could instead be due to decoder capacity. The necessary control — the same autoregressive decoder fed with perception-stage fused features — is missing. Moreover, the 'w/o AR Decoder' variant (0.628 ADE) is worse than the CoDriving baseline (0.619 ADE), suggesting the fixed-fusion ablation implementation is not a strong baseline. Without this control, the paper's core design principle is not cleanly isolated.","section":"§4-D, Table V"},{"comment":"The closed-loop evaluation protocol is underspecified. No seeds, number of evaluation runs, variance, or confidence intervals are reported, and it is not stated whether the CoDriving baseline numbers (DS=77.15, RC=92.34, IS=0.82) were re-run under identical conditions or taken from [4]. The reported gap of 2.57 DS points (3.33%) could be within run-to-run variability of the simulator or due to protocol differences (e.g., route selection, infraction definitions). Please provide full protocol details and repeated-run statistics to establish the closed-loop improvement.","section":"§4-A, §4-B, Table II"},{"comment":"The ablation study reports only open-loop ADE/FDE, but the paper's headline result is the closed-loop driving score (DS/IS/RC). Without closed-loop ablations, there is no evidence that the proposed components (MoE tokenizers, channel-wise purification, autoregressive fusion) improve the closed-loop metrics that constitute the main claim. Please add closed-loop DS/IS/RC for each ablation configuration, or justify why open-loop errors are sufficient for the central claim.","section":"§4-D, Table V"},{"comment":"Key hyperparameters are hand-set without sensitivity analysis: α=0.1 (channel retention), λ0=0.03 (router guidance weight), expert count=6, Top-3 routing, K=6 (maneuver clusters). Because the method's components are justified by these choices, the reported gains may hinge on fragile parameter settings. Please provide sensitivity sweeps or a principled selection procedure for α, λ0, expert count, and K.","section":"§3-C2, §3-E, Table V"}],"minor_comments":[{"comment":"The abstract claims 'maintaining communication efficiency,' but no bandwidth or communication cost metric is reported; the latency table (Table I) reflects inference time, not message size. Please specify what communication efficiency is claimed and how it is measured.","section":"Abstract, §4-C"},{"comment":"The problem formulation introduces P^{j→i} and a bandwidth constraint ||P^{j→i}|| ≤ B, but the paper never maps this notation to actual message sizes or evaluates the constraint. Please connect the formulation to the experiments or remove the unmeasured constraint.","section":"Eq. (1)"},{"comment":"Several baselines have missing entries: TCP has no ADE/FDE, V2X-ViT has no closed-loop DS/IS/RC, Coopernaut has no closed-loop entries. For fair comparison, either fill these entries or state why they are unavailable.","section":"Table II"},{"comment":"The decoder is called 'LLaMA-Style Parallel Autoregressive Decoder,' but LLaMA is not typically described as parallel; the hybrid causal mask enables parallel decoding in this paper. Please clarify the terminology to avoid confusion.","section":"§3-D2"},{"comment":"The claim that Top-3 routing among 6 experts 'reduces inference cost while maintaining model capacity' depends on how the MoE layer is implemented (e.g., whether experts are sequentially activated or batched). Please specify the implementation to substantiate the efficiency statement.","section":"§3-C3"},{"comment":"The K-means maneuver clustering procedure is described only as 'clustering on trajectory curvature.' Please provide details: how is curvature computed, what trajectories are clustered, and how the K=6 clusters correspond to the MoE expert count.","section":"§3-C3, §4-E"},{"comment":"There are minor typographical issues, e.g., 'two-stage paradigm' (roman type), 'anend-to-end' in the introduction, and inconsistent spacing. Please proofread.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the proposed direction is timely. However, the central claim requires substantial strengthening of the evaluation and ablation methodology. I did not identify any indication of intentional misrepresentation; the issues are standard rigor concerns that should be fixable within a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the planning-stage adaptive fusion idea is real and worth taking seriously; the paper's own ablation doesn't cleanly prove it, and the closed-loop numbers come without error bars. Still, this is one of the more sensible V2X-AD papers I've read recently.\n\nWhat's actually new: deferring multi-agent fusion to a planning-stage autoregressive decoder that learns scenario-dependent attention weights, instead of fixed perception-stage fusion. That's a meaningful design change, and the hybrid causal mask (bidirectional context, causal planning queries) is a sensible way to parallelize generation. The dual-stream tokenization with channel-wise purification and MoE routing guided by trajectory-curvature K-means is also new to this setting, and the expert activation analysis suggests the routing actually specializes. The paper is well-written and the relation to CoDriving is clear.\n\nWhere it gets soft: the stress-test concern is legitimate. Table V's \"w/o Autoregressive Decoder\" removes the AR decoder entirely, which changes both the fusion mechanism and the capacity/architecture of the trajectory head. The improvement credited to 'adaptive fusion' could just be a bigger, better planner. The missing control is the same AR decoder fed with perception-stage fused features. Without it, the central claim about fusion timing isn't isolated. The fact that the w/o-AR variant (0.628 ADE) is worse than the CoDriving baseline (0.619) also suggests their fixed-fusion implementation is weak, which further muddies attribution.\n\nOn top of that, the reader's point stands: no seeds, no error bars, no rollout counts, and it's unclear if the CoDriving DS=77.15 was re-run or copied. A 3.33% closed-loop gap can easily be run-to-run noise in CARLA-based benchmarks. Hyperparameters like α, expert count, λ0 are hand-set with no sensitivity analysis, and ablation is only reported on open-loop ADE/FDE, not DS/IS.\n\nNone of this kills the paper. The core design idea is plausible and the open-loop gains are directionally consistent. But the evidence as presented supports 'the full system works better than CoDriving' more than it supports 'deferring fusion to planning is the reason.' The authors should add the decoder-capacity control, re-run or at least clearly state the provenance of all baseline numbers, and report variance.\n\nBottom line: worth a serious referee, but as is I'd want substantial revision. If you work on V2X or end-to-end driving, read it for the design idea; don't quote the DS numbers yet.","headline":"Planning-stage adaptive fusion is a real design idea, but the key ablation doesn't isolate it from decoder capacity, and the closed-loop numbers lack error bars.","tokens_in":8799,"tokens_out":2490,"would_cite":true,"duration_ms":23273,"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":"Deferring multi-agent fusion to the planning stage lifts V2X driving score by 3.33% over the prior best.","keywords":["V2X driving","collaborative driving","end-to-end planning","autoregressive decoder","multi-agent fusion","mixture-of-experts","closed-loop evaluation","trajectory prediction"],"falsifier":"Re-run the exact closed-loop benchmark protocol for both the proposed system and CoDriving with at least ten random seeds and report mean plus standard deviation; if the driving-score gap falls below the run-to-run variance, the claimed improvement is not established. Alternatively, freeze the decoder's fusion weights to fixed equal weights during evaluation: if the driving score stays essentially the same, the adaptive weighting is not doing the work.","tokens_in":7872,"feed_emoji":"🚗","tokens_out":3228,"duration_ms":27838,"temperature":0.7,"pith_summary":"The paper argues that in vehicle-to-everything (V2X) autonomous driving, fusing features from other vehicles and roadside sensors should happen at the planning stage, not the perception stage. The proposed system uses an autoregressive decoder to learn scenario-dependent weights for combining ego and road tokens, so the relative importance of roadside information can shift from an occluded intersection to an open highway. On a closed-loop benchmark, the method achieves a driving score of 79.72 versus 77.15 for the prior state-of-the-art, with fewer infractions and lower trajectory error. The author's core case is that fusion timing, not just fusion architecture, is the key design choice for end-to-end collaborative driving.","feed_headline":"Deferring multi-agent fusion to planning boosts V2X driving score","feed_subtitle":"Autoregressive planning-stage fusion beats perception-stage baselines in closed-loop V2X driving.","key_machinery":"The central mechanism is the autoregressive decoder with a hybrid causal mask. It concatenates a target embedding, ego tokens, road tokens, and planning queries into one sequence; context tokens attend to each other bidirectionally, while each planning waypoint attends to the full context and only preceding waypoints. This lets the decoder dynamically re-weight ego versus roadside information during trajectory generation, replacing fixed perception-stage fusion. Supporting machinery includes a dual-stream tokenizer with channel-wise top-10% purification of roadside features and a mixture-of-experts layer with trajectory-guided routing to prevent expert collapse.","core_discovery":"The central claim is that multi-agent feature fusion should be deferred to the planning stage, where an autoregressive decoder learns adaptive, scenario-dependent fusion weights, rather than being performed at the perception stage with fixed weights. The system combines channel-wise purification of roadside features, mixture-of-experts tokenization, and an autoregressive decoder with a hybrid causal mask to jointly reason over context tokens and generate waypoints. On the V2Xverse closed-loop benchmark, this design raises the driving score from 77.15 to 79.72 and the infraction score from 0.82 to 0.88, while also degrading less than the baseline under pose noise and communication latency.","pith_inferences":["A testable extension would be to replace the autoregressive decoder with a simpler per-scenario weighting function (for example, a small MLP conditioned on scene features) to isolate whether the gain comes from adaptivity or from the decoder's representational capacity.","The adaptive attention weights over ego versus road tokens could serve as an interpretability signal, letting engineers see which agent's information drove a braking or turning decision.","The same 'defer fusion to the planning stage' principle may generalize to single-vehicle sensor fusion, where camera and LiDAR features are still often fused early; planning-stage adaptive weighting could be tested there.","The paper's assumption of a fixed communication topology leaves open a dynamic version where the planner selects which agents to attend to per scenario, which could cut bandwidth further."],"forward_implications":["If the claim holds, perception-stage fusion methods in V2X driving are leaving measurable performance on the table, even when their perception metrics look strong.","The reported infraction-score gain (0.82 to 0.88) suggests that planning-stage adaptive fusion produces safer trajectories with fewer traffic violations.","Robustness results under pose noise and latency indicate that adaptive re-weighting can compensate for degraded or stale collaborative information, which matters for real-world deployment.","Ablation studies attribute the largest single improvement to the autoregressive decoder, implying that fusion timing itself is a primary source of gain, not just feature processing.","The measured planning latency (18.4 ms total, 10.2 ms for the decoder) suggests the approach is compatible with real-time control on a single GPU."],"fun_headline_variants":["Defer fusion to planning, V2X driving score jumps 3.3%","Plan-time fusion beats perception fusion in V2X driving","Adaptive autoregressive fusion lifts V2X driving to 79.7","Fuse at planning stage for better V2X autonomous driving","V2X driving: deferring fusion to planning boosts score"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim rests on the closed-loop comparison against the CoDriving baseline being conducted under identical evaluation conditions; the paper reports no seeds, rollout counts, or confidence intervals, so if the baseline numbers came from a different protocol or run, the 3.33% driving-score gap may not be real.","fun_headline_variants_meta":{"raw":{"variants":["Defer fusion to planning, V2X driving score jumps 3.3%","Plan-time fusion beats perception fusion in V2X driving","Adaptive autoregressive fusion lifts V2X driving to 79.7","Fuse at planning stage for better V2X autonomous driving","V2X driving: deferring fusion to planning boosts score"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1049,"prompt_tokens":669,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":413,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":413,"tokens_out":380,"duration_ms":4488,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:43:49.574741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact closed-loop benchmark protocol for both the proposed system and CoDriving with at least ten random seeds and report mean plus standard deviation; if the driving-score gap falls below the run-to-run variance, the claimed improvement is not established. Alternatively, freeze the decoder's fusion weights to fixed equal weights during evaluation: if the driving score stays essentially the same, the adaptive weighting is not doing the work.","supporting_citations":[],"review_version":1}