Conditioning channels that don’t condition: shortcut reliance in instructed policies

2026-08-08 lit slice. Two papers: Robust Skills, Brittle Grounding (2602.24143) and DISC (2605.20856). Read because this afternoon’s frame-mining read (post) left the meta-report with a fact in need of a frame: the oracle-subgoal gain is FLAT across the aliasing spectrum — the conditioning channel helps everywhere equally rather than where the image underdetermines the action. The literature has a name for conditioning channels that don’t do the conditioning they advertise, and diagnostics for proving it. Fed: the fieldcond-subgoal-meta-report interpretation section, #6, #17.

Robust Skills, Brittle Grounding (2602.24143)

Contribution. A controlled diagnostic that decomposes what a single benchmark success rate conflates: can the policy execute the motion (reach, grasp-anything) vs did it do the instructed thing (instruction-conditioned success). Multi-object picking with a ladder of placement variability, up to 100k scripted demonstrations (10M frames), testing SmolVLA and π₀.₅.

Experiments. The decomposition is the story. SmolVLA at small jitter: 90% success, ~100% reach. Widen placement and execution survives while selection dies: large jitter 41% success but still 100% reach and 50% grasp-anything; full workspace randomization 2% success. With a single object present (no selection needed) full randomization recovers to 15% — the failure is grounding, not motor skill. The compositional cell is brutal: hold out object–region pairings (every object and every region seen individually in training) and success goes 44% → 0% at unchanged spatial difficulty. And the “more data” escape hatch is closed: 10k → 100k demos under full randomization moves SmolVLA 2% → 1% and π₀.₅ 4% → 6%. The policies read the instruction as a coarse region prior, not a per-scene referent.

Honest limits. Two policies, one synthetic picking domain, scripted demos; the abstract-level conclusion (“benchmarks measure manipulation, not following”) is argued from this one family.

DISC (2605.20856)

Contribution. Names the mechanism — task-state entanglement — and removes it structurally. When instruction and observation share network parameters, and scenes correlate with tasks in the data (they always do), the net learns scene→action shortcuts that bypass language entirely: their demos show Octo approaching the microwave when told to fetch a white bowl, π₀.₅ skipping instructed stove-activation because the scene context matches a related task. DISC’s fix: a two-stage hypernetwork maps the instruction to the parameters of a compact observation-only policy (coarse generation + learned iterative refinement, all feed-forward). The instruction cannot leak through the observation pathway because there is no shared pathway.

Experiments. LIBERO-90 from scratch 94.3% vs the best entangled baseline 86.6%, with the margin growing with task complexity (+0.9 easy → +8.4 long-horizon); beats pretrained π₀ (91.6%) without pretraining. The diagnostic that matters for us: a real-robot combinatorial benchmark — 9 tasks as 3 objects × 3 containers in an identical visual scene, so behavior variation can only come from language. DISC 86.4%, Octo 78.5% (failing by coarse scene→action association), plain hypernetwork 18.5%. Paraphrase-robustness (50 rephrasings per task) holds at 85.4%. Limits: heavier training, a compact target policy that costs fine placement precision on one task (53.3% vs 96.7%), and under-specified instructions surface uncertainty instead of being silently resolved by the scene — which the authors frame, fairly, as a feature.

What transfers to us, and what doesn’t

1. The meta-report gets its interpretive frame (direct feed). Our concentration null — Δ_oracle flat across aliasing, ~zero only on fully-determined frames — is the offline signature of what these papers measure online: a conditioning channel consumed as a coarse prior rather than a per-frame referent. Brittle Grounding’s decomposition (execution vs instructed-selection) is structurally our decomposition (pooled gain vs aliased-frame gain); their region-prior finding is our “uniform style/phase prior” reading. The report can now say: the −0.29 oracle-subgoal gain behaves like the coarse-prior regime the diagnostic literature documents, and the aliased-frame error floor (+29%) is exactly the part a prior cannot fix — DSSP says only extra information moves it, and these papers say the information must arrive through a channel the policy can’t shortcut around.

2. A cheap offline leakage read exists for our stack (candidate, not queued). Their strongest diagnostics are counterfactual conditioning at fixed scene (DISC’s combinatorial bench; the held-out pairings). Our panel apparatus already does forced-outcome counterfactuals (Q3); the subgoal analog — score frames under a wrong-episode subgoal and measure prediction sensitivity vs the true-subgoal pass — would quantify how much the slot’s content (vs its mere presence) moves the action. Rung (a) already hinted: the same text moved through the suffix channel scored +0.043 worse than the slot, and dropped-vs-present is −0.29 — presence is worth far more than placement. A swap read closes the triangle (presence vs content vs channel) for ~1 panel pass; it belongs in the meta-report’s open-questions, gated on its own pre-reg if it graduates.

3. Entanglement is an argument in the #4/#17 design space, not a recipe. DISC’s hypernetwork is far from our decoder-stack reality (and its precision regression on fine placement is disqualifying for a manipulation trunk). What transfers is the principle now backed by numbers: conditioning delivered through a separate structural path grounds better than conditioning mixed into shared tokens — worth one line when the attach-screen seam variants are debated, not a new arm.

4. The scaling null is a prior worth banking. 10k → 100k demos buying ~nothing under randomization is the cleanest published “more data doesn’t fix grounding” cell — relevant whenever a fieldgen or subgoal shortfall tempts a bigger-data response.

Doesn’t transfer: all success rates (sim/lab picking vs our offline teleop-corpus MAE); DISC’s from-scratch training regime (we warm-start VLM trunks precisely for the semantics entanglement erodes); their instruction-selection failure mode in its pure form — our panel conditions on episode-true fields, so we measure prior-vs- referent, never wrong-object grasping.

Where it fed

• fieldcond-subgoal-meta-report: the interpretation section’s external frame — the flat gain = coarse-prior consumption, the documented failure family; plus the presence/content/channel triangle as the open-questions candidate.
• #6 aux attribution: dated hook — the subgoal-swap sensitivity read named as the missing cell after today’s concentration null.
• #17 new trunks: structural-decoupling prior for future conditioning-path debates (one line, no arm).