{"id":"cce2b2b1-8c5f-4965-894e-47da46c4296f","arxiv_id":"2606.08653","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"FiberTune is a new fine-tuning objective that preserves action-fiber visual residuals in VLA policies, yielding performance gains on simulation and physical robot tasks.","lead":"The paper introduces FiberTune, a training objective for vision-language-action fine-tuning that preserves visual residuals along action fibers by using an online probe to filter action-predictive directions and align residuals to a frozen teacher. Smart generalists might care if this leads to more reliable robot policies for tasks like picking and placing without added runtime cost.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Online action probe may fail to isolate action-predictive directions without removing task-relevant residuals","rationale":"The reader's weakest_assumption directly identifies the unverified component required for the method to work as described; the abstract provides no additional evidence (e.g., probe diagnostics or ablation) that would move the verdict from UNVERDICTED.","tokens_in":1780,"tokens_out":297,"duration_ms":13461,"concrete_test":"Ablate the learned probe by replacing it with a fixed random linear projection of the same dimensionality; retrain under identical conditions on one CALVIN ABC-to-D run and one SO-101 physical run. If the SR(5) or success-rate gains disappear or reverse while the rank-regularization term remains, the probe estimation step is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the online action probe, trained jointly with the VLA policy, accurately identifies and filters only action-predictive directions from intermediate visual tokens so that the residual alignment to the frozen teacher preserves useful structure. If the probe (whose architecture, loss, and update schedule are not specified in the abstract) under- or over-estimates these directions, the filtered residuals either retain collapse or discard task information; the reported correlation between gains and increased alignment/rank does not rule out that the probe itself is the source of the observed difference rather than a neutral estimator.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that standard action-supervised fine-tuning of VLA policies constrains only action-changing directions and allows visual structure to collapse along local action fibers. It proposes FiberTune, a training-time objective that deploys an online action probe to estimate and filter action-predictive directions from intermediate visual tokens, aligns the resulting residuals to a frozen visual teacher, and regularizes their effective rank. Under identical training conditions the method yields consistent gains over task-loss-only baselines across six simulation settings (two benchmarks, two architectures) plus a physical SO-101 pick-place task, with representative improvements of +10.7 pp SR(5) on CALVIN ABC-to-D and 72.7 % to 78.1 % success on the physical task; these gains correlate with higher probe-filtered residual alignment and rank.","tokens_in":1878,"tokens_out":511,"duration_ms":16681,"significance":"If the central claim holds, FiberTune offers a practical, inference-free way to mitigate an under-appreciated form of visual collapse in VLA fine-tuning. The multi-benchmark, multi-architecture, and physical-robot validation is a clear strength, as is the explicit link between the proposed objective and the observed residual diagnostics. The work directly addresses a concrete failure mode that arises under standard imitation objectives.","major_comments":[{"comment":"Abstract: the reported performance gains (e.g., +10.7 pp SR(5) on CALVIN ABC-to-D and the physical SO-101 lift from 72.7 % to 78.1 %) are presented without error bars, number of runs, statistical significance tests, or explicit data-split details. This information is load-bearing for the claim of consistent improvement across all six controlled settings.","section":null},{"comment":"Method description of the online action probe: the architecture, loss function, and update schedule of the probe are not specified. Because the central claim rests on the probe correctly isolating only action-predictive directions so that residual alignment preserves task-relevant structure, the absence of these details leaves open the possibility that the probe itself (rather than the alignment step) is responsible for the observed difference, exactly as flagged by the stress-test concern.","section":null}],"minor_comments":[{"comment":"Abstract: the phrase 'representative gains' would be clearer if accompanied by a short table or explicit list of all six simulation results rather than a single example.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on result presentation and methodological transparency. We respond point-by-point to the major comments below.","responses":[{"response":"We agree that these details are necessary to support the consistency claim. In the revised manuscript we will report the number of runs (typically three random seeds), include error bars or standard deviations on the key metrics, note any statistical significance tests performed, and make data-split details explicit in the experimental section; a concise reference to run count will be added to the abstract where space permits.","revision_made":"yes","referee_comment":"Abstract: the reported performance gains (e.g., +10.7 pp SR(5) on CALVIN ABC-to-D and the physical SO-101 lift from 72.7 % to 78.1 %) are presented without error bars, number of runs, statistical significance tests, or explicit data-split details. This information is load-bearing for the claim of consistent improvement across all six controlled settings."},{"response":"We acknowledge that the current manuscript does not provide sufficient specification of the online action probe. We will expand the method section in the revision to fully describe the probe architecture, its loss function, and update schedule, and we will add analysis or controlled experiments to address the stress-test concern and clarify the distinct contribution of the alignment step.","revision_made":"yes","referee_comment":"Method description of the online action probe: the architecture, loss function, and update schedule of the probe are not specified. Because the central claim rests on the probe correctly isolating only action-predictive directions so that residual alignment preserves task-relevant structure, the absence of these details leaves open the possibility that the probe itself (rather than the alignment step) is responsible for the observed difference, exactly as flagged by the stress-test concern."}],"tokens_in":1467,"tokens_out":403,"duration_ms":19103,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a training objective that tries to keep visual residuals from collapsing in directions that do not affect predicted actions. The authors formalize this as collapse along local action fibers, then use an online action probe to identify and remove those directions from intermediate visual tokens before aligning the rest to a frozen teacher and adding a rank regularizer.\n\nThe experiments show the method beats plain task-loss fine-tuning in every one of the six controlled simulation runs (two benchmarks, two architectures) and on the physical SO-101 pick-place task. The reported numbers include a +10.7 point lift in SR(5) on long-horizon CALVIN and a jump from 72.7 % to 78.1 % success on the real robot. The residual diagnostics track the claimed mechanism: higher teacher alignment and effective rank after filtering.\n\nThe main soft spot is that the abstract supplies no error bars, no statistical tests, and no information on how the data splits or seeds were chosen. Without those, it is hard to judge whether the gains are stable or sensitive to post-hoc choices. The online probe itself is the load-bearing piece; if it mis-estimates the action-predictive directions, the filtered residuals could either keep the collapse or drop task-relevant structure. The paper correlates the gains with the alignment metrics, but that does not yet rule out the probe as the source of the difference rather than a neutral estimator.\n\nThis work is aimed at groups already fine-tuning VLA policies for robotics and who have seen visual features degrade on long-horizon tasks. It is worth sending to peer review because it names a concrete failure mode, ships controlled comparisons on multiple platforms, and includes some mechanistic checks, even though the statistical reporting and probe validation will need tightening.","headline":"FiberTune adds a residual alignment objective during VLA fine-tuning and reports gains across several settings, but the gains rest on an online probe whose reliability is not strongly validated in the abstract.","tokens_in":2356,"tokens_out":439,"would_cite":false,"duration_ms":10608,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Action-supervised fine-tuning of vision-language-action policies allows visual structure to collapse along action fibers, but FiberTune prevents this by filtering action-predictive directions and aligning the remaining residuals to a frozen","keywords":["vision-language-action","fine-tuning","visual residuals","action fibers","policy learning","feature collapse","robot learning","residual alignment"],"falsifier":"An ablation that applies the full FiberTune objective but disables the residual alignment loss and finds that performance gains over task-loss-only fine-tuning disappear would indicate the preservation mechanism is not responsible for the reported improvements.","tokens_in":2679,"feed_emoji":"🤖","tokens_out":777,"duration_ms":16203,"temperature":0.7,"pith_summary":"Standard action-supervised fine-tuning constrains only the feature directions that change predicted actions, leaving visual representations consistent across action-equivalent states free to collapse. FiberTune counters this collapse with an online action probe that identifies and removes action-predictive directions from visual tokens, then aligns the filtered residuals to a frozen teacher while regularizing their effective rank. The method runs only at training time and produces consistent gains over task-loss-only baselines across multiple simulation benchmarks, architectures, and a physical robot task. A sympathetic reader would care because the approach shows how to retain useful visual structure during policy adaptation, supporting better performance on long-horizon tasks without changing inference cost.","feed_headline":"FiberTune stops visual collapse during VLA fine-tuning","feed_subtitle":"Filtering action-predictive directions then aligning residuals to a teacher raises success on CALVIN and physical robots with no inference c","key_machinery":"The probe-filtered residual alignment objective, which removes estimated action-predictive directions from visual tokens before aligning the residuals to a frozen teacher and regularizing rank.","core_discovery":"Action-supervised fine-tuning of vision-language-action policies fits demonstrations effectively but constrains only the directions that change predicted actions, leaving visual structure consistent across action-equivalent states free to collapse. This is formalized as residual visual collapse along local action fibers. FiberTune uses an online action probe to estimate action-predictive feature directions, filters them from intermediate visual-token representations, and aligns the resulting probe-filtered residuals to a frozen visual teacher while regularizing their effective rank, preserving teacher-structured visual residuals without adding inference-time overhead.","pith_inferences":["The same filtering-plus-alignment pattern could be tested in other multimodal fine-tuning settings where supervision on outputs risks collapsing input representations that are invariant to those outputs.","A static rather than online probe might simplify implementation while retaining similar benefits if the action-predictive directions are stable across training.","The rank regularization term may combine with existing techniques such as weight decay or dropout to further stabilize visual features during adaptation.","Applying the method to additional robot platforms would test whether the action-fiber structure is consistent across embodiments."],"forward_implications":["Performance improves over task-loss-only fine-tuning in every one of six controlled simulation settings spanning two benchmarks and two architectures.","Physical robot task success on SO-101 pick-place rises from 72.7% to 78.1%.","Gains coincide with increased probe-filtered residual teacher alignment and effective rank.","The method applies to pi_0.5 and OpenVLA-OFT without inference overhead.","Long-horizon success rate SR(5) on CALVIN ABC-to-D increases by 10.7 percentage points."],"fun_headline_variants":["FiberTune preserves action-fiber residuals during VLA fine-tuning","FiberTune filters action-predictive directions to retain visual residuals","FiberTune aligns probe-filtered residuals to frozen visual teacher","Avoiding residual collapse along action fibers with FiberTune VLA tuning"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"An online action probe can reliably estimate action-predictive feature directions so that filtering them and aligning the residuals actually prevents harmful collapse without losing task-relevant information.","fun_headline_variants_meta":{"raw":{"variants":["FiberTune preserves action-fiber residuals during VLA fine-tuning","FiberTune filters action-predictive directions to retain visual residuals","FiberTune aligns probe-filtered residuals to frozen visual teacher","Avoiding residual collapse along action fibers with FiberTune VLA tuning"]},"model":"grok-4.3","cost_usd":0.005799,"raw_usage":{"total_tokens":2774,"prompt_tokens":694,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":57987000,"prompt_tokens_details":{"text_tokens":694,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2012,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":694,"tokens_out":68,"duration_ms":13940,"temperature":1.0,"reasoning_tokens":2012,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T18:40:14.098790+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An ablation that applies the full FiberTune objective but disables the residual alignment loss and finds that performance gains over task-loss-only fine-tuning disappear would indicate the preservation mechanism is not responsible for the reported improvements.","supporting_citations":[],"review_version":1}