Today
-
Preserve decision-aware residual learning evidence
-
Preserve frozen readout and fixed budget transfer diagnostics
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve twelve-group sampled credit trajectories and actual policy evidence
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve twelve-group sampled credit trajectories and actual policy evidence
-
Preserve twelve-group query and STOP training with zero-update transfer paths
-
Preserve three image-disjoint feedback groups eval
Yesterday
-
Preserve twelve-group actual decoder fit and fixed-history evidence
-
Preserve twelve-group actual decoder fit and fixed-history evidence
-
Preserve eight-group frozen visual transfer readout
-
Preserve eight-group frozen visual transfer eval
-
Preserve full-network credit and strong supervision evidence
-
Preserve full-network credit and strong supervision evidence
-
Preserve full-network parameter step matching and exact cumulative replay evidence
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve full-network matched credit and strong supervision evidence
-
Preserve matched frozen policy coordinate conditioning evidence
-
Preserve fixed-decoder matched encoder trainability experiment
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve frozen-feature decoder-policy results
-
Preserve task-reward training controls and fixed-history evaluation
-
Preserve shared-decoder feasibility search and replay evidence
-
Preserve observation-conditioned ceilings and signed decoding regret
-
Preserve direct-decision matching evidence and LP verification
-
Preserve fixed-history task-supervised decoding evidence
-
Preserve shared answer-channel training and frozen-policy evidence
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve frozen-feature soft-warm results
-
Preserve frozen-feature soft-learning results
-
Preserve frozen soft-feedback diagnostic and hard-history controls
-
Preserve frozen-feature blind-refit results
-
Preserve complete blinded visual panel eval
-
Preserve single geometric feedback intervention and frozen-policy paths
-
Preserve matched actual-step credit controls
-
Preserve frozen root-tail crossover paths and paired task intervals
-
Preserve frozen-feature rank-rl results
-
Preserve frozen-feature value-gap ranking results
-
Preserve matched signed immediate and full value regression records
-
Preserve matched signed immediate and full value regression records
-
Preserve endpoint regret and single-step gradient diagnostic
-
Preserve visual training-reward optimization and matched controls
-
Preserve fixed-order stopping reference separately from learned visual policies
-
Preserve visual training-reward optimization and matched controls
-
Preserve visual training-reward optimization and reward-matched controls
-
Preserve learned visual policy evidence with model and routed scores separated
-
Preserve exact-answer-tie removal sensitivity without rewriting primary visual results
-
Preserve frozen visual questions, exact cost ledger and ternary planning replay
-
Preserve five-target candidate-query structural diagnostic
-
Preserve exact multi-step reward optimization and matched continuation controls
