{"id":"f883c5c5-6d10-4e30-b9e2-be9d5a62e870","arxiv_id":"2608.07596","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A local cross-layer routing mechanism for VLA models, feeding each action-decoder block features from a small window of adjacent VLM layers, improves success rates on LIBERO, CALVIN, and LIBERO-Plus.","lead":"This paper introduces LIRA, a way to route information between layers of a vision-language model and a robot action decoder. On the LIBERO-Plus benchmark, it raises zero-shot task success from 59.1% to 78.0% compared with the VLA-Adapter baseline at the same model size.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The depth-locality claim rests on the layer index i being semantically meaningful, but the paper never tests whether permuting or reversing the routing window preserves the gain; if it does, the centered-window mechanism may be an artifact of fixed indexing.","rationale":"The reader's weakest assumption is exactly the one that carries the paper's conceptual weight: that the layer index i in Eq. (1) is semantically aligned with decoder depth. My concern is not about the internal consistency of the architecture, which is coherent, nor about the controlled comparison to VLA-Adapter, which is reasonably designed. The empirical ablations (Tables 6 and 7) rule out simple capacity increases and global aggregation, which is real evidence for a local-window benefit. What they do not rule out is that the benefit depends on having some small fixed set of neighboring layer indices, rather than on those indices corresponding to semantically adjacent VLM depths. A permutation or reversal test would directly separate these two explanations. The reader's CONDITIONAL verdict already appropriately flags the need for verification, and this test should be part of that condition; it does not change the verdict, so I leave it UNCHANGED.","tokens_in":12159,"tokens_out":5889,"duration_ms":64055,"concrete_test":"Re-run the LIBERO-Long and CALVIN-5 ablations (§4.3) with two additional routing assignments while keeping all other hyperparameters identical: (1) reverse layer order (PFB i receives a window centered at layer L−i+1), and (2) a fixed random permutation of layer indices for the windows, preserving the same 3-layer/2-layer boundary clipping. If reversed/permuted routing matches the identity-centered 97.6/80.5 numbers, the depth-locality prior is not supported and the gains are attributable to fixed-index feature mixing; if performance drops substantially, the centered alignment is doing genuine semantic work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the identification of VLM layer index with decoder block index in Eq. (1): W_r(i) is centered on the matched layer i, and every ablation—window width, topology, query budget—preserves this identity. The paper's core conceptual claim is that local depth structure, not merely feature diversity or token count, explains the gains (Table 7: 3-layer window 97.6/80.5 vs matched 95.4/77.9, global 93.7/76.9). But the same numbers would arise if adjacent layer indices were a convenient fixed indexing that supplies complementary features for reasons unrelated to VLM depth. VLM layers are not guaranteed to be monotonic in semantic abstraction after LoRA fine-tuning; a reversed or randomly permuted assignment of the same window sizes could perform equally well if the benefit comes from having a small fixed set of nearby-but-distinct VLM states rather than from genuinely depth-local structure. Since the contribution is specifically a depth-aware routing prior, this untested identity is the central load-bearing assumption. The absence of released code/checkpoints and of seed-level variance makes the issue unverifiable from the paper alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LIRA, a local cross-layer routing mechanism for vision-language-action (VLA) models. Rather than conditioning each action-decoder block on a single matched VLM layer (as in VLA-Adapter), LIRA aggregates query-token features from a small window of adjacent VLM layers centered on the matched depth, while retaining the layer-aligned task-token branch. The mechanism concatenates features along the token dimension and therefore introduces no new trainable parameters. The authors evaluate LIRA against VLA-Adapter and other systems on LIBERO, CALVIN ABC→D, LIBERO-Plus (zero-shot and fine-tuned), and a Franka real-robot setup, reporting consistent gains, most notably a 59.1% to 78.0% improvement in zero-shot LIBERO-Plus success. Ablations cover conditioning source, window width, routing topology, and query-token budget.","tokens_in":12563,"tokens_out":5788,"duration_ms":55887,"significance":"If the results hold, LIRA is a simple, parameter-free architectural improvement that exposes complementary intermediate VLM features to a compact action decoder, with a particularly large robustness gain under distribution shift. The controlled comparison at the same 0.5B scale, the ablation that rules out query-token budget as the sole cause of the gain, and the explicit no-additional-parameters property are clear strengths. However, the central depth-locality claim—that adjacency in VLM depth, rather than merely having a small set of distinct VLM states, explains the improvement—is not directly tested, and the absence of error bars makes the smaller gains difficult to assess. The paper's appended limitation statement is admirably candid about the single-backbone and single-robot scope.","major_comments":[{"comment":"The central claim is that local depth structure explains the gains, but the paper never varies the identity of the VLM layer matched to each decoder block. Since W_r(i) is always centered on the matched layer i, the ablations in Tables 5–7 cannot distinguish 'adjacent VLM layers are semantically complementary' from 'any small set of nearby-but-distinct VLM states helps.' Please add a permutation or reversal test: for example, route W'_r(i)={i-1,i+1,i+2}, random triples of layers, or a reversed layer ordering, and compare on LIBERO-Long and CALVIN-5. If these variants perform as well as the centered window, the depth-locality prior is not supported; if they degrade, the claim is strengthened.","section":"§3.2, Eq. (1)"},{"comment":"No error bars or multiple-seed runs are reported. The LIBERO average gain is 1.6 points (97.3% to 98.9%) and the CALVIN five-task gain is 4.0 points (76.5% to 80.5%); without training-seed variance it is impossible to tell whether these improvements are stable. Please report mean±std over at least three seeds for the main LIRA vs. VLA-Adapter comparisons, or explicitly justify why the evaluation protocol has negligible variance.","section":"§4.1, §4.2"},{"comment":"VLA-Adapter is called a 'controlled architectural baseline' and the paper states that the same objective and training recipe are used, but it is not explicitly stated whether the VLA-Adapter numbers were obtained by retraining in the same codebase or taken from the original publication. Please state this clearly. If the numbers are from the original paper, the 'controlled' claim is not supported and the comparison should be re-run under identical training conditions.","section":"§4.1"}],"minor_comments":[{"comment":"The routing topologies Flashback, Column-wise, and Q-Former are not defined in the main text; please define them in the text or refer to the appendix in the table caption.","section":"§4.3, Table 6"},{"comment":"The sentence 'All ablations follow the same compact training and evaluation protocol unless stated otherwise' is vague; please specify which ablations deviate, such as the query-budget ablation in Table 5.","section":"§4.1"},{"comment":"With 10 trials per task, differences of one success are well within sampling noise; please add per-task confidence intervals or explicitly frame these results as feasibility demonstrations.","section":"§4.5, Table 9"},{"comment":"The limitation statement is appropriate, but the single-backbone and 10-trial real-world caveats should also appear in the abstract or introduction so that readers do not overinterpret the headline numbers.","section":"§5"},{"comment":"The boundary clipping of the window is described only in prose; please write the full definition of W_r(i) with explicit boundary handling.","section":"§3.2, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for cs.RO and the core idea is interesting, but the missing permutation test for the depth-locality claim, the lack of seed-level variance, and the ambiguity about whether the VLA-Adapter baseline was retrained prevent acceptance in current form. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid architecture paper, not a breakthrough. The change is simple: instead of feeding each Parallel Fusion Block only the matched VLM layer's query features, LIRA concatenates query features from the matched layer and its immediate neighbors (r=1). No extra parameters, same training recipe, controlled against VLA-Adapter at the same 0.5B scale. That is the right way to test an interface change, and they do it well. Gains are consistent across LIBERO, CALVIN, and LIBERO-Plus; the zero-shot LIBERO-Plus jump from 59.1 to 78.0 is the headline. The ablations matter: local three-layer window beats matched-layer-only, global aggregation, and alternative topologies, and the query-budget sweep shows the gain is not just more tokens. Credit where due: this is a genuine comparison, not benchmark tourism.\n\nSoft spots, in rough order of importance. First, no error bars or multiple seeds anywhere; given the ablations are close (95.4 vs 97.6 on LIBERO-Long), we don't know if the local-window margin is stable. Second, code and checkpoints are promised but not released, so the key numbers can't be checked. Third, the default window radius and query budget were chosen with the evaluation benchmarks; the appendix and ablations mitigate this but don't eliminate it. Fourth, the real-robot results are 10 trials per task and should stay preliminary, as the paper itself says.\n\nThe stress-test concern about depth-locality is real and worth stating plainly: they never test whether the centered, depth-aligned window is doing the work or whether any small fixed set of nearby-but-distinct VLM states would do as well. A reversed or shuffled layer assignment would settle this. That said, it does not sink the empirical claim. The local-window topology still outperforms the alternatives; the question is only what story to tell about why. The paper currently over-interprets \"depth-aware\" as if the semantics were established.\n\nBottom line: for people working on compact VLA models this is a useful data point. It deserves a serious referee. I'd send it to review and ask for seeds/error bars, code release, and a permutation/reversal ablation before acceptance.","headline":"LIRA is a clean, well-controlled empirical paper on a small but real architectural change—routing each action-decoder block a local window of adjacent VLM layers—and its main unresolved issue is interpretive, not evidentiary.","tokens_in":12974,"tokens_out":2026,"would_cite":true,"duration_ms":20058,"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":"This paper claims that conditioning each action-decoder block on a small centered window of adjacent VLM layers, rather than one matched layer or all layers, improves VLM-to-action routing and raises zero-shot LIBERO-Plus success from…","keywords":["vision-language-action models","VLM-to-action interface","cross-layer routing","depth-aligned window","action decoding","robot manipulation","zero-shot transfer","LIBERO-Plus"],"falsifier":"A concrete check: retrain the same LIRA protocol with the routing window center deliberately offset from the matched layer (for example, centering block $i$ on VLM layer $i+3$) and compare LIBERO-Plus zero-shot success; if the offset preserves the 78.0% level, then depth alignment is not load-bearing, while a sharp drop would confirm that exact layer identity matters. A second check is to permute VLM layer indices while keeping the visual features fixed: if the gains persist after permutation, the locality prior is not about true depth ordering.","tokens_in":1961,"feed_emoji":"🤖","tokens_out":1958,"duration_ms":78754,"temperature":0.7,"pith_summary":"LIRA proposes a specific answer to a previously under-studied question: which intermediate layers of a pretrained vision-language model should feed the action decoder of a robot policy. The paper argues that neither a single matched layer nor the full VLM hierarchy is the right source: each decoder block should receive LIRA Query features from a small window of neighboring VLM layers centered on its matched layer. This depth-aware local routing adds no trainable parameters to the VLA-Adapter baseline, and the paper reports consistent success-rate gains on LIBERO, CALVIN ABC→D, LIBERO-Plus, and a physical robot. The headline result is a zero-shot LIBERO-Plus success improvement from 59.1% to 78.0% at the same 0.5B-parameter scale. The paper's discovery, if correct, is that local cross-layer structure, not layer count or token budget, is what the VLM-to-action interface is missing.","feed_headline":"Local VLM layer windows lift zero-shot robot success to 78%","feed_subtitle":"At the same 0.5B scale, a three-layer centered window beats one-layer and all-layer conditioning.","key_machinery":"The load-bearing object is the depth-aligned local routing window of Eq. (1): for Parallel Fusion Block $i$, the window is $W_r(i)=\\{\\ell\\in\\{1,\\dots,L\\}: |\\ell-i|\\le r\\}$, and the LIRA Query features $\\mathbf{R}_t^{(\\ell)}$ from those layers are concatenated along the token dimension to form $\\tilde{\\mathbf{R}}_t^{(i)}$. With $r=1$ this is a centered three-layer neighborhood, and the concatenation keeps the mechanism parameter-free because the decoder's cross-attention already consumes variable-length token sequences. The matched task-token features $\\mathbf{T}_t^{(i)}$ remain the depth-aligned semantic anchor, while $\\tilde{\\mathbf{R}}_t^{(i)}$ supplies complementary nearby-depth context. This machinery carries the paper's locality prior over VLM depth and defines the ablation space (window width, routing topology, and query budget).","core_discovery":"On the paper's own terms, the central claim is that VLM-to-action conditioning is best understood as depth-aware information routing, and that the optimal routing policy is a depth-aligned local window: for the i-th Parallel Fusion Block, the routing window selects all VLM layers within distance r of layer i, with r=1 by default, giving a centered three-layer window (two layers at the boundaries). The LIRA Query features from the selected layers are concatenated along the token dimension and fed through the block's existing cross-attention, so no new projection layers or trainable parameters are introduced. The paper asserts that this centered three-layer window outperforms matched-layer-only conditioning, two- and four-layer neighborhoods, and global aggregation on LIBERO-Long and CALVIN-5, and that it improves the controlled VLA-Adapter baseline on every principal aggregate metric. The strongest reported outcome is zero-shot LIBERO-Plus success rising to 78.0% from the baseline's 59.1%.","pith_inferences":["Editorial inference: the depth-locality prior likely transfers to any multi-layer backbone feeding a depthwise decoder (for example, diffusion policies with per-block conditioning), not only VLM-based action decoders; a direct test would be to route local windows of a frozen vision encoder into a diffusion action head.","Editorial inference: the LIBERO-Plus gains may reflect a cheap ensembling effect, since neighboring VLM layers are partially decorrelated; this could be tested by measuring whether the success gain tracks inter-layer feature similarity.","Editorial inference: the design effectively restricts the VLM-to-decoder connectivity matrix to a diagonal band of width $2r+1$; a testable extension is to tune $r$ per task or per benchmark rather than fixing $r=1$.","Editorial inference: because all headline comparisons use one 0.5B backbone, the optimal window width may depend on model depth; sweeping $r$ on a larger VLM would show whether the locality prior is scale-invariant."],"forward_implications":["At a fixed 0.5B parameter budget, replacing one matched VLM layer with a centered three-layer window raises LIBERO-Plus zero-shot success from 59.1% to 78.0%, implying that the routing interface itself, not model capacity, drives the robustness gain.","Long-horizon tasks inherit the largest benefit: LIBERO-Long success rises from 95.0% to 97.6% and CALVIN five-task completion from 76.5% to 80.5%, so local cross-layer context matters most when actions depend on sustained task evidence.","Query-token count alone cannot explain the gain: even 256 last-layer tokens reach only 92.8% on LIBERO-Long, below LIRA's 64-token, three-layer result of 97.6%.","Window-width ablations show a centered three-layer window outperforms one-, two-, four-layer, and global aggregation, so the optimal routing is neither single-layer nor full-hierarchy mixing.","On the physical robot, LIRA improves pick-and-place from 7/10 to 9/10 and two-stage task completion from 6/10 to 8/10, suggesting the benefit carries across subgoals, though the paper calls these results preliminary."],"supporting_citations":[{"why":"VLA-Adapter is the controlled architectural baseline; it supplies the one-to-one layer-aligned Parallel Fusion Block interface and the supervised continuous-action training recipe that LIRA keeps unchanged.","marker":"(Wang et al. 2026b)"},{"why":"LIBERO-Plus is the 10,030-task perturbation benchmark where LIRA's largest zero-shot gain (59.1% to 78.0%) is measured.","marker":"(Fei et al. 2026)"},{"why":"LIBERO supplies the four manipulation suites, including LIBERO-Long where the single-layer-to-local-window gains are largest.","marker":"(Liu et al. 2023)"},{"why":"CALVIN ABC→D supplies the long-horizon sequential evaluation where LIRA improves five-task completion from 76.5% to 80.5%.","marker":"(Mees et al. 2022)"},{"why":"OpenVLA-OFT is the strongest zero-shot comparison system on LIBERO-Plus and represents the final-layer action-head interface LIRA is measured against.","marker":"(Kim, Finn, and Liang 2025)"},{"why":"π0.5 shows a multi-layer, depth-conditioned VLM-to-action interface, providing a design contrast for LIRA's local-window routing.","marker":"(Black et al. 2025)"},{"why":"Qwen2.5-0.5B is the language backbone used in the matched 0.5B configuration, so all controlled comparisons share this backend.","marker":"(Yang et al. 2024)"}],"fun_headline_variants":["Zero-shot robot success hits 78% with local VLM windows","Depth-aware routing lifts zero-shot robot success to 78%","Local cross-layer windows boost robot action success to 78%","LIRA's local window routing pushes zero-shot robot success to 78%"],"cache_read_input_tokens":15104,"weakest_assumption_plain":"The load-bearing premise is that VLM layer $i$ and decoder block $i$ are meaningfully aligned in depth, so a window centered on layer $i$ has a principled center; if the VLM's layer ordering is not semantically monotonic, the center is arbitrary and the reported gains could be an artifact of this matching.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot robot success hits 78% with local VLM windows","Depth-aware routing lifts zero-shot robot success to 78%","Local cross-layer windows boost robot action success to 78%","LIRA's local window routing pushes zero-shot robot success to 78%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000937,"raw_usage":{"total_tokens":4035,"prompt_tokens":1002,"completion_tokens":3033,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2958}},"tokens_in":618,"tokens_out":3033,"duration_ms":21555,"temperature":1.0,"reasoning_tokens":2958,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:30:52.643161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: retrain the same LIRA protocol with the routing window center deliberately offset from the matched layer (for example, centering block $i$ on VLM layer $i+3$) and compare LIBERO-Plus zero-shot success; if the offset preserves the 78.0% level, then depth alignment is not load-bearing, while a sharp drop would confirm that exact layer identity matters. A second check is to permute VLM layer indices while keeping the visual features fixed: if the gains persist after permutation, the locality prior is not about true depth ordering.","supporting_citations":[],"review_version":1}