acceptodds
Under review as a conference paper at ICLR 2027

R²-WAM: Functional Reuse and Refinement of Historical Adaptations for World-Action Models

Abstract

World-Action Models (WAMs) offer a promising foundation for robot manipulation, yet existing adaptation methods typically learn each new task independently from the same base policy. We study a history-aware adaptation setting where task-specific adaptations accumulated during deployment are reused to accelerate future learning. We propose R²-WAM, which organizes historical adaptations into Video, Context, and Action functional routes, selects compact historical support using target-loss signals, and jointly refines reused residuals with new task-specific capacity. Experiments on LIBERO and RoboCasa365 show that the proposed function-structured adapter is data-efficient under limited demonstrations. In our history-aware LIBERO evaluation, historical reuse improves overall success from 57.9% with fresh structured adaptation to 70.4%, surpassing Full Fine-Tuning at 61.9%. Further analysis shows that effective reuse depends on task-group structure, support organization, and controlled refinement.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World-Action Models, Robot Adaptation, Historical Parameter Reuse, Few-Shot Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_q6ogOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces R$^2$-WAM, which is built on a function-structured adapter for world-action models (WAMs) that enables reusable experience for new few-shot tasks. It ranks historical residuals route-by-route with first-order target-loss scores, learns task-specific compositions, and adds a group-shared new residual with task-specific exposure for missing capabilities. R$^2$-WAM requires only saved adapters, without needing historical data. Experimental results on LIBERO and RoboCasa365 demonstrate that R$^2$-WAM achieves improved success rates compared to fresh adapters, pooled-data initialization, and random historical support.

Critical strengths
  1. The paper addresses an important problem in few-shot robot policy adaptation by exploring how previously learned adapters can be reused to facilitate adaptation to new tasks without retaining historical training data.

  2. The experiments on LIBERO and RoboCasa365 demonstrate the potential of the proposed approach to improve few-shot adaptation performance.

Critical weaknesses

1. Insufficient comparisons and experimental evidence.

The paper lacks direct comparisons with existing approaches for historical adapter reuse, composition, and continual learning. The experiments do not sufficiently establish the advantages of R$^2$-WAM over existing methods under matched training conditions. Furthermore, the ablation results do not fully clarify the contributions of the proposed components. For example, the new-residual-only variant performs substantially worse than fresh structured adaptation, making it difficult to isolate the benefits of historical reuse from differences in parameter sharing and optimization.

2. Limited validation of the claimed long-term experience reuse and accumulation.

The main experiments rely on a fixed, manually constructed historical adapter bank, without evaluating how the framework performs as new adapters are progressively added. Although Section 6 discusses experience persistence, folding, and bank expansion, these mechanisms are not validated through corresponding closed-loop experiments. Moreover, the evaluation relies on manually organized task groups with shared operation structures, leaving the applicability to more diverse and less structured task distributions unclear. Consequently, the experiments support few-shot adaptation from a fixed historical bank but do not fully substantiate the broader claims of continual experience accumulation and reuse.

3. Unclear presentation.

The paper is difficult to follow due to unclear explanations, excessive technical details, and an overly complicated methodological narrative. In particular, the relationships between the function-structured adapter, historical residual composition, and target-group refinement are not clearly motivated, making it difficult to distinguish the core technical contributions from implementation details. The discussion of related work is also insufficient. Although several relevant methods are mentioned, their differences from the proposed framework are not adequately explained. These presentation issues make it difficult to assess the novelty and significance of the proposed contributions.

Questions for authors
  1. How does R$^2$-WAM compare with existing methods for historical adapter reuse and composition, such as LoRAHub and continual-learning approaches?

  2. How sensitive is R$^2$-WAM to the size, diversity, and quality of the historical adapter bank? Does performance consistently improve as more historical adapters become available, or can irrelevant adapters introduce negative transfer?

  3. Why does the new-residual-only variant substantially underperform fresh structured adaptation?

10 comments

We thank the reviewer for recognizing the importance of historical experience reuse in few-shot robot adaptation. We provide additional comparisons, examine historical-bank sensitivity, and clarify the method's technical contributions and evaluation scope. For your convenience, please also refer to our revised manuscript for the corresponding updates.

 

A1 (W1, Q1): Motivation, problem formulation, and comparison with prior methods (Part 1)

Motivation and adaptation-space design. Robot demonstrations are costly to collect, making few-shot adaptation of large WAMs challenging. Rather than incorporating all historical data into a single embodied pretraining stage, we organize accumulated experience by task groups and preserve learned capabilities as reusable adapters. New-task adaptation can then be formulated as recombining historical capabilities and learning corrections for missing ones.

This motivates a parameter-efficient, function-structured adaptation space. Unlike LoRA's low-rank weight updates, our adapter aligns nonlinear residual modules with WAM's video, context, and action routes to enhance expressivity. Since nonlinear residuals cannot generally be composed by directly merging their parameters, we combine their outputs in representation space. The resulting adapter achieves 97.2% versus 92.0% for LoRA on LIBERO-Object with 80 demonstrations (Section 5.2).

Three perspectives on historical experience. We distinguish three related optimization problems:

DimensionLoRAHubR²-WAM (Ours)Continual Learning
Problem settingHistorical CompositionHistory-Aware AdaptationSequential Model Evolution
Task settingTarget task, frozen base and bankTarget group, frozen anchor and bankSequential tasks, evolving system
Optimized variablesComposition weights wtw_tαt,i,m,ηt,ϕm\alpha_{t,i,m},\eta_t,\phi_mSystem state Sk\mathcal S_k
Learning objectiveCompose historical modulesReuse and learn missing capabilitiesLearn new and retain prior capabilities

LoRAHub — Historical Composition. Given a frozen base model and historical LoRA bank B\mathcal B, LoRAHub optimizes target-specific composition coefficients:

wt∗=arg⁡min⁡wtL(Dt;fbase,B,wt).w_t^*=\arg\min_{w_t}\mathcal L(D_t;f_{\mathrm{base}},\mathcal B,w_t).

R²-WAM — History-Aware Adaptation. We formulate history-aware target-group adaptation over a function-structured residual space, jointly optimizing task-specific historical compositions and group-shared learnable corrections:

(α∗,η∗,ϕ∗)=arg⁡min⁡α,η,ϕ1M∑t=1ML(Dt;fanchor,B,αt,ηt,ϕ).(\alpha^{\ast},\eta^{\ast},\phi^{\ast})=\arg\min_{\alpha,\eta,\phi}\frac{1}{M}\sum_{t=1}^{M}\mathcal L(D_t;f_{\mathrm{anchor}},\mathcal B,\alpha_t,\eta_t,\phi).

Here, αt,i,m\alpha_{t,i,m} learns task- and route-specific historical combinations, while the group-shared ϕ\phi learns missing capabilities with task-specific exposure ηt\eta_t.

Continual Learning — Sequential Model Evolution. Given a task stream D1→D2→⋯→DKD_1\rightarrow D_2\rightarrow\cdots\rightarrow D_K, continual learning updates a persistent system from Sk−1\mathcal S_{k-1} to Sk\mathcal S_k. A representative objective is Sk∗=arg⁡min⁡Sk[L(Dk;Sk)+λRretain(Sk,Sk−1)].\mathcal S_k^*=\arg\min_{\mathcal S_k}\left[\mathcal L(D_k;\mathcal S_k)+\lambda\mathcal R_{\mathrm{retain}}(\mathcal S_k,\mathcal S_{k-1})\right].

The two terms represent new-task learning and preservation of prior capabilities. Historical retention can be enforced through replay or frozen modules rather than an explicit regularization loss. DMPEL and CLARE study such evolving modular systems through expert routing and expansion.

Key distinction and motivation. LoRAHub and R²-WAM share a fixed-bank, target-given adaptation setting that reuses saved modules without accessing historical demonstrations during target adaptation. Their key difference lies in the optimization space. LoRAHub composes frozen historical modules, whereas R²-WAM jointly optimizes function-structured historical reuse and new learnable corrections to address unseen objects, semantics, and interactions that existing capabilities may not fully cover. Continual learning instead addresses sequential system evolution, balancing new-task acquisition with historical capability retention.

We study fixed-bank adaptation to isolate historical reuse from task-arrival order and old-task retention. Historical adapters are independently trained from a common anchor and treated equally during target-specific selection, regardless of acquisition order. This separates experience reuse from sequential bank maintenance and expansion, discussed in Section 6.

A1 (W1, Q1): Motivation, problem formulation, and comparison with prior methods (Part 2)

Matched composition comparisons. All methods use a Top-4 candidate budget per route, with the same anchor, historical bank, target demonstrations, and Stage-2 refinement. Uniform and NGOpt share R²-WAM's route-wise candidates, while WQ uses a single whole-adapter candidate set shared across routes.

MethodOverall SR (35 tasks)
Uniform60.6%
NGOpt (LoRAHub-style)64.5%
Whole-adapter WQ61.4%
R²-WAM (GSC)72.2%

Uniform fixes equal historical coefficients, NGOpt performs gradient-free coefficient optimization, and whole-adapter WQ shares historical candidates and coefficients across functional routes. R²-WAM outperforms these baselines by +11.6 pp, +7.7 pp, and +10.8 pp, respectively. Its advantage over whole-adapter WQ holds across all six target groups, supporting task-specific coefficient optimization and function-structured historical reuse.

A2 (W1, Q3): Component contributions and the residual-only performance gap

Historical selection, composition, and refinement. We distinguish the contributions of three key components:

  • Historical reuse. The original trained-bank mean is +15.5 pp above the single-seed output-scale-matched random-bank result in task-weighted Overall SR (72.2% vs. 56.7%). This cross-protocol comparison is descriptive. Target-loss candidate selection improves success over random trained candidates by +6.5 pp (72.2% vs. 65.7%) in the original paper's mean results (Table 1).
  • Task-specific composition. Table 2(a) shows that making composition coefficients task-specific, while retaining group-shared candidates and task-specific exposure, raises success from 58.7% to 72.2% (+13.5 pp), demonstrating the value of personalized historical combinations.
  • New-residual refinement. Unlike prior NLP work where frozen-adapter composition alone enables effective few-shot transfer, bank-only composition achieves just 0.9% on our embodied targets (Table 1), highlighting the importance of learning capabilities absent from the historical bank.

Why new-residual-only underperforms fresh adaptation. Both methods train the same group-shared structured residuals, but new residual only retains the full method's task-specific learnable exposure ηt\eta_t, initialized at 0.1, after removing history. Fresh adaptation instead fixes η=1\eta=1. We compare the two exposure settings with and without history in Appendix D.5:

Target groupNo history, learned η\etaNo history, η=1\eta=1 (Fresh)With history, learned η\etaWith history, η=1\eta=1
Object59.0%72.6%85.6%65.4%
Goal27.7%75.0%76.7%85.0%
Long-C8.0%5.0%30.0%5.0%

Without history, fixing η=1\eta=1 recovers fresh adaptation and improves success on Object and Goal. This shows that their residual-only deficits depend on the exposure setting, rather than a different residual architecture. Object reverses this preference with history, favoring learned exposure over η=1\eta=1. Long-C also favors learned exposure with history, while full exposure alone does not overcome its low no-history success. Goal instead performs best with full exposure even when history is present. Thus, the preferred exposure setting depends on both historical support and the target group.

We therefore interpret new residual only as a history-removal ablation under the full method's exposure setting, and use fresh adaptation as the standalone baseline. The gain over new-residual-only should not be read as the gain over standalone adaptation. Section 5.3 now makes this distinction explicit.

Together, these experiments distinguish the contributions of historical candidate selection, task-specific composition, and exposure-controlled refinement.

Reviewer_32uYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

R2-WAM adapts FastWAM to new tasks by reusing a bank of saved adapters. It selects historical components separately for the video, context, and action routes using target-loss gradients, then learns task-specific mixtures and a new residual shared across the target group. With three demonstrations per task, it reaches 72.2% success on 35 LIBERO tasks, compared with 57.9% for a fresh structured adapter and 66.1% for pooled initialization. All experiments, including those on RoboCasa365, are in simulation.

Critical strengths
  • Reusing different parts of past adapters makes sense for WAMs: the best source for visual features need not be the best source for action generation. The new residual lets the model learn what the bank cannot provide, and reuse does not require keeping the original historical trajectories.

  • The experiments include useful controls. Target-loss selection improves success from 65.7% with random support to 72.2%, while pooled initialization reaches 66.1%.

Critical weaknesses
  • My main concern is the lack of real-robot experiments. The paper is motivated by robots reusing experience accumulated during deployment, but all results are from LIBERO and RoboCasa365. I would like to see whether the gain over fresh adaptation holds on a physical robot before drawing conclusions about its deployment value.

  • The comparison with target-loss-based whole-adapter composition only covers four Long-B tasks (Section 5.4.2). Without this baseline across the 35-task evaluation, using the same refinement and optimization budget, it is hard to tell how much of the overall gain comes from routing the three components separately.

Questions for authors
  • Are any real-robot results available? A comparison with fresh structured adaptation under the same demonstration budget would be particularly useful.

  • Does the advantage over whole-adapter composition hold across all 35 tasks with matched support and optimization budgets and the same refinement procedure?

  • Which formula and initialization were used for η\eta? Section 4.3 defines ηt=0.1+0.9σ(st)\eta_t = 0.1 + 0.9\sigma(s_t), but Section 5.3 reports values as low as 0.06. Could differences in residual scaling or optimization also explain why “new residual only” scores 35.5%, compared with 57.9% for fresh structured adaptation?

5 comments

We thank the reviewer for recognizing function-separated historical reuse and the matched controls. We address real-robot validation, extend the whole-adapter comparison to all 35 tasks, and clarify the exposure parameterization and its effect on refinement. For your convenience, please also refer to our revised manuscript for the corresponding updates.

 

A1 (W1, Q1): Real-robot validation

Setup and historical bank. We add real-robot experiments on NERO with head, top, and wrist RGB observations (Section 5.5; Appendix A.2). Both methods share a frozen FastWAM anchor trained on 57,732 DROID trajectories using DreamZero preprocessing. We independently train four historical adapters with 75 demonstrations each: H1, pen into holder; H2, black box onto pink label; H3, tape into box; and H4, remote control onto box. These form a frozen bank for the held-out target, placing tape on a box (Figure 5: task layouts).

Matched target-data protocol. We compare fresh structured adaptation and R²-WAM (GSC Top-2) on identical, nested training sets of 75, 30, and 20 target demonstrations. GSC uses the same fixed three demonstrations for route-wise ranking and Stage-1 composition fitting. These are included in every training budget, and Stage 2 uses the full corresponding training set. Three additional demonstrations are reserved for offline validation and excluded from optimization. Both methods train for 10 epochs at the 75-demonstration budget and 15 epochs at the 30- and 20-demonstration budgets, using global batch size 64 and equally weighted action and video losses. Each method is evaluated in 10 physical closed-loop trials per budget.

Target training demonstrationsFresh structured adaptationR²-WAM (GSC Top-2)Gain (pp)
7510/10 (100%)10/10 (100%)0
300/10 (0%)5/10 (50%)+50
201/10 (10%)3/10 (30%)+20

Historical reuse helps under limited target data. Both methods succeed in all 10 trials with 75 demonstrations, while GSC improves success by 50 pp at 30 demonstrations and 20 pp at 20 demonstrations (Figure 4). In low-data trials, fresh adaptation often repeatedly descends without completing the grasp-and-place sequence. GSC more consistently follows the intended trajectory, with remaining failures involving imprecise grasps. The recorded 30-demonstration trials show a Fresh failure and a GSC success. These observations suggest that historical reuse supports coarse motion structure, while grasp precision remains a limitation on this target.

Selected historical support. The route-wise Top-2 candidates for tape-on-box are shown below in ranking order (Appendix D.7, Table 28).

RouteFirst candidateSecond candidate
Video (V)H4: Remote control onto boxH3: Tape into box
Context (C)H3: Tape into boxH2: Black box onto pink label
Action (A)H3: Tape into boxH4: Remote control onto box

H3 shares the target object, while H4 shares placement onto a box, consistent with combining related experience across functional routes. Recorded historical-task examples are available for H2: black box onto pink label, H3: tape into box, and H4: remote control onto box. H1 did not successfully complete its own task after training. It remains in the bank but is not selected on any route. As a support-size check at the 75-demonstration budget, using all 75 demonstrations for ranking and Stage-1 fitting yields the same candidate selections and ordering as the fixed-three protocol, and both achieve 10/10 successes. These experiments provide real-robot evidence for historical reuse on one held-out target under matched target-data budgets.

A2 (W2, Q2): Full-benchmark comparison with whole-adapter composition

Matched comparison across all 35 tasks. We extend the whole-adapter q-ranked (WQ) baseline from Section 5.4.2 to all 35 held-out tasks. Both use the same anchor, historical bank, three demonstrations per task, Top-4 candidate budget per route, coefficient optimization, and Stage-2 refinement. WQ shares historical candidates and coefficients across video, context, and action routes, whereas R²-WAM selects and composes them separately.

MethodSpatialObjectGoalLong-ALong-BLong-COverall (35 tasks)
Whole-adapter WQ Top-462.8%75.4%67.0%49.3%47.0%15.0%61.4%
R²-WAM (GSC Top-4)78.0%85.6%76.7%56.7%50.0%30.0%72.2%
Gain (pp)+15.2+10.2+9.7+7.4+3.0+15.0+10.8

Benefit of functional granularity. The +10.8 pp task-weighted Overall gain is accompanied by higher SR in all six groups, extending the Long-B finding to the full benchmark and demonstrating the benefit of independently selecting and combining historical capabilities across functional routes.

A3 (Q3): Exposure parameterization and residual-only adaptation

Correct exposure parameterization. We apologize for the notation error in Section 4.3, where ηinit\eta_{\mathrm{init}} was mistakenly used in place of ηmin⁡\eta_{\min}. The implemented parameterization is

ηt=ηmin⁡+(1−ηmin⁡)σ(st),ηmin⁡=0.05,ηt(0)=0.1.\eta_t=\eta_{\min}+(1-\eta_{\min})\sigma(s_t),\qquad \eta_{\min}=0.05,\quad \eta_t^{(0)}=0.1.

We initialize st(0)=log⁡[(0.1−0.05)/(1−0.1)]≈−2.8904s_t^{(0)}=\log[(0.1-0.05)/(1-0.1)]\approx-2.8904, yielding 0.05<ηt<10.05<\eta_t<1, consistent with the reported values of 0.06–0.10. We will correct Section 4.3 accordingly.

Why new-residual-only underperforms fresh adaptation. Both methods train the same group-shared structured residuals. However, new residual only removes history while retaining the full method's task-specific learnable exposure ηt\eta_t, initialized at 0.1, whereas fresh adaptation fixes η=1\eta=1. The former is therefore a history-removal ablation under the full method's exposure setting, while the latter is the standalone adaptation baseline.

Without history, fixing η=1\eta=1 raises success from 59.0% to 72.6% on Object and from 27.7% to 75.0% on Goal, recovering fresh adaptation. These paired controls identify exposure settings as a contributing factor to the gaps in these groups, rather than a difference in residual architecture. Long-C changes from 8.0% to 5.0%, showing that the effect is group-dependent. With history, Object reverses its preference: learned exposure reaches 85.6%, versus 65.4% with η=1\eta=1. Thus, an exposure setting suitable for standalone learning need not suit refinement on top of historical support. We report the full three-group controls in Appendix D.5 and clarify this distinction in Section 5.3, without attributing the entire overall gap to a single optimization mechanism.

Reviewer_KAY2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Within the framework of World-Action Models, this paper addresses the problem of transferring adapters trained on previous tasks to adapt to new tasks with limited demonstrations, without access to the original training data. The authors constructed a structured adapter on a frozen FastWAM model, with video, context and action routes. They first select historical residuals along each route by the first-order gradient of the target loss, and learn task-specific combinations. Then they add a new residual shared across all routes by learnable exposure to compensate for missing capabilities in the library. On 35 LIBERO held-out tasks with 3 demonstrations, the method obtained a 72.2% success rate.

Critical strengths
  1. The problem setting is practical and the protocol is well defined. This study examines few-shot adaptation in a setting close to real-world deployment, where only the adapter history is kept, and no historical data is preserved. Experiments were performed on a separate development set with measures in place to prevent data leakage.

  2. Control experiments confirm the effectiveness of the method. Training from scratch achieved 57.9% performance under consistent procedures, randomly chosen historical candidates achieved 65.7%, and candidates chosen based on target loss achieved 72.2%. It also show through ablation studies that both historical reuse and the new residual component are indispensable.

Critical weaknesses
  1. The gains may not be due to capability reuse. With bank composition only, success is only 0.9%, while random candidates are already 7.8 pp better than training from scratch. Thus, most of the benefit may come from extra capacity or regularization rather than reuse of specific capabilities (a similar finding is reported by Liu et al., 2026, arXiv:2602.12323). No frozen randomly initialized residuals or uniform averaging are considered in the paper. Thus, the central claim remains unsupported.

  2. No comparisons with other existing composition methods were made. The main experiments do not compare against LoRAHub, uniform averaging or layer-wise learned composition (e.g., CAT, MoLE). Evidence for route-wise composition being better is only from a single four-task group (Long-B). , the core contribution is not convincingly established.

Questions for authors

Could the author clarify the differences between seeds 42–44 in Table 21? The closed-loop success rate for each task may fluctuate over 50 or more iterations, but the fluctuations shown in the Table 21 are very small.

4 comments

We thank the reviewer for recognizing the practical history-aware few-shot setting, the disjoint development protocol, and the matched experimental controls. We add experiments to clarify the contribution of learned historical capabilities and strengthen comparisons with existing composition methods. For your convenience, please also refer to our revised manuscript for the corresponding updates.

 

A1 (W1): Evidence for historical capability reuse and complementary refinement

Value of learned historical capabilities. Following the reviewer's suggestion, we replace the trained bank with frozen, output-scale-matched randomly initialized residuals under the same reuse–refinement procedure.

MethodSpatialObjectGoalLong-ALong-BLong-COverall (35 tasks)
Scale-matched random bank56.5%68.0%84.2%33.3%33.8%0.0%56.7%
Trained bank (R²-WAM)78.0%85.6%76.7%56.7%50.0%30.0%72.2%

The new random-bank run has higher SR than the paper's fresh-adaptation mean on Goal, Long-A, and Long-B, suggesting that random residual directions can sometimes benefit subsequent refinement. The original trained-bank mean exceeds the random-bank single-run SR in five of six groups, by +15.5 pp in task-weighted Overall SR (72.2% vs. 56.7%). This is a descriptive cross-protocol difference. Goal is the exception (76.7% vs. 84.2%), but this does not imply a lack of transferable historical capabilities: the trained bank reaches 85.0% with fixed η = 1 and 87.7% with task-specific Top-2 selection. This suggests that the effectiveness of historical reuse also depends on how learned capabilities are selected and refined, particularly for heterogeneous target groups such as Goal-Placement.

Value of target-aligned selection. The random candidates baseline in Table 1 samples from trained historical adapters, which may already contain transferable capabilities, explaining its improvement over fresh adaptation (65.7% vs. 57.9%). LoRAHub similarly reports 34.7% with randomly selected trained LoRAs versus 27.0% zero-shot on NLP tasks, rising to 35.4% with loss-based candidate filtering (Huang et al., 2024). In our setting, target-loss selection improves success from 65.7% to 72.2% (+6.5 pp), demonstrating the benefit of identifying target-aligned historical support.

Complementary roles of reuse and refinement. While frozen-module composition can enable cross-task generalization in NLP without learning new adapters, embodied targets may require capabilities absent from the bank. Table 1 reports 0.9% for bank composition only and 35.5% for new residual only, compared with 72.2% when combined. Here, new residual only removes history while retaining the full method's learnable exposure, whereas fresh adaptation fixes η=1\eta=1 and serves as the standalone baseline (57.9%). The low composition-only success therefore does not imply ineffective reuse. Historical adapters provide transferable partial capabilities, while refinement learns missing ones. Both contribute to successful embodied adaptation.

A2 (W2): Comparisons with existing composition methods and full-benchmark evidence for functional reuse

Comparison protocol. Existing methods such as LoRAHub, CAT, and MoLE primarily study adapter composition in language models. To compare these strategies in embodied adaptation, we keep the anchor, bank, target demonstrations, and Stage-2 refinement fixed, varying only the optimization method or granularity of historical composition.

1. Composition optimization. We compare three strategies using the same q-selected route-wise Top-4 candidates. Uniform fixes equal coefficients across candidates and tasks. NGOpt (LoRAHub-style) uses 32 Nevergrad loss evaluations per task for gradient-free Stage-1 coefficient optimization, followed by joint refinement. R²-WAM instead optimizes task-specific coefficients using target-loss gradients.

MethodSpatialObjectGoalLong-ALong-BLong-COverall (35 tasks)
Uniform61.8%81.6%58.3%50.0%32.0%29.0%60.6%
NGOpt (LoRAHub-style)75.4%85.4%53.7%50.0%31.5%25.0%64.5%
R²-WAM (GSC Top-4)78.0%85.6%76.7%56.7%50.0%30.0%72.2%

Uniform reaches 60.6%, highlighting the importance of learned, task-specific historical compositions over fixed equal-weight mixtures. The gaps are largest on the more heterogeneous Goal-Placement and Long-B groups (18.4 and 18.0 pp), consistent with their greater need for differentiated historical combinations. NGOpt improves to 64.5%, confirming the benefit of coefficient optimization. R²-WAM reaches 72.2% (+7.7 pp over NGOpt), demonstrating the effectiveness of gradient-based target-loss optimization in this few-shot setting.

2. Composition granularity. We compare whole-adapter, functional-route, and layer-wise composition.

MethodSpatialObjectGoalLong-ALong-BLong-COverall (35 tasks)
Whole-adapter WQ Top-462.8%75.4%67.0%49.3%47.0%15.0%61.4%
Whole-adapter layer-wise (CAT-style)†61.2%69.8%68.3%51.7%36.3%15.0%58.6%
Route-conditioned layer-wise†65.4%71.8%84.2%60.0%37.5%10.0%63.6%
R²-WAM (GSC Top-4)78.0%85.6%76.7%56.7%50.0%30.0%72.2%

† The two layer-wise rows use single seed.

Whole-adapter versus route-wise composition. We extend the whole-adapter q-ranked (WQ) baseline from Section 5.4.2 to all 35 held-out tasks, with matched Top-4 support budgets, coefficient optimization, and Stage-2 refinement. WQ shares historical candidates and coefficients across video, context, and action routes, whereas R²-WAM selects and composes them separately. R²-WAM outperforms whole-adapter WQ by +10.8 pp in task-weighted Overall SR, with higher SR in all six groups. This extends the original Long-B finding and demonstrates the advantage of independently selecting and combining historical capabilities along functional routes.

Route-conditioned versus whole-adapter layer-wise composition. We evaluate two static layer-wise parameterizations across six target groups. Route-conditioned layer-wise learns separate coefficients for each task, layer, and video/context/action route, retaining each route's Top-4 support. The whole-adapter layer-wise (CAT-style) control shares coefficients and one q-derived Top-4 support across routes at each layer. It is an adaptation to our structured bank, not a reproduction of CAT. MoLE requires an input-conditioned gating network with additional trainable parameters, so we do not include it in this static layer-wise composition comparison.

Both layer-wise variants achieve only nearly 0% composition-only SR across the six groups, so historical composition alone does not solve these embodied tasks. Refinement remains necessary in this setting. Within the layer-wise comparison, route-conditioned composition has higher SR than the whole-adapter CAT-style control in five of six groups. It also exceeds R²-WAM on Goal (84.2% vs. 76.7%) and Long-A (60.0% vs. 56.7%), but its task-weighted Overall SR is lower (63.6% vs. 72.2%). The latter comparison uses different evaluation protocols and is descriptive.

A3 (Q1): Clarification of training seeds and reported standard deviations

Independent training seeds. Seeds 42–44 correspond to independent training runs with the same demonstrations and hyperparameters. Changing the seed alters new-adapter initialization, training-window order, and stochastic sampling, yielding independently trained weights.

Reported variability. The original Table 21 reports mean and sample SD across seeds 42–44. For each seed, Overall first aggregates 35 tasks with 50 rollouts per task (1,750 total). The sample SD is then computed across the three seed-level Overall SRs, not across individual binary rollouts. Aggregation across tasks can yield a small Overall SD even when individual task outcomes vary. We retain the original means in this rebuttal without repeating SD values. Added single-seed controls are identified separately.

Dear Reviewer KAY2,

Thank you again for your thoughtful feedback. As the discussion period draws to a close, we would like to briefly follow up on our responses.

Following your suggestions, we added random-bank controls and comparisons with LoRAHub-style, whole-adapter, and layer-wise composition. We also clarified the roles of reuse and refinement, training seeds, and reported variability. These results and clarifications are detailed in our rebuttal and revised manuscript.

We greatly value this opportunity to discuss and improve our work. If your time permits, we would sincerely appreciate your feedback on whether these additions address your concerns. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,
The Authors of Submission #18362

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all reviewers for their constructive feedback. Below we summarize the main clarifications and additional experiments; details are available in our revised manuscript and reviewer responses.

R²-WAM enables history-aware few-shot adaptation through functional reuse and refinement: composing frozen historical capabilities across video, context, and action routes while learning shared corrections. With three demonstrations per task, it achieves 72.2% success across 35 held-out tasks, versus 57.9% for fresh adaptation. New comparisons show gains of 10.8 pp over whole-adapter composition and 7.7 pp over LoRAHub-style optimization under matched refinement. NERO experiments further support reuse under limited target data on one held-out physical task.

Original denotes submitted evidence; New denotes rebuttal experiments. Clarifications are labeled separately. Locations refer to the detailed reviewer responses.

ConcernRaised byEvidence sourceResponse and key evidenceLocation
Contribution and prior workKAY2, q6ogOriginal (clarified)Fixed-bank adaptation jointly learns task- and route-specific compositions and new corrections for missing capabilities. Sequential model evolution instead addresses task-stream learning and retention. Composition alone achieves 0.9%, versus 72.2% with refinement.KAY2 A1; q6og A1–A2
Historical capabilities and selectionKAY2, q6ogOriginal + NewOriginal: Target-loss selection improves over random trained-adapter selection by 6.5 pp (72.2% vs. 65.7%). New: The trained-bank mean exceeds the scale-matched random bank by 15.5 pp (72.2% vs. 56.7%), with higher success in five of six groups.KAY2 A1; q6og A2–A3
Composition strategiesKAY2, q6ogNewWith matched route-wise candidates and refinement, uniform / NGOpt composition reaches 60.6% / 64.5%, versus 72.2% for R²-WAM. Whole-adapter CAT-style / route-conditioned layer-wise controls reach 58.6% / 63.6%. The latter exceeds R²-WAM on Goal and Long-A, but is lower overall.KAY2 A2; q6og A1
Functional reuse across the benchmarkKAY2, 32uY, q6ogNewExtended whole-adapter q-ranked composition (WQ) to all 35 targets. Route-wise composition achieves 72.2% versus 61.4% (+10.8 pp), with gains in all six groups.KAY2 A2; 32uY A2; q6og A1
Physical-robot validation32uYNewAdded NERO experiments with four historical adapters and one held-out tape-on-box target. Under matched data budgets, both methods achieve 10/10 at 75 demonstrations; R²-WAM improves success from 0/10 to 5/10 at 30 and 1/10 to 3/10 at 20.32uY A1
Exposure and the residual-only gap32uY, q6ogOriginal (clarified) + NewClarified η_min = 0.05 and initial η = 0.1. Residual-only retains learned exposure; fresh adaptation fixes η = 1. Without history, fixing η recovers fresh success on Object (59.0% → 72.6%) and Goal (27.7% → 75.0%). With history, Object favors learned exposure (85.6% vs. 65.4%), showing dependence on historical support.32uY A3; q6og A2
Bank sensitivity and heterogeneityq6ogOriginal + NewNew: Bank scaling is non-monotonic: Long-C peaks at 15 adapters (30.0%); Long-B peaks at 8 (56.5%) and drops to 44.0% at 12. Original: Task-specific Top-2 selection improves Goal success over group-shared selection (87.7% vs. 71.7%), supporting flexibility for heterogeneous targets.q6og A3
Persistence and consolidationq6ogOriginal (clarified) + NewExact task-wise folding yields 0.0000382% action relative L2 error in FP32 on one Long-B task, with identical outcomes in three paired trials. Group-level F-mean reduces success on Long-B and Object, with lower retained success in the more heterogeneous Long-B group. The loss arises from coefficient consolidation; subsequent folding preserves success.q6og A3
Seeds and variabilityKAY2Original (clarified)Seeds 42–44 are independent training runs. Each seed's Overall aggregates 35 tasks × 50 rollouts; sample SD is computed across the three seed-level rates.KAY2 A3

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.