Analogical RL Representation Comparison

{
  "query": "Compositional Transduction with Latent Analogies offline goal-conditioned reinforcement learning limitations invariant displacement bilinear transduction OGBench",
  "categories": [
    "cs.LG",
    "cs.RO"
  ],
  "max_results": 10
}

{
  "query": "contrastive GCRL analogy displacement invariance information bottleneck out-of-combination generalization CTA",
  "max_results": 5
}

{
  "title": "Preliminary JAX mechanism screen: spontaneous versus engineered context invariance",
  "research_question": "In a deliberately simple contextual goal-reaching problem, does vanilla online CRL spontaneously learn a context-invariant displacement code and OOC generalization, and do VIB or CTA-like analogy conditioning change that mechanism? This screen cannot support claims about original CTA or real tasks.",
  "method": "Build a JAX-native contextual 2D point-mass with explicit task displacement, 4 binary exogenous context bits, and a diagonal goal-quadrant×context-bucket holdout. Train vanilla CRL; CRL plus repository-derived VIB with beta in {0.001,0.01,0.1}; and a clearly labeled CTA-like analogy/bilinear control with bottleneck 16 and representation dimension 64. Compare online training with an offline retrain on the frozen online buffer.",
  "controls": [
    "Before training, unit-test OOC split absence/presence, invariance score on hand-crafted invariant and random codes, VIB KL at the unit-normal reference and away from it, bilinear gradient flow, encoder and ensemble shapes",
    "Use 1 seed for beta and horizon selection; use 3 seeds for vanilla, selected VIB, selected analogy, offline CRL, CTA-like offline, and any causal ablation",
    "Keep the held-out evaluation combinations fixed across methods",
    "Label CTA-like as a mechanism control, never as the original CTA baseline"
  ],
  "evaluation": "OOC and in-distribution success, mean pairwise cosine invariance of the displacement across contexts at fixed task, effective rank, and context decodability from the displacement. Inspect pilot dynamic range before committing the sweep.",
  "experiments": [
    "pilot and unit-test gate",
    "vanilla CRL baseline",
    "one-seed VIB beta screen followed by 3-seed confirmation of the selected setting",
    "one-seed analogy horizon screen followed by 3-seed confirmation",
    "offline CRL and CTA-like controls on frozen online data"
  ],
  "deliverables": [
    "unit-test report",
    "fixed-split metric table",
    "invariance-versus-OOC analysis",
    "explicit limitations statement"
  ]
}

{
  "proposal_id": "P-a17c"
}

{
  "run_id": "R-b24d",
  "section": "controls",
  "detail": "unit-test gate and pilot dynamic range before the full sweep"
}

{
  "run_id": "R-b24d",
  "instruction": "Replace the weak nuisance with 4 context bits plus 6 larger-magnitude nonlinear distractor dimensions that are fixed functions of context and inert to dynamics/reward. Re-run the relevant unit tests. Replace saturated full-representation context decoding with context decoding from the displacement at fixed task. Pilot all variants before the sweep; stop if the harder environment does not learn."
}

{
  "run_id": "R-b24d",
  "section": "summary",
  "detail": "recalibrated pilot and completed main sweep"
}

{
  "run_id": "R-b24d",
  "section": "metrics"
}

{
  "run_id": "R-b24d",
  "instruction": "Add the decisive fixed-context ablation: train online CRL with one context only and retrain offline on its frozen buffer, each with 3 seeds. Keep goal variation, optimization, and evaluation fixed. Report training-context success, varied-context/OOC success, invariance, and effective rank."
}

{
  "run_id": "R-b24d",
  "section": "controls",
  "detail": "fixed-context causal ablation"
}

{
  "title": "Synthetic CRL acquires invariant displacement from context variation",
  "statement": "In a contextual point-mass mechanism screen, vanilla CRL learns invariant displacement structure and OOC generalization without the CTA-like add-ons; removing context variation collapses OOC despite high cosine invariance.",
  "evidence": "Vanilla online CRL: OOC 0.996±0.006, invariance 0.991 over 3 seeds. Fixed-context online/offline CRL: OOC 0.270±0.035 / 0.169±0.037 and effective rank 1.24±0.10 / 1.23±0.04, while invariance remains 0.904±0.020 / 0.938±0.009. VIB beta 0.001→0.1 lowers invariance 0.983→0.589 and OOC 0.986→0.838.",
  "caveats": "Preliminary low-dimensional synthetic environment only. The CTA-like agent is not the original CTA implementation. The setting has a clean, well-covered context marginal and cannot support real-task or paper-level claims."
}

{
  "query": "synthetic CRL context variation invariant displacement CTA-like limitation",
  "max_results": 3
}

{
  "query": "original CTA code OGBench scene-play CRL HIQL per-environment comparison",
  "categories": [
    "cs.LG",
    "cs.RO"
  ],
  "max_results": 10
}

{
  "title": "Original CTA versus HIQL and CRL on the minimal real OGBench case",
  "research_question": "Does the synthetic coverage-substitution result survive in the simplest real high-dimensional offline environment where original CTA works and CRL fails, or does CTA retain an advantage even with dense data?",
  "method": "Use the original CTA release and its OGBench agents. Screen scene-play first, then a simpler alternative only if it yields a cleaner factored case. Compare original CTA, HIQL as the shared-backbone control, and OGBench offline CRL. After a full-data correctness gate, train on random trajectory fractions 1.0, 0.3, 0.1, and 0.03. Probe each exposed displacement-like code for exogenous-context invariance and effective rank, but treat the probe as secondary.",
  "controls": [
    "Run a storage check and inspect individual dataset sizes before downloading",
    "Require an end-to-end pilot with the original code and environment",
    "Use 3 seeds for full-data CTA/HIQL/CRL; use 1 seed per coverage point as an initial curve and label it provisional",
    "Sanity-check every probe rank against its representation dimension and inspect ensemble axes",
    "Attempt a targeted context×task holdout only after auditing whether trajectory-level removal preserves usable transitions"
  ],
  "evaluation": "Primary: fixed-suite success for original CTA, HIQL, and CRL, full data and versus random trajectory coverage. Secondary: exogenous-context invariance and effective rank. A targeted combination holdout is reportable only if each marginal remains supported and the filtered dataset remains a genuine play dataset.",
  "experiments": [
    "storage and original-code pilot",
    "minimal-environment screen and full-data 3-seed comparison",
    "single-seed random coverage curve",
    "representation probe with shape audit",
    "feasibility audit for targeted context×task holdout"
  ],
  "deliverables": [
    "minimal real benchmark identification",
    "full-data table with seed variation",
    "coverage curve with explicit seed caveat",
    "probe table",
    "holdout feasibility verdict"
  ]
}

{
  "proposal_id": "P-c39e"
}

{
  "run_id": "R-d41f",
  "section": "summary",
  "detail": "storage, original-code pilot, environment dimensions, and early screening gate"
}

{
  "run_id": "R-d41f",
  "instruction": "Fix scene-play-v0 as the minimal real benchmark. Complete original CTA, HIQL, and CRL at 500k matched steps with 3 seeds on full data. Run the 1-seed random trajectory coverage curve at 100%, 30%, 10%, and 3%. Save parameters for the representation probe. Audit retained transitions before treating any targeted context×task holdout as valid evidence."
}

{
  "run_id": "R-d41f",
  "section": "metrics",
  "detail": "full-data headline and coverage curve"
}

{
  "run_id": "R-d41f",
  "section": "qualitative",
  "detail": "raw representation probe and tensor-shape sanity checks"
}

{
  "run_id": "R-d41f",
  "instruction": "Fix the CRL representation probe by selecting one critic ensemble member consistently rather than flattening the (2,B,512) ensemble axis. Recompute CRL invariance and effective rank; assert effective_rank <= representation_dim before accepting. Leave CTA and HIQL probes unchanged."
}

{
  "run_id": "R-d41f",
  "section": "controls",
  "detail": "corrected probe and targeted-holdout feasibility audit"
}

{
  "run_id": "R-d41f",
  "reason": "Stop and discard the targeted context×task holdout phase: filtering removes 95% of transitions and all substantive play trajectories, so zero performance is a dataset-deletion artifact. Valid full-data, random-coverage, and corrected-probe phases are complete."
}

{
  "title": "Original CTA beats HIQL and CRL on scene-play; targeted combination deletion is infeasible",
  "statement": "On the minimal real OGBench scene-play case, original CTA outperforms the shared-backbone HIQL control and CRL at full data. The toy coverage-substitution story does not transfer cleanly: CRL stays at zero and CTA retains a full-data advantage, while the CTA-HIQL coverage gap is non-monotonic. A trajectory-level context×task holdout is invalid because it deletes nearly all substantive data.",
  "evidence": "At 500k matched steps over 3 seeds: CTA 0.820±0.05, HIQL 0.498±0.05, CRL 0.002±0.003. Single-seed 100/30/10/3% curves: CTA 0.787/0.767/0.193/0.020; HIQL 0.520/0.300/0.053/0.007; CRL 0 throughout. Corrected invariance/effective-rank probes: CRL 0.586/319, CTA 0.743/117, HIQL 0.822/9.1. Targeted filtering removes 955000 of 1001000 transitions and leaves 44-step stubs, so its all-zero outcome was discarded.",
  "caveats": "Training used 500k rather than 1M steps. CRL's floor may reflect known scene-play weakness, undertraining, or tuning rather than isolated OOC failure. Coverage curve is single-seed. Probe representations are not architecturally identical. No valid targeted context×task holdout was obtained."
}