- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ThoughtVLA, a framework for continual adaptation of pretrained vision-language-action models. It introduces Continually Evolving Thought Trajectories (CETT), implemented through Evolving Skill Reasoning (ESR) for pre-action skill identification and Complementary Knowledge Transfer (CKT) for reusing previous LoRA updates. The method separates skill selection from execution and attempts to prevent changes in expert routing from interfering with previously acquired skills. Evaluation covers multiple LIBERO streams, RoboTwin, humanoid manipulation, mixed embodiments, different VLA backbones, and real-robot experiments.
Critical strengths
-
The paper addresses a meaningful failure mode in continual VLA adaptation: previously learned skill policies can remain frozen while their effective behavior changes because of routing decisions. Separating skill identification from execution is a well-motivated design. The strong results across different embodiments and the real-robot evaluation make this contribution consequential.
-
The two principal components are supported by targeted ablations. On LIBERO-Long, ESR improves final success from 52.8% to 70.2% when conditional refinement is enabled, while CKT further improves it to 81.0%. The comparison with alternative knowledge-transfer criteria helps substantiate the proposed low-rank complementarity principle.
Critical weaknesses
W1. Limited conceptual novelty of Continually Evolving Thought Trajectories.
Although CETT is presented as an evolving reasoning mechanism, its implementation largely relies on frozen semantic embeddings, accumulated feature statistics, instruction matching, and predefined attribute-based score refinement. The reasoning rules themselves remain fixed throughout continual learning. Consequently, it is unclear whether CETT constitutes a substantially new reasoning paradigm or a structured, knowledge-augmented skill retrieval mechanism. A stronger comparison against prototype-based retrieval, discriminative skill classification, and attribute-aware routing would help isolate its conceptual contribution.
W2. Strong assumptions about task structure and skill availability.
The continual-learning protocol assumes known stage boundaries, explicitly grouped demonstrations, and an identifiable skill associated with each training task. At evaluation, the model selects among previously acquired skills rather than learning to recognize and compose arbitrary new behaviors. These assumptions substantially simplify continual adaptation compared with real-world scenarios involving unsegmented experience, overlapping skills, and evolving task definitions. The generality of the proposed framework under less structured task arrivals remains unclear.
W3. Limited support for within-episode skill switching and compositional execution.
ESR selects a skill before execution, and its corresponding LoRA composition is retained throughout the episode. This design may be insufficient for long-horizon tasks requiring different manipulation primitives at different execution stages. For example, a task involving opening a container, grasping an object, and performing a precise placement may benefit from dynamically switching between specialized skills. The paper does not establish whether episode-level skill selection can handle such compositional requirements as effectively as more flexible routing mechanisms.
W4. The additional contribution of the CKT criterion is relatively modest compared with a simpler residual baseline.
The CKT ablation on LIBERO-Long reports 81.0% final success, compared with 79.4% for full-residual transfer. Although the proposed rank-aware criterion is theoretically motivated, its incremental improvement over this simpler alternative is only 1.6 percentage points. The paper would benefit from uncertainty estimates, sensitivity to LoRA rank, and an analysis of when the rank-aware score selects a different source skill and whether those differences consistently improve execution.
W5. The theoretical approximation underlying CKT may not reflect actual nonlinear adaptation dynamics.
CKT estimates transfer complementarity using gradients and compressed activations evaluated at frozen base weights. Its low-rank approximation argument characterizes a local response rather than the full nonlinear optimization process. Moreover, the candidate-selection procedure validates only the highest-scoring proposal and does not systematically explore alternative candidates if confirmation fails. The robustness of this approximation across different adaptation regimes and increasing skill-library sizes remains insufficiently characterized.
W6. Limited evidence for scalability to substantially longer continual-learning streams.
The largest evaluated mixed-embodiment stream contains 55 skills. Both the retained knowledge and the LoRA composition library grow as additional skills are acquired. Although the reported storage and parameter overheads are favorable, the paper does not investigate substantially longer streams in which semantically overlapping skills accumulate. Such settings may introduce increasing ambiguity, retrieval errors, and interference between closely related skill representations.
Questions for authors
-
How much of ESR's improvement depends on its evolving knowledge representation rather than simply replacing learned expert routing with a strong frozen-encoder prototype or attribute-aware skill classifier under a matched information budget?
-
How robust is skill reasoning when task descriptions are ambiguous, contain contradictory attributes, or depart substantially from the training instruction distribution? Could the authors provide uncertainty estimates for the real-robot results?
-
See the above weaknesses.
4 comments
Response to Reviewer hSqq
W3: Within-episode skill switching and compositional execution.
A3: Thank you for the insightful comment. Our streams already contain such multi-stage tasks, each learned from full demonstrations. Table R11 selects them by a fixed rule: all ten LIBERO-Long tasks; GR-1 tasks 1–6, which place an item into a cabinet, drawer or microwave and close it; RoboTwin tasks whose instructions contain "then"; and the real-robot multi-step tasks 17–20.
Table R11. Final success in % on multi-stage task groups, pooling in-domain and out-of-domain rollouts for the real robot; the baseline is the strongest by overall FSR in the corresponding table.
| Task group | Tasks | ThoughtVLA | Strongest baseline | Baseline |
|---|---|---|---|---|
| LIBERO-Long, all tasks | 10 | 81.0 | SkillsCrafter, Table 1 | 67.6 |
| GR-1 tasks 1–6 | 6 | 40.8 | DMPEL, Table 3 | 28.0 |
| RoboTwin 11, 26, 33, 42, 45 | 5 | 48.4 | AtomicVLA, Table 3 | 25.0 |
| Real robot 17–20 | 4 | 23.8 | SkillsCrafter, Table 5 | 16.3 |
As shown in Table R11, ThoughtVLA outperforms the strongest baseline on every group, including AtomicVLA on RoboTwin, 48.4% vs. 25.0%, although AtomicVLA chains atomic skills within each episode. Since each skill is learned from full demonstrations and the action head responds to each observation, execution spans subtask boundaries within the selected policy. Switching between separately acquired skills within an episode is a natural extension, which we add to Appendix D. We have added Table R11 to Appendix C.4 as Table A18.
[1] Liu et al. Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning. arXiv:2603.03818, 2026. (https://arxiv.org/abs/2603.03818)
[4] Römer et al. CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion. RA-L 2026. (https://arxiv.org/abs/2601.09512)
[5] Wu et al. Continually Evolving Skill Knowledge in Vision Language Action Model. arXiv:2511.18085, 2026. (https://arxiv.org/abs/2511.18085)
W4: Uncertainty, rank sensitivity and source selection of CKT versus full-residual transfer.
A4: Thank you for this careful comment. As described in Appendix C.6, the two strategies differ only in the ranking score and the Δr > 0 gate.
Uncertainty. Table R12 adds training seeds 43 and 44 for full-residual transfer; all runs share the 500 initial states of evaluation seed 7, so the comparison is paired.
Table R12. Paired LIBERO-Long comparison at rank 16, with published values from Tables 8 and A15.
| Method | Seed 42 | Seed 43 | Seed 44 | Mean |
|---|---|---|---|---|
| Full residual FSR (%) | 79.4 | 78.6 | 78.8 | 78.9 |
| Rank-aware CKT FSR (%) | 81.0 | 81.0 | 80.6 | 80.9 |
| Paired difference (pp) | +1.6 | +2.4 | +1.8 | +1.9 |
As shown in Table R12, CKT outperforms full residual under every seed, by 1.9 pp on average, more than the seed-to-seed spread of either method, 0.4 and 0.8 pp.
When the criteria select different sources. Full residual ranks candidates by the total error after transfer, ‖R‖²_F, while Δr uses ε_r(R), the error left after the best rank-r new update. CKT can thus prefer a larger but low-rank residual, which the new update can absorb, over a smaller but spread-out one: in Eq. A8 of Appendix A, full residual prefers B since 4.5 < 25.25, but CKT prefers A since ε₁ = 0.25 < 2.25. CKT also proposes transfer only when the maximum Δr is positive, whereas full residual always proposes its top candidate.
Whether the differences improve execution. They improve it consistently: besides leading under every seed in Table R12, CKT raises CSR, which averages performance over all stages, by 2.2 pp in Table 8 and lowers forgetting from 1.7 to 1.4 pp.
Rank sensitivity. Full residual is the r = 0 case of Δr, since ε₀(R) = ‖R‖²_F by Appendix A, so the two ranking scores differ only through the rank budget. Table R13 repeats the Table 8 comparison with new-update ranks 8 and 32, sharing the stage-1 checkpoint and all other settings; Δr uses the adapter's rank and the scaling factor is 2r.
Table R13. LIBERO-Long results in % for different new-update ranks; rank 16 is from Table 8.
| Rank | Method | FSR ↑ | CSR ↑ |
|---|---|---|---|
| 8 | Full residual | 69.4 | 70.9 |
| 8 | Rank-aware CKT | 72.8 | 73.7 |
| 16 | Full residual | 79.4 | 81.2 |
| 16 | Rank-aware CKT | 81.0 | 83.4 |
| 32 | Full residual | 80.0 | 81.8 |
| 32 | Rank-aware CKT | 81.4 | 83.5 |
As shown in Table R13, CKT improves on full residual at every rank, by 3.4, 1.6 and 1.4 pp in FSR and 2.8, 2.2 and 1.7 pp in CSR at ranks 8, 16 and 32, so the advantage is not specific to rank 16. The gain is largest at rank 8, where the rank budget is tightest, and narrows as a higher-rank update absorbs more of the residual itself. We have added Tables R12 and R13 to Appendix C.8 as Tables A25 and A26.
Response to Reviewer hSqq
W5: Local approximation versus nonlinear adaptation, and rejected proposals.
A5: Thank you for this insightful comment. We agree that the low-rank analysis characterizes the local response at the frozen base weights, not the full nonlinear optimization, so CKT uses it only to rank candidates and lets actual adaptation decide.
First, Δr ranks all candidates from gradients and compressed activations without training an adapter per candidate, keeping scoring affordable as Table R2 shows.
Second, the top-ranked proposal is tested by paired adaptation trials, which run the actual nonlinear fine-tuning with and without transfer, and is adopted only if the held-out action-loss reduction meets the threshold and more than half of the trajectories improve, following Eq. A11. We have added this clarification to Appendix A.1 at l. 837–840.
W6: Scalability to substantially longer streams with overlapping skills.
A6: Thank you for this constructive comment.
First, on stream length, our RoboTwin stream learns 50 bimanual tasks and our mixed-embodiment stream 55 single-arm, bimanual and humanoid skills over 51 stages, as listed in Table A1; for reference, the longest streams in Liu et al. [1], Stellar VLA [5] and CLARE [4] contain 10, 30 and 40 single-arm LIBERO tasks. On the 55-skill stream, ThoughtVLA reaches 67.3% FSR with 3.4 pp forgetting, versus 51.5% and 5.7 pp for AtomicVLA, the strongest baseline in Table 3. Memory also stays compact: unlike the replay of [1] and [5], ThoughtVLA stores no raw trajectories, and it has the lowest parameter growth and storage in Table A11, 4.8M/4.9M per stage and 53.0/53.5 MB. We agree that even longer streams merit further study, as noted in Appendix D.
Second, for ambiguity among closely related skills, the stress test in Table R3, repeated in Q2 below, evaluates ESR after all 50 bimanual skills are learned, under unseen instructions and their edits, including confusable pairs such as stacking three vs. two blocks in tasks 45/46. Moreover, previous LoRA updates are frozen and each skill keeps its own composition, so new skills do not overwrite earlier ones.
Q1: Contribution of evolving knowledge under a matched information budget.
A1: Thank you for the question. We assess the contribution of evolving knowledge from two angles.
First, the same information with and without evolution in Table 6. Base + SR (static) and Base + SR share the same records, demonstrations, LoRA updates and reasoning rules; they differ only in whether the shared statistics S_s are recalibrated as skills accumulate, which isolates the evolving part of the knowledge.
Table 6, excerpt. LIBERO results without oracle IDs; FSR and CSR are in % and forgetting in pp.
| Variant | Goal FSR ↑ | Goal CSR ↑ | Goal Forg. ↓ | Long FSR ↑ | Long CSR ↑ | Long Forg. ↓ |
|---|---|---|---|---|---|---|
| Base + SR (static) | 75.4 | 78.9 | 18.9 | 47.6 | 50.5 | 22.4 |
| Base + SR | 80.2 | 83.8 | 12.5 | 52.8 | 58.1 | 18.7 |
Recalibrating S_s improves FSR by 4.8 and 5.2 points and reduces forgetting by 6.4 and 3.7 points.
Second, alternative selectors on the same policies in Table R9. All selectors share the frozen encoder and CKT-trained policies, and the ESR ablations remove only attribute refinement or its gating.
Table R9, repeated from A1. RoboTwin, Table 7 protocol; FSR is closed-loop.
| Category | Selector | Evidence | SR acc. (%) | FSR (%) |
|---|---|---|---|---|
| Prototype-based retrieval | Semantic prototype c0 | per-skill mean VLM features | 58.2 | 28.5 |
| Prototype-based retrieval | HiDe-LLaVA [2] | per-skill image and text anchors | 64.2 | 36.1 |
| Discriminative classification | RanPAC [3] | random projection, ridge classifier | 54.7 | 26.4 |
| ESR ablation | ESR w/o attributes c1 | no attribute refinement | 82.3 | 44.2 |
| ESR ablation | ESR w/o gating | refinement always applied | 80.5 | 41.3 |
| Ours | Full ESR | full evidence, updated statistics | 91.0 | 52.8 |
As shown in Table R9, ESR leads prototype retrieval and discriminative classification on the same policies by at least 26.8 pp in SR acc. and 16.7 pp in FSR. Removing attribute refinement or its gating lowers SR acc. to 82.3% and 80.5%, so the gain comes from conditional reasoning over attributes rather than attribute evidence alone.
Response to Reviewer hSqq
Q2: Ambiguous, conflicting or out-of-distribution instructions, and real-robot uncertainty.
A2: Thank you for the question. For ambiguous and out-of-distribution instructions, we repeat the RoboTwin stress test from our response to Reviewer weVY on W3.
First, attribute refinement is conditional and falls back safely: it applies only when the initial correction was triggered, the two leading scores are within a margin and attribute evidence separates them, following Eq. 3 and Appendix B.2; otherwise, including when no supported attribute is parsed, the scores after instruction matching are kept. Contradictory attributes are averaged in the attribute agreement of Eq. A10, so if they no longer separate the leading candidates, the gate stays closed and the same fallback applies.
Second, we freeze the final RoboTwin checkpoint and test all 50 tasks with RoboTwin's own unseen instruction R0, five single edits of it and one random pair of edits, keeping scene, target skill and parsed attribute set fixed, so every G1–G6 instruction is an unseen variant of R0.
- R0: RoboTwin's unseen evaluation instruction for each episode, unchanged.
- G1, verb paraphrase: one verb in R0 is replaced by a synonym from a fixed lexicon, such as "rotate" → "spin".
- G2, object alias: one object is renamed with another of RoboTwin's unseen descriptions of the same asset.
- G3, syntactic restructuring: R0 is reordered by fixed rewriting rules without adding or removing content words.
- G4, irrelevant background: the fixed sentence "The tabletop is visible." is prepended to R0.
- G5, out-of-schema attribute: one object receives a real color, material or shape adjective from its asset descriptions, unseen in that task's training instructions and not read by the parser.
- G6, combined edits: two edits from different categories in G1–G5 are applied together.
Table R3, repeated from our response to Reviewer weVY. Skill reasoning accuracy in % on all 50 RoboTwin tasks after the final stage; all selectors use the same policies and instructions.
| ID | Instruction condition | Semantic prototype c0 | HiDe-LLaVA | ESR |
|---|---|---|---|---|
| R0 | RoboTwin unseen instruction | 58.2 | 64.2 | 91.0 |
| G1 | Verb paraphrase | 68.6 | 71.4 | 89.7 |
| G2 | Object alias | 71.9 | 67.0 | 90.8 |
| G3 | Syntactic restructuring | 62.4 | 68.0 | 90.4 |
| G4 | Irrelevant background | 59.8 | 67.9 | 88.2 |
| G5 | Out-of-schema attribute | 79.1 | 65.4 | 86.3 |
| G6 | Combined edits, two of G1–G5 | 53.1 | 57.5 | 87.6 |
As shown in Table R3, ESR reaches 91.0% on RoboTwin's unseen instructions, stays at 86.3–90.8% under every single edit, including out-of-schema attributes, and at 87.6% under combined edits. It leads both alternatives under every condition by 7.2–30.1 pp over the stronger one, with the largest margin under combined edits, where they fall to 53.1% and 57.5%.
For real-robot uncertainty, we evaluate the final policy two more times under the in-domain and out-of-domain protocols of Table 5, with 10 rollouts per task per run.
Table R14. ThoughtVLA real-robot FSR in % over three evaluation runs.
| Condition | Run 1, Table 5 | Run 2 | Run 3 | Mean ± std |
|---|---|---|---|---|
| In-domain | 53.0 | 49.5 | 52.5 | 51.7 ± 1.9 |
| Out-of-domain | 40.5 | 43.0 | 42.5 | 42.0 ± 1.3 |
As shown in Table R14, FSR is 51.7 ± 1.9% in-domain and 42.0 ± 1.3% out-of-domain. Even the lowest runs, 49.5% and 40.5%, stay above the strongest baselines in Table 5, SkillsCrafter and AtomicVLA, both at 43.5% in-domain and 31.0% out-of-domain. We have added Table R14 to Appendix C.5 as Table A19.
Response to Reviewer hSqq
A1: Thank you for this thoughtful question. As defined in Sec. 1 at l. 72–82, thought in CETT is a structured pre-action process that interprets the task against acquired skills and refines ambiguous interpretations before identifying a skill. It goes beyond skill retrieval in three ways:
Table R9 compares ESR with prototype retrieval and discriminative classification on the same CKT-trained policies and frozen encoder; its first row, ESR's own semantic score, is plain prototype retrieval.
Table R9. RoboTwin, Table 7 protocol: SR acc. on RoboTwin's unseen evaluation instructions for all 50 tasks after the final stage; FSR is closed-loop.
As shown in Table R9, ESR leads prototype retrieval and discriminative classification on the same policies by at least 26.8 pp in SR acc. and 16.7 pp in FSR. Removing attribute refinement or its gating lowers SR acc. to 82.3% and 80.5%, so the gain comes from conditional reasoning over attributes rather than attribute evidence alone. We have added Table R9 to Table 7 and to Appendix C.7 as Table A21.
[2] Guo et al. HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model. ACL 2025. (https://arxiv.org/abs/2503.12941)
[3] McDonnell et al. RanPAC: Random Projections and Pre-trained Models for Continual Learning. NeurIPS 2023. (https://arxiv.org/abs/2307.02251)
A2: Thank you for this question.
Our setting. ThoughtVLA learns tasks over a sequence of stages from current-stage demonstrations only and keeps no raw trajectories. After each stage, it is evaluated on all tasks seen so far without oracle skill or stage IDs, so it must identify the required skill itself before acting.
Relation to prior work. Our protocol follows recent continual VLA work such as Liu et al. [1], CLARE [4] and Stellar VLA [5], which evaluate mainly on LIBERO; our validation extends it in three respects:
Less structured arrivals. We rerun LIBERO-Long with irregular stages of 3, 1, 4 and 2 tasks and no task labels within a stage: a grouping step clusters each stage's demonstrations into skills with ESR's instruction-matching features, while stage boundaries remain given. All methods use 110K formal steps; SkillsCrafter and DMPEL are the two strongest baselines on LIBERO-Long in Table 1.
Table R10. LIBERO-Long with irregular stages. Stage columns give final success in % on the tasks introduced at that stage; FSR, CSR and Forg. follow Table 1.
As shown in Table R10, ThoughtVLA reaches 80.6% FSR without task labels, within 0.4 pp of standard stages, with 1.3 pp forgetting, and leads SkillsCrafter and DMPEL at every stage and by 14.8 and 21.0 pp in FSR, while their forgetting reaches 14.4 and 17.3 pp. We have added this experiment to Appendix C.4 with Table A17.