{"id":"c2c28724-9232-41c6-97f0-ddb9954093d7","arxiv_id":"2607.10206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Orthogonal Source Lifting makes flow-matching imitation intervenable by selecting only the source endpoint, changing future routes in 91.1% of matched-prefix tests while keeping free-deployment performance strong.","lead":"SL-FM turns passive noise in flow-matching robot policies into a selectable source handle that chooses among valid futures from the same state. It does this without mode-specific velocity fields, using orthogonal lifting so branches keep their identity through path crossings.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged weakest assumption.","rationale":"The paper's strongest claim is empirical and interface-focused, not a universal theorem. The same-prefix design (identical environment seed, matched stochastic draws outside a five-call window, same-handle control at 0.000 m / 0% route change) already isolates the causal effect of the source handle. Orthogonal Source Lifting keeps targets on the zero-lift plane (Eq. 11), so the demonstrated action marginal is preserved by construction; the shared field never receives z. Free-deployment tables and the Kitchen subtask-handle correlation supply independent support that the handles remain usable. The only real soft spot is precisely the one the reader already identified: whether the modest mixture + orthogonal anchors continue to prevent identity collapse when crossings become denser or action dimensionality rises. That is a scope limitation, not a contradiction of the reported evidence, so the CONDITIONAL verdict with medium correctness risk and moderate confidence needs no adjustment.","tokens_in":15010,"tokens_out":537,"duration_ms":5144,"concrete_test":"Re-run the D3IL Avoiding same-prefix protocol of Table 3 after ablating lift scale λ to 0 (recovering SL-FM w/o lift) and after increasing action-chunk dimension or obstacle density; if route-change rate collapses toward the FM noise-resample baseline while free-deployment success stays high, the identity-preservation claim is regime-limited exactly as the reader flags.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that source geometry alone (state-conditioned Gaussian mixture + Orthogonal Source Lifting, Eqs. 10–13, with responsibility floor Eq. 14) turns passive flow-matching randomness into a reliable, source-only intervention handle without ever feeding z into the shared velocity field. The manuscript already supplies the right causal evidence for that claim on the stated regime: same-prefix 91.1% route change with same-handle control at 0, w/o-lift ablation drop, free-deployment competitiveness, and selector study. The reader's weakest assumption correctly names the soft spot—whether modest K and orthogonal lifts continue to separate identity under denser crossings or higher-dimensional action chunks—but that is an ordinary scale/regime limit for a methods preprint, not an internal inconsistency or missing control that undermines the reported results. No stronger load-bearing flaw (e.g., hidden z-conditioning, broken target-marginal preservation, or non-causal intervention protocol) is present in the text.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Source-Lifted Flow Matching (SL-FM), a conditional flow-matching policy for multimodal imitation that exposes a discrete source handle z without conditioning the velocity field on z. A state-dependent Gaussian mixture defines selectable source endpoints; Orthogonal Source Lifting embeds handle-specific sources into auxiliary orthogonal coordinates while keeping all targets on the zero-lift plane (Eqs. 10–13), so one shared latent-free field can transport different branches without identity collapse at crossings. A responsibility floor (Eq. 14) keeps redundant handles trainable. Free deployment samples z from the learned prior; intervention sets z at a decision state and integrates the same field. On D3IL Avoiding, same-prefix interventions change future routes in 91.1% of pairs (same-handle control 0%), free-deployment success is competitive or best among same-harness baselines (Tables 1–2), and a frozen-policy high-level selector can use z for route control.","tokens_in":15303,"tokens_out":664,"duration_ms":5592,"significance":"If the result holds, SL-FM supplies a clean, source-only intervention interface for generative imitation policies that is stricter than mode-conditioned fields or latent steering: the handle never enters the velocity field, the target action marginal is preserved, and same-prefix counterfactuals with a same-handle null control provide causal evidence of local controllability. The constructive method (Algorithm 1), w/o-lift ablation, MI diagnostics, and selector study make the contribution concrete and usable for planning or human-in-the-loop control on multimodal robot tasks. Strengths include the matched-prefix protocol, explicit preservation of pdata(a|s), and competitive free-deployment averages on Avoiding/Aligning/PushT.","major_comments":[{"comment":"The central claim that source geometry alone keeps branch identity through crossings rests on modest K and orthogonal lifts (Methodology, Eqs. 10–13; responsibility floor Eq. 14). The manuscript already shows this works on the reported regime (Avoiding 91.1% route change, w/o-lift drop 0.825→0.733, same-handle control 0%). However, the weakest assumption is left largely untested: denser crossings, higher-dimensional action chunks, or poorly aligned mixtures could still cause the shared field to average branches. A short stress test (e.g., synthetic denser-crossing diagnostic or higher-da chunk ablation) would substantially strengthen the load-bearing claim that the handle remains reliable without ever feeding z into vθ.","section":null},{"comment":"Table 3 and the same-prefix protocol are the primary causal evidence, yet the intervention window is fixed to five consecutive policy calls at the first high-entropy pre-obstacle band. It is not shown how sensitive the 91.1% / 57.8% both-success rates are to window length, decision-point selection, or later obstacles (where free-rollout MI for digit2 is weaker). Clarifying robustness of the intervention protocol would make the causal claim more transferable beyond the Avoiding geometry.","section":null}],"minor_comments":[],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know is that this paper turns passive source noise in flow-matching imitation into a local, selectable handle without ever feeding the discrete code into the velocity field. Orthogonal Source Lifting is the real move: lift only the sources into auxiliary orthogonal coordinates, keep every target on the zero-lift plane so the demonstrated action marginal is untouched, and train one shared field. That is cleaner than the usual mode-conditioned or latent-steering alternatives they cite.\n\nWhat they do well is the evidence package for the interface claim. Same-prefix counterfactuals on D3IL Avoiding change future routes in 91.1% of pairs, the same-handle control sits at 0%, the w/o-lift ablation drops success from 0.825 to 0.733, free-deployment stays competitive (best average among the same-harness baselines on Avoiding/Aligning/PushT), and a frozen-policy selector can drive commanded routes. Crossing-flow diagnostics and MI peaks at decision phases line up with the story. Math is standard CFM plus a mixture prior, lift, and responsibility floor; citations cover the right FM, OT-coupling, and multimodal imitation lines without obvious gaps.\n\nSoft spots are ordinary for a methods preprint, not load-bearing. The weakest assumption is real: modest K and orthogonal lifts may not keep identity under denser crossings or higher-dimensional action chunks. Several free hyperparameters (K, floor α, lift scale, weights, Euler steps) and no public code yet. Kitchen gains are modest and Aligning is basically flat. None of that breaks the reported regime or the causal protocol.\n\nThis is for people building hierarchical or human-in-the-loop robot policies who want a continuation dial on top of a shared flow field. It deserves a serious referee. I would engage with it and expect it to survive review with the usual scale and ablation requests.","headline":"Solid robotics methods paper: source-only handles plus Orthogonal Source Lifting make multimodal flow policies intervenable without mode-conditioned fields, with the right causal controls.","tokens_in":15937,"tokens_out":468,"would_cite":true,"duration_ms":4807,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A shared flow-matching policy can turn passive source noise into a real choice among robot futures by selecting only where the flow starts, not a mode-specific field.","keywords":["flow matching","imitation learning","multimodal policies","source intervention","orthogonal source lifting","robot control","generative policies"],"falsifier":"On a same-prefix intervention protocol like D3IL Avoiding, if changing only the local source handle fails to redirect future routes in a large fraction of successful pairs—or produces composite averaged trajectories at known crossings—while free deployment still succeeds, the claim that source geometry alone yields an actionable handle would be falsified.","tokens_in":15884,"feed_emoji":"🤖","tokens_out":923,"duration_ms":17076,"temperature":0.7,"pith_summary":"Flow-matching policies can represent several valid robot actions for the same observation, yet that diversity is usually only random sampling: a user cannot deliberately pick which continuation happens from a fixed state. This paper argues that the choice can be exposed without splitting the model into mode-conditioned dynamics. The discrete handle selects only the source endpoint of a single shared, latent-free velocity field. Orthogonal Source Lifting raises handle-specific sources into auxiliary orthogonal coordinates while keeping every demonstrated target in the original action subspace, so branch identity survives path crossings and the demonstrated action distribution is left intact. A state-dependent source mixture plus a responsibility floor keeps every handle trainable. On crossing diagnostics and robot benchmarks, changing the handle redirects future routes in 91.1% of matched-prefix interventions while free-deployment performance remains competitive.","feed_headline":"Source geometry alone steers robot futures from one shared field","feed_subtitle":"Lifting handle starts into orthogonal coordinates redirects routes in 91% of matched-prefix tests without mode-conditioned dynamics.","key_machinery":"Orthogonal Source Lifting: each handle’s source is embedded with an auxiliary orthogonal coordinate while every target action stays at zero lift in the original action space; this keeps source identity through crossings so one shared velocity field can transport different branches without merging.","core_discovery":"Source-Lifted Flow Matching shows that a discrete handle which chooses only the source endpoint of a conditional flow—never conditioning the velocity field—is enough to make multimodal imitation intervenable, when handle-specific sources are lifted into orthogonal auxiliary coordinates and all targets remain on the original action subspace. One shared latent-free field then carries distinct branches through crossings without averaging them into composite trajectories, converting passive source randomness into an actionable local control variable.","pith_inferences":["Language, goal images, or other semantic signals could bias the state-dependent source prior to ground high-level intent without ever conditioning the shared velocity field on a mode label.","If lift dimension and mixture capacity scale with denser crossings in high-dimensional action chunks, the same source-only interface may extend to long-horizon generalist policies that currently rely on latent or mode-conditioned heads.","Same-prefix causal tests of the kind used here could become a standard check for whether any discrete “mode” variable in a generative policy is actually intervenable rather than merely correlated with outcomes.","The responsibility floor pattern may generalize to any overcomplete discrete interface attached only to generative sources, as a simple way to keep unused handles trainable."],"forward_implications":["A planner or operator can set a local source handle at a decision state to choose among valid futures that share the same prefix.","Multimodal robot policies need not condition the velocity field on a discrete mode to expose controllable branches.","Free sampling from the learned state-dependent prior recovers ordinary stochastic imitation while the same interface supports deliberate intervention.","A high-level selector can treat the source handle as a compact discrete action space over routes or subtasks with the low-level field frozen.","Crossing-induced composite trajectories that arise under a shared field can be removed by the lift construction without partitioning the target action distribution."],"fun_headline_variants":["Shared field steers routes via source handles alone","Orthogonal source lift makes flow matching intervenable","Handle picks source endpoint not mode-conditioned field","Source lifting removes composites keeps one velocity field","State-dependent sources turn passive multimodality into control"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a modest number of state-conditioned Gaussian source components plus orthogonal lift coordinates is enough to keep chosen branches distinct through real closed-loop crossings without ever feeding the discrete choice into the velocity field.","fun_headline_variants_meta":{"raw":{"variants":["Shared field steers routes via source handles alone","Orthogonal source lift makes flow matching intervenable","Handle picks source endpoint not mode-conditioned field","Source lifting removes composites keeps one velocity field","State-dependent sources turn passive multimodality into control"]},"model":"grok-4.5","effort":"low","cost_usd":0.0045,"raw_usage":{"total_tokens":1344,"prompt_tokens":843,"num_sources_used":0,"completion_tokens":54,"cost_in_usd_ticks":45000000,"prompt_tokens_details":{"text_tokens":843,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":447,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":843,"tokens_out":54,"duration_ms":4444,"temperature":1.0,"reasoning_tokens":447,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T13:29:34.061720+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a same-prefix intervention protocol like D3IL Avoiding, if changing only the local source handle fails to redirect future routes in a large fraction of successful pairs—or produces composite averaged trajectories at known crossings—while free deployment still succeeds, the claim that source geometry alone yields an actionable handle would be falsified.","supporting_citations":[],"review_version":1}