Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. Updating shared parameters to learn new capabilities can degrade previously acquired ones, resulting in catastrophic forgetting. As a trajectory unfolds, different stages may require different capabilities, yet conventional joint training and static adapter merging maintain a fixed parameter configuration throughout execution. We introduce WeAct (Weight-level Internal Actions), which trains independent capability-specific LoRA adapters over a frozen backbone and adds adapter selection to the agent's action space. These internal actions use the model's native tool-calling interface for on-demand capability composition in a shared context. Across eight deep-research benchmarks, Qwen3-32B with WeAct improves average accuracy by 6.2% relative to full fine-tuning counterparts despite using less compute, and by 27.5% relative to the strongest static LoRA-merging method. WeAct also supports modular continual adaptation, by adding a new code computation adapter to expand the agent's capabilities and updating the existing writing adapter to improve report quality. Both updates leave the backbone and other adapters unchanged, supporting modular continual adaptation in agents.
Figures & tables
Figure 1: Internal actions extend the agent’s action space. WeAct selects LoRA adapters internally and interacts with the environment through external actions.
Figure 2: WeAct overview. (a) Capability learning builds the adapter library. (b) Tool-mediated inference switches adapters through internal tool calls and interacts with the environment through external tools. B/P/S/W nodes denote base/Plan/Search/Writing states; solid and dashed arrows indicate selected and alternative execution paths.
Figure 3: Capability retention under sequential training. Detailed results are reported in Appendix Table 7 .
Method
General
Multi-Hop QA
Search
GAIA
2Wiki
Bam- boogle
MuSiQue
Web WalkerQA
xbench- DS
Browse Comp
FRAMES
Avg.
Prompt & Multi-Agent-Based
ReAct
35.92 ± 4.73
46.20 ± 1.58
53.60 ± 4.46
13.98 ± 0.71
27.35 ± 1.71
35.00 ± 4.77
4.00 ± 1.39
48.67 ± 1.74
33.09 ± 1.09
Least-to-Most
37.86 ± 4.78
47.40 ± 1.58
48.00 ± 4.47
14.40 ± 0.71
27.21 ± 1.71
39.00 ± 4.88
5.00 ± 1.54
51.70 ± 1.74
33.82 ± 1.10
ManuSearch
41.75 ± 4.86
53.50 ± 1.58
60.80 ± 4.37
16.84 ± 0.76
28.38 ± 1.73
45.00 ± 4.97
7.50 ± 1.86
62.14 ± 1.69
39.49 ± 1.12
Single-Model Weight Adaptation
Table 1: Results on eight benchmarks (%). Subscripts give nominal task-based binomial standard-error approximations. Bold and underlined values denote the best and second-best scores per benchmark.
Figure 4: SearchAtlas trace diagnostics. Query utilization: fraction of queries supporting the answer; constraint coverage: fraction of question constraints covered by those queries; PK → answer: share of answers with unsupported content attributed to prior knowledge. Error bars are 95% task-bootstrap intervals.
Figure 5: Accuracy versus historical mean cumulative model-path time on four Search benchmarks. Times include saved retries.
Variant
MegaScience
SciBench
DS-1000
Δ GAIA
w/o LoRA-code
26.5
40.0
14.0
0.0
w/ LoRA-code
29.0
40.5
16.0
+1.0
Table 2: Results of code-adapter insertion. Δ GAIA is relative to the original K=3 system.
Method
ResearchQA
DeepResearch Bench
LoRA-write
hard
58.33
26.51
soft
54.14
27.92
LoRA-write++
hard
60.72 (+2.40)
27.71 (+1.21)
soft
55.97 (+1.83)
28.52 (+0.61)
Table 3: Writing refresh scores. Bold values show gains over Original.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Use
Source
Count
Main
MiroVerse-v0.1
5,887
Main
OpenSeeker-v1-Data
1,075
Main total
6,962
Code
MiroVerse-MegaScience
3,413
Writing++
DR-Tulu / OpenScholar
1,000
Appendix
Table 4: Training-data sources. Main-corpus counts refer to source trajectories before capability splitting; Code and Writing++ counts refer to capability samples used in Section 4.5 .
Policy change
Changed
Spans (before → after)
Strengthen targeted verification
4/80
Plan 20→20 ; Verif. 21→21
Narrow recap; broaden verification
36/78
Recap 30→12 ; Verif. 9→41
Revise tool/span rules
11/78
Evid. 19→19 ; Verif. 38→37
Appendix
Table 5: Policy sensitivity on fixed inputs. Changed : relabeled/total assistant messages, including base-only transitions. Slices overlap and are not independent.
Figure 7: Nominal training compute. (a) Total recipe compute, with WeAct split by capability. (b) Dense and causal-attention FLOPs per token. Error bars show length-statistic uncertainty (details in text).
Figure 8: Accuracy versus model-path time on General and Multi-hop QA. Points pair Table 1 accuracy with historical mean cumulative time, including retries. Colors and markers identify methods; line styles group benchmarks. Time axes are logarithmic in seconds.
Figure 9: Historical task-record wall-clock distributions. Colors and markers identify methods; line styles identify benchmarks. Each panel uses its own logarithmic time axis in seconds and retains all tails. Curves include all valid task-record times and saved retries, regardless of correctness; logged wall time can include post-answer judging.
Benchmark
N
ReAct
Full SFT
Single LoRA
TIES-R64
WeAct -soft
WeAct -hard
GAIA
103
0.00
16.50
26.21
14.56
4.85
1.94
2Wiki
1,000
0.70
1.00
6.40
4.70
1.00
0.60
WebWalkerQA
680
0.29
9.26
11.32
12.79
9.26
3.82
Appendix
Table 6: Tasks with saved retries (%). The denominator is the task count N for each method and benchmark.
Figure 10: Runtime decomposition on WebWalkerQA with 50 runs per method. Bars show stage-aligned means. The ellipsis denotes repeated capability switching; all durations are retained. ManuSearch* is a historical reference with unmatched deployment and tool conditions.
Plan
Search
Writing
Adapter organization
G
L
%
G
L
%
G
L
%
Single-LoRA (rank 32)
14.28
1.67
11.7
6.95
1.57
22.6
3.05
1.41
46.2
Single-LoRA (rank 96)
14.31
1.82
12.7
6.81
1.68
24.7
3.06
1.50
49.2
Random partition
13.36
1.32
9.9
6.16
0.96
15.6
2.58
1.09
42.4
Capability adapters
14.22
0.00
0.0
6.84
0.00
0.0
3.09
0.00
0.0
Appendix
Table 7: Acquired gain (G) and lost gain (L) in points of content-token accuracy, and fraction of acquired gain lost (%), pooled over the four orders in which each capability precedes another. Fractions use unrounded means.
Figure 11: Retention during subsequent training. Panels track the first capability in each order on monitoring trajectories; shading marks later stages, and 100% denotes the gain acquired in its own stage.
Figure 12: Gradient alignment and magnitude in the Single-LoRA pilot. (a) Pairwise cosines and (b) gradient norms at ten checkpoints of a rank-96 Single-LoRA run on frozen Qwen3-32B . Most pairwise cosines are positive, while gradient magnitudes differ across capabilities.
Figure 13: Local mismatch on capability-matched activations. Curves show trajectory-weighted layer means of Eq. 10 ; shading gives pointwise 95% trajectory-bootstrap confidence intervals. The dashed zero line is Individual’s identity reference. Full Additive uses no final rank compression.
Figure 14: Paired target-token NLL gaps relative to Individual. Points show validation records, boxes show medians and interquartile ranges, and diamonds mark means with 95% confidence intervals.
Method
QU ↑
Cov ↑
PK ↓
ReAct
0.365
0.336
0.590
Least-to-Most
0.386
0.360
0.600
ManuSearch
0.305
0.385
0.520
Single LoRA
0.315
0.332
0.630
Full SFT
0.410
0.478
0.530
TIES-R64
0.293
0.311
0.680
Appendix
Table 8: SearchAtlas diagnostics by method over the 100 tasks of Figure 4 . Values are task means; QU = query utilization, Cov = question-constraint coverage (answer graph), PK = share of answers with a prior-knowledge-to-answer edge. Bold marks the best value per column (lowest for PK).
MegaScience
SciBench
DS-1000
GAIA
#Tasks
200
575
100
103
Appendix
Table 9: The number of selected tasks used for evaluating code insertion tasks. The MegaScience and DS-1000 tasks were randomly sampled based on a balanced consideration of cost and representativeness.
Method
MMLU- Pro
IFEval
MT- Bench
Struct Flow
Human Eval
Base
58.03
77.82
9.3875
98.12
90.85
Full SFT
−19.99
−19.48
−0.93
−1.60
−6.86
DARE-TIES
−0.55
−4.61
+0.53
−0.95
−1.34
Single LoRA
−1.15
−3.80
0.00
−0.76
−5.37
WeAct -hard
+1.60
−4.51
+0.80
−0.08
−2.69
WeAct -soft
+1.21
−4.28
+0.80
−0.04
0.00
Appendix
Table 10: General-capability retention (external tools disabled). Base: scores; others: 100(S/SBase−1)% . Green/red mark the best/worst changes per column, including ties.
Rubric
WeAct -Hard
WeAct -Soft
Original
LoRA-write++
Δ
Original
LoRA-write++
Δ
Comprehensiveness
25.32
26.33
+1.01
26.50
26.78
+0.28
Depth
22.46
23.59
+1.13
23.82
24.19
+0.37
Instruction following
29.65
31.22
+1.57
31.36
32.45
+1.09
Readability
32.33
33.81
+1.48
33.32
35.10
+1.77
Appendix
Table 11: DeepResearchBench rubric scores before and after the Writing refresh (higher is better). Δ is the reported score-point change from Original to LoRA-write++ .
Evaluation Set
Rows
NLL
Perplexity
Original
LoRA-write++
Original
LoRA-write++
OpenScholar validation
205
1.27
1.02
3.56
2.76
Historical Writing validation
200
0.46
0.50
1.59
1.66
Appendix
Table 12: Supervised-token fit before and after the Writing refresh (lower is better).
Post-training can unlock new capabilities and improve performance on specialized tasks, but sometimes at the cost of catastrophic forgetting in other domains. This poses a problem in long agent trajectories that compose different capabilities. We reject this tradeoff by giving an agent a tool to switch between specialized LoRA adapters mid-trace. To test its effectiveness, we compose two synthetic coding tasks that are logically simple but require specialization. We find that this allows the model to solve problems it previously could not, that the model is able to switch autonomously (and find a new strategy that beats our human heuristic baseline on one task), and that this incurs an up to an 18x reduction in capability tax compared to an agent using only one specialized adapter. Our approach also substantially outperforms spawning subagents in both task capabilities (solving 4 of our hardest tasks versus none) and token usage (46.1x fewer tokens in some scenarios).
Kenneth Ge
Independent San Francisco, CA, United States of America
Agent performance depends jointly on the model parameters and the executable harness code that manages context and control flow. Optimizing either component in isolation can leave the system bottlenecked by its frozen counterpart: weight updates can change which harness is effective, while harness updates can change which model capabilities are exposed. Existing joint-adaptation methods optimize weights and textual prompts but leave the broader harness fixed. We propose Weight-Harness Alternating LEarning (WHALE), a simple recipe that alternates two phases: updating the model under the current harness, then searching for a better harness under the updated model. We instantiate these two phases with online rejection-sampling fine-tuning and Meta-Harness, respectively. When to switch is a key design choice: to separate real improvements from noise without over-optimizing against a changing counterpart, WHALE uses either fixed phase durations or an adaptive patience rule over training signals. Using Qwen3.5-2B/4B agents across three domains (search question answering, mathematical reasoning, and chess puzzles), WHALE outperforms weight-only, harness-only, and Fast-Slow Training by 4.15-24.38 percentage points in best mean@8 accuracy. Either component can be the bottleneck: harness search matches peak weight-only accuracy with far fewer rollouts in SearchQA, but improves math accuracy only after a weight update. Small interleaved updates also outperform stagewise weight-then-harness optimization in accuracy and rollout cost. The code is available at https://github.com/krafton-ai/WHALE.
Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.
Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov +21
Toyota Motor Europe · University of Bremen · University of Florence +17