From a Prompt to Repertoires: Evolving Functional REpertoires Enable LLM Continual Learning

Fengyuan Liu; Yue Wang; Hangxi Guo; Fengyuan Liu; Chenxu Wu; Yanguang Liu; Mengnan Du

cs.LGOct 8, 2026

From a Prompt to Repertoires: Evolving Functional REpertoires Enable LLM Continual Learning

Authors: Fengyuan Liu, Yue Wang, Hangxi Guo, Fengyuan Liu, Chenxu Wu, Yanguang Liu, Mengnan Du

Organizations: The Chinese University of Hong Kong, Shenzhen · Shanghai AI Laboratory · University of Science and Technology of China · Independent Researcher · New Jersey Institute of Technology

Abstract

Continual learning remains challenging for large language models, which must enable models to acquire new skills and knowledge without degrading existing capabilities. Existing approaches typically address this challenge by carefully designing how model parameters are updated. In contrast, prompt optimization avoids costly parameter updates while achieving competitive or even superior performance to reinforcement learning methods such as GRPO on individual knowledge-intensive and reasoning tasks. This raises a natural question: \textit{Can prompt optimization, as an efficient adaptation approach, be directly applied to continual learning?} Our analysis shows that, under sequential task adaptation, it suffers from catastrophic forgetting, while optimized prompts accumulate rules that overfit to local task distributions. To address these limitations, we propose \emph{Evolving Functional REpertoires} (EFRE), which replaces a single prompt with a repertoire of functions that evolves as new tasks arrive: compatible updates refine existing functions, while conflicting updates trigger the emergence of new ones. On a three-task continual-learning stream, EFRE achieves a final average performance 7.50 percentage points higher than GRPO. Moreover, after adaptation to the Bio task, its performance on FinQA decreases by only 1.56 percentage points, compared with 25.10 percentage points for the base prompt optimization method. We further instantiate EFRE in a minimal agent system and observe consistent improvements across different backbone models. Overall, these results demonstrate EFRE's strong performance in continual learning for large language models and highlight its substantial potential for continual learning in advanced agent systems.

Figures & tables

Figure 1: Overview of Evolving Functional REpertoires (EFRE). Each function is represented by a prompt and evolves through feedback-driven refinement on incoming data. Candidates that improve current performance undergo a historical consistency check: compatible updates refine the current function, while incompatible updates trigger a sequential search for another function that can absorb them. If no compatible update is found, a new function emerges. The repertoire thus evolves through functional refinement, reuse, and emergence, enabling continual adaptation while preserving previously learned capabilities.

Figure 2: Functional reuse patterns in continual-learning and single-task settings. Curves show utilization of the six most-used functions. (a), (b) report utilization on ToolUse, FinQA, and SciKE-Bio. (c) compares both refinement configurations for single-task optimization, with a separate vertical scale.

\dagger

denotes GPT-5.6-Luna as the refinement model; Qwen3-8B is used otherwise.

Figure 3: Continual-learning performance along the ToolUse

\rightarrow

FinQA

\rightarrow

SciKE-Bio sequence. Each panel shows one method, with curves tracking performance on all three tasks before adaptation and after each learning stage. Stars mark performance immediately after training on the corresponding task, and shaded regions indicate the task introduced at each stage. The two color palettes distinguish Qwen3-8B and GPT-5.6-Luna backbones. EFRE variants are grouped in the rightmost column; EFRE uses Qwen3-8B for refinement by defualt;

\dagger

indicates GPT-5.6-Luna instead.

Setting	Cross-task reuse	Shared Func.	Func.	Example learned function
Continual Learning	41.7%	2	7	Identify the next required action, select a tool, and check its arguments against the tool schema before execution.
Continual Learning †	56.3%	6	19	Extract the relevant financial values and units, write the calculation, and verify its result against the source table.
Single-task	–	–	8	Decompose the question into two linked subquestions, gather evidence, and synthesize a supported answer.
Single-task †	–	–	12	Translate the conditions into equations, solve them, and verify candidates against the original constraints.

Table 1: Functional organization of the learned repertoires. For continual learning, Cross-task reuse denotes the percentage of evaluation instances routed to functions used by at least two tasks, averaged equally across ToolUse, FinQA, and SciKE-Bio. Shared Func. counts functions used by at least two tasks, while Func. denotes total repertoire size. Cross-task reuse is not defined for single-task settings, where each benchmark is optimized independently. The final column illustrates a learned function. Rows marked with

\dagger

use GPT-5.6-Luna for refinement; others use Qwen3-8B.

History Retention
Setting	after ToolUse		after FinQA		after Bio
Setting	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$
GEPA	+1.07 ± 0.42	–	+30.48 ± 1.73	-0.49 ± 0.28	+4.59 ± 0.86	-11.82 ± 1.24
GRPO	+10.28 ± 0.94	–	+39.46 ± 1.65	-0.40 ± 0.31	+20.58= ± 1.29	-8.32 ± 1.07
Last-20	+8.60 ± 0.73	–	+44.58 ± 1.36	+1.49 ± 0.24	+19.52 ± 1.08	-0.31 ± 0.19
Last-50	+6.76 ± 0.91	–	+40.19 ± 1.62	+1.48 ± 0.31	+16.04 ± 1.27	-1.33 ± 0.38

Table 2: Ablation studies on history retention, parameter settings, and function selection along the ToolUse

\rightarrow

FinQA

\rightarrow

Bio stream.

\mathcal{J}_{\mathrm{plas}}

measures the score change on the current task, while

\mathcal{J}_{\mathrm{stab}}

measures the mean score change over previously encountered tasks with equal weights. Values are reported in percentage points; higher is better. Stability is undefined at the first stage and denoted by –. Baselines are shown in light beige, default configurations are shaded light blue, and the oracle reference is shown in gray. Bold means with pale gold shading indicate the best result in each column within each ablation group, excluding the baseline and oracle references.

Method	HotpotQA	IFBench	AIME-2025	LiveBench Math	Aggregate	Improvement
Base	42.33	36.90	27.33	48.70	38.82	–
GRPO	43.33	35.88	38.00	51.26	42.12	+3.30
ACE	60.00	35.72	31.00	52.76	44.87	+6.06
GEPA	62.33	38.61	32.00	51.95	46.22	+7.41
GEPA+Merge	64.33	28.23	32.00	51.95	44.13	+5.31
EFRE	70.67	39.26	34.33	51.88	49.04	+10.22

Table 3: Performance comparison across four benchmarks. Both EFRE configurations use Qwen3-8B as the backbone. EFRE and EFRE † use Qwen3-8B and GPT-5.6-Luna as the refinement model, respectively. Aggregate is the unweighted mean across benchmarks, and Improvement is the percentage-point gain over Base. EFRE and EFRE † are shaded light blue and darker blue, respectively. Bold means indicate the best result in each column.

Method	Blind Spectrum Monitoring	Codebase Adaptation	Cohort Studies	Database Exploration	Exploitable Poker	Sales Prediction
Qwen3-8B
+ ICL	12.0 ± 5.6	-32.0 ± 8.4	-28.0 ± 7.9	10.0 ± 5.1	-8.0 ± 6.7	25.0 ± 6.2
+ GEPA	26.0 ± 4.8	-15.0 ± 7.2	-18.0 ± 6.8	22.0 ± 4.6	0.0 ± 5.9	42.0 ± 5.3
+ ACE	2.0 ± 4.4	-10.0 ± 6.9	-24.0 ± 7.3	5.0 ± 4.8	-5.0 ± 5.6	15.0 ± 5.9
+ EFRE	35.0 ± 4.1	-5.0 ± 6.2	-10.0 ± 6.0	30.0 ± 4.2	6.0 ± 5.1	52.0 ± 4.7
GPT-5.6-Luna

Table 4: Minimal agent results across six interactive environments with two backbones. Each backbone uses ICL, GEPA, ACE, or EFRE. Values are mean ± standard deviation; higher is better. EFRE rows are light blue. Bold means indicate the best result within each backbone, including ties.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Setting
Micro-batch size	3
History capacity $H$	20 examples per function
Historical consistency tolerance $\epsilon$	0.1
Maximum number of functions	50
Backbone model	Qwen3-8B, GPT-5.6-Luna
Refinement model	Qwen3-8B, GPT-5.6-Luna

Appendix

Table 5: Default EFRE hyperparameters. Model choices correspond to the experimental configurations reported in the main text.

Method	Optimizer	Batch size
SFT	AdamW	16
SDFT	AdamW	16
SDPO	AdamW	16
GRPO	AdamW	16
In-Place TTT	AdamW	16
GEPA	–	3

Appendix

Table 6: Optimizer and batch-size settings for comparison methods. A dash indicates that the setting is not applicable.

Qwen3-8B
Method	Base			After ToolUse			After FinQA			After Bio			Final Avg.
	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio
SFT	58.17	24.31	29.84	60.26	21.73	26.18	56.79	29.24	28.36	57.82	10.27	36.74	34.94
SDFT	57.28	24.16	30.22	63.76	28.34	34.81	64.21	69.74	29.36	62.18	68.83	43.27	58.09
SDPO	58.11	23.79	30.26	64.83	30.21	35.17	65.24	62.78	33.69	64.72	58.31	53.18	58.74
GRPO	57.86	26.23	33.78	68.14	26.37	35.82	67.74	65.83	29.28	51.22	65.71	49.86	55.60

Appendix

Table 7: Full continual-learning results. Each stage reports ToolUse, FinQA, and SciKE-Bio accuracy (%); Bio denotes SciKE-Bio. Final Avg. is the unweighted mean after Bio. Bold values indicate the highest score in each column within each backbone, including ties. Light and dark blue rows denote EFRE and EFRE † , respectively.

Qwen3-8B
Method	After ToolUse		After FinQA		After Bio
	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$	$\mathcal{J}_{\mathrm{plas}}\uparrow$	$\mathcal{J}_{\mathrm{stab}}\uparrow$
SFT	+2.09	—	+7.51	-3.47	+8.38	-8.97
SDFT	+6.48	—	+41.40	+0.45	+13.91	-1.47
SDPO	+6.72	—	+32.57	+0.41	+19.49	-2.50
GRPO	+10.28	—	+39.46	-0.40	+20.58	-8.32

Appendix

Table 8: Stage-local plasticity and stability calculated from Table 7 , in percentage points (higher is better). Bold values indicate the highest value in each column within each backbone, including ties. First-stage stability is undefined and shown as a dash. Light and dark blue rows denote EFRE and EFRE † , respectively.

Source $\rightarrow$ Target	Base			After ToolUse			After FinQA			After Bio			Final Avg.
	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio	ToolUse	FinQA	Bio
Q $\rightarrow$ Q	55.18	23.21	31.82	63.78	25.24	32.16	65.27	69.82	35.31	66.21	68.26	54.83	63.10
Q $\rightarrow$ L	66.27	29.16	38.23	72.86	31.94	40.58	74.12	75.46	43.17	72.38	71.84	57.17	67.13
L $\rightarrow$ Q	55.18	23.21	31.82	65.42	27.06	33.18	67.83	72.14	37.62	68.47	70.31	56.28	65.02
L $\rightarrow$ L	66.27	29.16	38.23	75.24	33.18	41.27	76.31	77.82	44.24	74.83	73.76	58.29	68.96

Appendix

Table 9: Cross-model evaluation of functional repertoires. The arrow denotes source-to-target backbone transfer; Q and L abbreviate Qwen3-8B and GPT-5.6-Luna. All configurations use GPT-5.6-Luna refinement. Each stage reports ToolUse, FinQA, and SciKE-Bio accuracy (%). Final Avg. averages the three scores after Bio. Bold values mark column maxima, including ties.

Qwen3-8B
Method	API calls	Output tokens (M)	API output cost (USD)	GPU hours	Accounted cost (USD)	Wall time (h)
SFT	0	0.00	0.00	18.2	36.40	4.8
SDFT	0	0.00	0.00	40.6	81.20	10.5
SDPO	0	0.00	0.00	50.4	100.80	13.0
GRPO	0	0.00	0.00	75.6	151.20	19.5
GEPA	1,400	9.60	11.52	6.0	23.52	3.7

Appendix

Table 10: Measured adaptation workloads and elapsed times for 12 configurations over a complete three-stage run. API output cost is calculated at

1.20 per million output tokens, and local GPU cost at

2.00 per GPU-hour. Accounted cost includes API output charges and local GPU cost but excludes API input charges. EFRE † uses GPT-5.6-Luna for refinement. Blue rows identify EFRE variants.

Figure 4: Prompt for selecting a function from the function library.

Figure 5: Prompt for refining a functional context using feedback and three queries.

Figure 16

Figure 17

Figure 18

Explore similar work

When Does Continual Learning Require Learning
Jul 8, 2026Anne Harrington, Nayan Saxena, Michael Murphy +7Continual Learning for LLMs Concept Drift
1UC Berkeley · 2Independent · 3Capital Fund Management
RECAP: Regression Evaluation for Continual Adaptation of Prompts
Jun 4, 2026Harsh Deshpande, Kushal Chawla, Sangwoo Cho +2Continual Learning for LLMs LLM Evaluation
Capital One
Harness Continual Learning: Continual Adaptation Beyond Model Parameters
Aug 19, 2026Borui Kang, Jinrui Gu, Junhan Lv +3Continual Learning for LLM Agents Continual Learning
State Key Laboratory for Novel Software Technology, Nanjing University, China · University of Wollongong, Australia

Jul 8, 2026, cs.LGJ/K move · Enter open · S save

When Does Continual Learning Require Learning

Anne Harrington, Nayan Saxena, Michael Murphy, Anastasia Borovykh+6

1UC Berkeley · 2Independent · 3Capital Fund Management

Continual Learning for LLMs Concept Drift Domain-Incremental Learning

Open paper AskPDF arXiv

As large language models (LLMs) become increasingly capable, the next question is how can we enable models to continually learn? Today, the field largely frames this as a problem of context management and mitigating forgetting. We argue this framing is incomplete: continual learning is fundamentally about increasing model competence as the world changes. We disentangle this change along two axes -- space, where the model encounters new domains, and time, where the underlying data drifts under a fixed task. This framing lets us study continual learning under realistic conditions: new domains arrive over time, facts drift past their training cutoff, and agentic interactions accumulate state across episodes. To evaluate methods under this setting, we recast widely used LLM benchmarks as sequential problems and introduce a single mechanism-agnostic protocol that compares prompt-based methods (GEPA, ACE), supervised learning (SFT, SDFT), reinforcement learning (GRPO, SDPO), and context compression (Cartridges, In-place TTT). Prompt-based methods fit each new stage quickly but degrade on future tasks. Distillation-based methods accumulate knowledge stably but struggle to update outdated facts. Context compression improves efficiency without substantially improving the ability to learn new tasks. Online reinforcement learning adapts most effectively to knowledge updates but remains sensitive to noisy reward signals. Overall, our results suggest that continual learning is not a single capability: different patterns of environmental change require fundamentally different update behaviors, determining when adaptation must be learned inside model weights and when it can be achieved through external scaffolding. We hope that understanding where each method succeeds and fails will guide the design of stronger continual learning systems.

Abstract

Explore similar work

When Does Continual Learning Require Learning

RECAP: Regression Evaluation for Continual Adaptation of Prompts

Harness Continual Learning: Continual Adaptation Beyond Model Parameters