- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies failed adaptation in compact policy representations and proposes using a successful adaptation as a behavioral witness for repair. The witness supplies action targets for fixed-capacity fitting and a parameter direction for expansion. A hierarchy tests stronger coordinate search, witness-informed fitting, decoder updates and joint paths, then aligned expansion, selecting the weakest tested intervention that meets target-performance and archive-preservation criteria. Experiments include controlled failure constructions, natural HalfCheetah reward variants, and SmolVLA/LIBERO adaptation. The MuJoCo experiments demonstrate different sufficient repairs, while the VLA experiments evaluate witness-aligned directions rather than the full hierarchy.
Critical strengths
1. Useful problem formulation: low return does not distinguish unsuccessful coordinate search from an inadequate policy representation, so expansion should not be the default response. The paper appropriately limits its conclusions to sufficient repairs within a declared intervention family and budget. The joint-only case in Table 7 also shows why testing individual interventions alone can miss a viable fixed-capacity repair.
2. Meaningful controls for witness information: archive-only refitting matches the structure and training budget of witness-conditioned fitting, while random expansion controls match direction norm and, in VLA, layer-wise rank and spectrum. On Random-20, aligned residuals achieve 69/200 successes versus 0/200 for each of three random controls (Table 23). This supports the utility of task-aligned evidence beyond additional capacity alone.
Critical weaknesses
1. The main practical claim is not convincingly established. In the original prospective panel, hierarchical selection and always-expand both pass 12/12 cases, but the hierarchy avoids only one expansion while requiring approximately 4.9 times as many interactions (Table 11). The later targeted panel resolves 10/12 cases without expansion, but consists of two rewards selected after development inspection. VLA does not evaluate the full hierarchy. Its efficiency comparison covers only 50 continuation steps after a shared 5,000-step candidate, and Aligned is slightly more memory-intensive and slower than Direct LoRA, with five additional successes out of 120 episodes (Tables 31–32). These results establish examples of successful repair, but not a compelling performance–cost–storage advantage for the proposed procedure.
Key controls and consistency checks remain incomplete. MuJoCo expansion uses the candidate-minus-base direction, so by construction; final confirmation coefficients remain close to one. Without fixed- witness replay or scalar-search-only controls, the contribution of the subsequent fitting/PPO procedure is unclear. In the untruncated VLA experiment, Aligned should equal Direct algebraically, yet seven paired rollout outcomes differ (Table 24). Loaded-weight and same-input action checks remain outstanding. The manuscript acknowledges both issues, but disclosure does not resolve them: the former limits the interpretation of expansion results, and the latter leaves part of the VLA validation unreliable.
Questions for authors
-
What advantage does the hierarchy offer over always-expand and direct witness reuse when candidate acquisition, diagnosis, evaluation, repair, and persistent storage are all included? Can this be demonstrated on a prospectively fixed task panel?
-
How do fixed- witness replay and scalar-search-only expansion compare with the full repair procedure under the same target and archive criteria?
-
What explains the Direct–Aligned discrepancies in Table 24? Please verify loaded weights, adapter scaling, same-input actions, and rollout agreement under matched inference randomness.
This paper decomposes adaptation failures in compact policy representations into representation loss and exploration error, and proposes using a verified successful adaptation (here, behavioral witness) to identify the minimum sufficient repair by testing weaker interventions first. The problem formulation is useful: low return alone cannot distinguish the causes of failure. However, as currently written, the paper makes it difficult to understand which hypothesis each experiment is intended to support. Although the questions about the discrepancies in Table 24 and the replay controls have been resolved, the ablations are insufficient to establish whether the higher-level stage that uses witnesses is necessary. The presentation is also still not intuitive, making it difficult to justify an Accept recommendation.