{"id":"2b6a0fb6-52cd-4bed-a569-e70a47cc387b","arxiv_id":"2512.15038","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"LADY shows that an end-to-end driving model using only linear attention can match transformer-based planners on NAVSIM/Bench2Drive while fusing arbitrary-length historical sensor frames at constant per-frame cost.","lead":"LADY is an end-to-end driving planner built entirely on linear attention (RWKV-7), adding a linear cross-attention block and a diffusion decoder, and it reports planner accuracy close to transformer-based state of the art on NAVSIM and Bench2Drive while keeping per-frame inference cost constant as history grows. Why read: it claims to remove the quadratic-attention bottleneck for real-time autonomous driving on edge devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LICA's causal RWKV-7 pass compresses all BEV-ego tokens into a fixed-size recurrent state before queries read them; no ablation shows this lossy cross-attention preserves the information needed for planning.","rationale":"Assessment: The paper is a coherent engineering contribution and I do not doubt the authors' good faith. The memory-complexity error in Table I (RWKV-7 state is d×d, so memory is O(d^2), not O(d)) is real and should be corrected, and the absence of code/error bars weakens reproducibility. But neither of those directly threatens the core claim 'constant inference cost regardless of history length' as strongly as the unvalidated LICA design. The reader identified LICA as the weakest assumption; I partially agree. The specific concern is not so much that 'feature tokens never see query tokens'—standard cross-attention only requires queries to attend to features—but that LICA makes each query read a fixed-size recurrent summary of all features rather than performing a direct query-to-feature lookup. Whether this compression is adequate for driving is an empirical question the paper does not isolate. The proposed ablation is feasible, since the rest of the model is unchanged, and would settle whether the fully-linear cross-attention is the source of the remaining gap to iPad (PDMS 90.9 vs 91.7) or is essentially free.","tokens_in":12962,"tokens_out":16706,"duration_ms":535010,"concrete_test":"On the NAVSIM navtest split, replace the second RWKV-7 block in LICA with standard softmax cross-attention: use the M encoded queries as Q and the fused BEV-ego tokens as K/V, keeping all other components, training epochs, batch size, and inference settings identical. Compare PDMS and the closed-loop success rate/driving score. If the gap is within run-to-run noise (or <~1 PDMS), the causal fixed-state LICA is not losing material information; if it exceeds that, the fully-linear design carries an unquantified accuracy cost that should be disclosed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on LICA (Sec. III-B, Fig. 4) being a sufficient linear replacement for cross-attention. In the design, fused BEV-ego tokens are followed by encoded query tokens in a single causal RWKV-7 pass, and only the last M outputs are taken. Although the query-side outputs can see a recurrent state that has summarized all feature tokens, that state is a fixed d×d matrix; it is not equivalent to letting each query attend directly to the full feature sequence, and it is not bidirectional. The paper's statement that the state 'naturally attend[s] to both earlier feature tokens and later query tokens' is misleading: the feature-token outputs are discarded, and query tokens never influence the feature representation. More importantly, no ablation compares LICA against standard cross-attention or a bidirectional linear variant, and no analysis shows that the fixed-size state preserves the spatial/safety-critical information needed for planning. If LICA's lossy compression discards information that transformer cross-attention retains, the reported competitive planning numbers understate the cost of the fully-linear design. This is load-bearing because the novelty and the efficiency/accuracy tradeoff of LADY both depend on LICA being 'effective cross-modal interaction' (Sec. III-B).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"LADY proposes a fully linear-attention end-to-end autonomous driving model that replaces transformer attention with RWKV-7 linear attention for multi-frame camera/LiDAR fusion, and introduces LICA, a causal RWKV-7-based linear cross-attention mechanism inside a diffusion-based trajectory decoder. The paper claims constant per-frame inference cost and memory independent of history length, reports a NAVSIM PDMS of 90.9, closed-loop Bench2Drive results showing the benefit of temporal context, and validates deployment on an NVIDIA Jetson edge device.","tokens_in":13298,"tokens_out":5267,"duration_ms":50889,"significance":"If the efficiency and accuracy claims hold, the contribution is useful: it extends modern linear-attention sequence models to cross-modal end-to-end planning and demonstrates a constant per-frame inference overhead as historical context grows. The paper provides direct benchmark measurements, a temporal-context ablation, and edge-device validation, which are concrete strengths. However, the headline complexity claim is overstated (the recurrent state is d×d, not O(d)), and the LICA design lacks an ablation against standard cross-attention, so the central claim that linear cross-attention is an effective replacement is not yet fully supported.","major_comments":[{"comment":"The memory complexity claim is misstated. Equation (1) defines S_t as a d×d matrix, so the recurrent temporal state requires O(d^2) memory per frame, not O(d) as listed in Table I. The abstract's 'O(1) regardless of the history length' is only true with respect to the history length T, not as an absolute constant. The time column O(Td) is a total cost over T frames, not a per-frame cost; the claimed constant per-frame cost should be stated as O(d^2) per step independent of T. Please correct the table and abstract accordingly.","section":"Table I, Eq. (1), Abstract"},{"comment":"LICA is a single causal pass: feature tokens are processed first, their outputs are discarded, and query tokens never affect the feature representation. The last M query outputs see feature information only through the fixed d×d recurrent state. The statement that the state 'naturally attend[s] to both earlier feature tokens and later query tokens' is misleading for the extracted query outputs. Since LICA is a core contribution supporting 'effective cross-modal interaction', the paper needs an ablation comparing LICA against standard transformer cross-attention (or a bidirectional linear attention variant) to quantify any accuracy loss. Without such an ablation, the fully-linear design's accuracy claim is not fully supported.","section":"Sec. III-B, Fig. 4"}],"minor_comments":[{"comment":"The text says 'the rows of S∈R^{B×d} are the hidden states S_0,...,S_{B-1}', but S_t is d×d. The dimensions in this sentence are inconsistent and should be clarified.","section":"Sec. III-A, Eq. (1), parallel form"},{"comment":"The frame-number row '15' is typeset as '1598.0', merging the frame count and the numeric score; please fix the formatting.","section":"Table III"},{"comment":"'start-of-the-art' should be 'state-of-the-art'; also, the footnote notation 'MT b' should be expanded and the table footnote numbering cleaned up.","section":"Table I caption"},{"comment":"The sentence on Mamba-2's decay, 'resulting them becoming a smaller proportion of the numerically increasing state', is grammatically unclear and should be rephrased.","section":"Sec. II-B"}],"recommendation":"major_revision","confidential_remarks":"The two major issues are fixable within the manuscript's scope: correct the complexity statements and add an ablation that directly tests whether LICA preserves the information needed for planning. I did not find evidence of data fabrication; the benchmark numbers are internally plausible. The 'first fully linear attention' novelty claim may warrant a brief literature check by the editor, since the paper itself surveys closely related linear-attention architectures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is best read as an engineering result, not a conceptual breakthrough. The composition is new: RWKV-7 for multi-frame sensor fusion, a simple causal-concatenation trick for linear cross-attention (LICA), and a diffusion decoder, all in one end-to-end driving pipeline. That is genuinely the first fully linear-attention driving model I know of. The positive side is real. Bench2Drive shows temporal context helps; the NAVSIM numbers are competitive; the edge-device latency measurements support the constant-time claim; and the Limitations section is unusually honest about the scorer bottleneck and the gap to iPad. I believe the authors are not overselling the planning results.\n\nThe soft spots are proportionate but real. First, Table I says memory is O(d) when the RWKV-7 state is a d×d matrix (Eq. 1); with a fixed feature dimension the memory is constant in sequence length, but the stated scaling in d is wrong. That should be corrected. Second, the stress-test concern about LICA is fair. Concatenating feature tokens and query tokens in a causal RWKV-7 pass means the query outputs see a fixed-size recurrent summary of the features, not per-query attention — and the feature tokens never see the queries. This is a plausible compression, but no ablation compares LICA to standard cross-attention or a bidirectional linear variant. Since the fully-linear design is the paper's main selling point, that omission is the weakest part of the evaluation. It does not sink the paper, but it limits how strongly the authors can claim LICA is an 'effective' replacement without more evidence.\n\nMissing code and error bars are minor for a systems paper at this stage, and the single-seed results are common in this subfield. The self-citation to FusionAD is background, not load-bearing.\n\nVerdict: this deserves a serious referee. The reviewer should push for a corrected complexity table, an ablation that swaps LICA for real cross-attention (or at least a bidirectional linear variant), and ideally a best-of-N comparison for LICA. I would bring it to a reading group for people interested in efficient end-to-end driving or recurrent state-space models. I would cite it in follow-up work. Send it to peer review.","headline":"A credible engineering contribution with a real constant-time inference story, but the memory claim needs fixing and LICA deserves an ablation.","tokens_in":13805,"tokens_out":1403,"would_cite":true,"duration_ms":15207,"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":"LADY proposes the first fully linear attention model for end-to-end autonomous driving, aiming to fuse arbitrarily long sensor history at constant inference cost.","keywords":["end-to-end autonomous driving","linear attention","linear cross-attention","temporal context fusion","diffusion policy","trajectory planning","edge deployment","multi-modal trajectories"],"falsifier":"Benchmark an identical model whose decoder cross-attention is replaced by standard quadratic attention with the same parameter count and training budget: if the quadratic version raises the planner's score on the open-loop benchmark by more than a noise-level margin, the causal design is losing information; if it does not, the linear shortcut is sufficient.","tokens_in":12834,"feed_emoji":"🚗","tokens_out":8541,"duration_ms":73970,"temperature":0.7,"pith_summary":"End-to-end driving systems today lean on attention blocks whose cost grows quadratically with the number of tokens, which makes long multi-frame history and on-vehicle deployment hard. This paper claims that a fully linear attention stack—where a fixed-size recurrent state carries all past camera and LiDAR information—can replace every transformer component in a driving planner, including decoder cross-attention, without sacrificing planning quality. To keep the whole pipeline linear, the authors introduce LICA, a lightweight cross-attention built from two causal passes and a concatenation trick, and pair it with a diffusion-based trajectory decoder. On standard driving benchmarks the model is competitive with transformer-based systems, and on the closed-loop benchmark an unbounded-history variant matches top baselines; at the same time, inference latency and memory stay flat as the number of past frames grows. If the claim holds, long-horizon temporal context becomes nearly free at run time, which is exactly the setting where today's methods are most taxed.","feed_headline":"LADY: constant-cost long-range planning without quadratic attention","feed_subtitle":"A fully linear attention driving planner matches top open- and closed-loop scores while latency stays flat as history grows.","key_machinery":"Two pieces of machinery carry the argument. The backbone is a recurrent linear-attention block with a delta-rule state update: it processes token sequences causally, compressing all past tokens into a fixed-size hidden state, so the per-token cost is constant in sequence length. The paper's own contribution is LICA (lightweight linear cross-attention): to let decoder queries attend to fused scene features, the queries are first encoded by one recurrent block, then concatenated after the BEV-ego token sequence and fed through a second recurrent block; the last query-position outputs are the cross-attended queries. This converts the typical quadratic query-to-feature softmax into a single caus","core_discovery":"LADY's central claim is that a fully linear attention architecture is sufficient for end-to-end planning: a recurrent block with a delta-rule hidden-state update compresses multi-frame camera and LiDAR tokens causally, and at inference each new frame is folded into a fixed-size state, so time and memory per step are constant regardless of how many frames have come before. The paper's own mechanism, LICA, performs cross-attention without a quadratic softmax: learnable query tokens are first encoded by one recurrent pass, concatenated after the fused bird's-eye-view and ego-feature sequence, and run through a second recurrent pass; the last query positions become the cross-attended queries tha","pith_inferences":["The causal, one-directional nature of LICA means features never directly see the queries; if bidirectional feature-to-query information proves necessary in tricky scenes, the constant-cost property may force a compromise. An ablation replacing LICA with quadratic cross-attention would quantify this.","The O(1) guarantee is an inference-time property; training still processes a fixed-length window in parallel, so the asymptotic advantage applies to deployment, not to the cost of fitting the model.","The 'infinite frames' variant effectively turns the driving model into a learned state-space memory; this suggests a connection to memory-augmented world models, where the hidden state could also be used for prediction and surprise detection beyond planning.","Since the best-of-N already exceeds human PDMS, the scoring head, not the trajectory generator, is the current ceiling; training the scorer on richer rule-based labels or closed-loop outcomes could unlock SOTA with this exact architecture."],"forward_implications":["Long sensor history becomes nearly free at inference: a vehicle can keep a running state from every frame it has seen, so behaviors that unfold over many seconds can inform the plan at constant marginal cost.","The same model fits on embedded vehicle computers, where transformer planners with long input windows would exceed real-time budgets.","The quality of the generated candidate trajectories is already near-human; a better learned scorer alone could lift the model past state-of-the-art results without changing the planner.","The framework's linear-attention backbone can be swapped for other recurrent linear-attention layers, so the efficiency property is not tied to one specific implementation."],"fun_headline_variants":["Linear attention driving: constant cost, zero quadratic blowup","LADY drives with O(1) memory as history grows","Autonomous driving without quadratic attention: LADY matches SOTA","Fully linear attention planner keeps latency flat over time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The design's success hinges on assuming that a single causal pass over the concatenated feature-and-query sequence passes all the information a bidirectional cross-attention would, even though features never look back at the queries.","fun_headline_variants_meta":{"raw":{"variants":["Linear attention driving: constant cost, zero quadratic blowup","LADY drives with O(1) memory as history grows","Autonomous driving without quadratic attention: LADY matches SOTA","Fully linear attention planner keeps latency flat over time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1304,"prompt_tokens":742,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":492}},"tokens_in":486,"tokens_out":562,"duration_ms":5443,"temperature":1.0,"reasoning_tokens":492,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:29:18.597986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Benchmark an identical model whose decoder cross-attention is replaced by standard quadratic attention with the same parameter count and training budget: if the quadratic version raises the planner's score on the open-loop benchmark by more than a noise-level margin, the causal design is losing information; if it does not, the linear shortcut is sufficient.","supporting_citations":[],"review_version":1}