KA2L: A Knowledge-Aware Active Learning Framework for LLMs

Haoxuan Yin; Chen Tang; Yangfan Wang; Lian Yan; Jingchi Jiang

cs.CLMar 18, 2026

KA2L: A Knowledge-Aware Active Learning Framework for LLMs

Authors: Haoxuan Yin, Chen Tang, Yangfan Wang, Lian Yan, Jingchi Jiang

Organizations: Faculty of Computing, Harbin Institute of Technology, Harbin, 150001, China · Institute for Advanced Algorithms Research, Shanghai, 201306, China · National Key Laboratory of Smart Farm Technologies and Systems, Harbin, 150001, China

Abstract

Fine-tuning large language models (LLMs) with high-quality knowledge has been shown to enhance their performance effectively. However, there is a paucity of research on the depth of domain-specific knowledge comprehension by LLMs and the application of targeted active learning to improve their expertise. To address this gap, we introduce the Knowledge-Aware Active Learning (KA2L) framework. This framework assesses LLMs' mastery of specific knowledge points to aid in constructing unanswerable or unknowable questions through latent space analysis. This active learning strategy enhances training efficiency by focusing on knowledge the model has yet to master, thereby minimizing redundancy in learning already acquired information. This study innovatively employs a knowledge distribution probing technique to examine the hidden states of specific Transformer layers and identify the distribution of known and unknown knowledge within the LLM. Additionally, a hidden-state decoding method is proposed to generate numerous unknown questions in natural language from the latent knowledge space. In our experiments, we selected nine open-source LLMs to validate the effectiveness of the proposed framework. Results indicate that KA2L not only significantly reduces 50% annotation and computation costs across two open-domain and one vertical-domain dataset but also achieves better performance, offering valuable insights into active learning strategies for LLMs. The code is available at https://github.com/greenjerry/KA2L.

Figures & tables

Figure 1: KA2L Workflow : (a) Knowledge Distribution Probing : Training Phase : For each question in the sampled question set, sample its hidden state once and its textual outputs multiple times. Perform semantic clustering on the textual outputs to calculate Semantic Entropy (SE). The SE is then binarized using a dynamic threshold to serve as labels for the classifier. An MLP classifier is trained using these hidden states and the binarized SE (BiSE). Inference Phase : For a new set of questions, sample their $l$ -th layer hidden states. These are then classified by the MLP classifier, representing the knowledge distribution as ”Known” and ”Unknown” knowledge. (b) Unknown Question Augmentation : Sample the hidden states ( $h_{unk}$ ) of questions identified as ”Unknown” from the knowledge distribution. These are then transformed and decoded into multiple similar questions. (c) Downstream Tasks : This knowledge distribution guides dataset construction and model fine-tuning. Many existing methods can be applied, such as LoRA ( Hu et al., 2022 ) and P-tuning ( Liu et al., 2022 ) .

Model	Hidden-State Layer Index
Model	TriviaQA Mix	NQ-Open Mix	MedMCQA Mix
DeepSeek-R1-Distill-Qwen-7B	28	28	28
glm4-9b-chat	38	38	24
Llama-2-7b-chat-hf	31	30	30
Llama-3.1-8B-Instruct	31	32	31
Mistral-7B-Instruct-v0.1	31	31	30

Table 1: Model-specific layer index choices based on probing accuracy.

Model	Learning Rate	Template	Batch Size	Gradient Accumu- lation Steps
DeepSeek-R1-Distill-Qwen-7B	5.0e-5	qwen	4	4
glm4-9b-chat	1.0e-4	glm4	2	8
Llama-2-7b-chat-hf	1.0e-4	llama2	4	2
Llama-3.1-8B-Instruct	1.0e-4	llama3	4	2
Mistral-7B-Instruct-v0.1	1.0e-4	mistral	4	4
Mistral-7B-Instruct-v0.3	1.0e-4	mistral	4	4

Table 2: SFT training parameters for different models.

Model	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)
DeepSeek-R1- Distill-Qwen-7B	None	0.02	0.52	1.31	76.29	10k Combine	2.44	12.55	7.92	83.70
	5k Known	1.85	10.41	6.50	83.19	10k Unknown	2.85	14.53	9.20	84.00
	5k Unknown	1.91	11.99	7.54	83.62	10k Augmented	2.62	13.55	8.91	84.00
glm4-9b-chat	None	0.08	1.71	4.21	78.48	10k Combine	6.90	28.27	19.67	86.80
	5k Known	4.17	20.62	13.86	85.17	10k Unknown	9.82	36.02	25.50	88.30
	5k Unknown	6.61	28.44	19.87	86.73	10k Augmented	8.48	29.92	21.67	87.13

Table 3: Active learning performance on MedMCQA. KA2L-selected data (Unknown) is compared against Known data and a mixed Combine setting. Bold values indicate the best performance.

Model	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)
DeepSeek-R1- Distill-Qwen-7B	None	0.02	0.46	1.26	76.39	10k Combine	6.10	13.17	9.68	86.45
	5k Known	4.69	11.15	7.87	86.19	10k Unknown	7.48	16.05	11.84	86.98
	5k Unknown	4.05	12.19	8.78	86.37	10k Augmented	2.85	13.56	10.04	86.49
glm4-9b-chat	None	0.22	3.68	8.63	79.94	10k Combine	28.78	45.94	36.50	91.33
	5k Known	19.55	35.78	27.67	89.62	10k Unknown	39.54	57.89	46.91	93.27
	5k Unknown	28.61	45.91	36.47	91.31	10k Augmented	22.49	46.00	36.62	91.34

Table 4: Active learning performance on the NQ_Open dataset. Bold values indicate the best performance.

Model	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)	SFT Dataset	BLEU	ROUGE-L	METEOR	BS(%)
DeepSeek-R1- Distill-Qwen-7B	None	0.02	0.49	1.24	76.13	10k Combine	9.98	23.65	16.38	87.05
	5k Known	7.79	18.77	12.82	86.32	10k Unknown	14.20	27.00	18.86	87.56
	5k Unknown	9.41	22.22	15.39	86.85	10k Augmented	6.68	24.46	17.11	87.16
glm4-9b-chat	None	0.15	3.62	8.16	79.13	10k Combine	21.51	41.26	30.71	88.18
	5k Known	14.68	31.65	23.03	86.34	10k Unknown	11.68	55.80	42.43	91.00
	5k Unknown	19.62	40.28	29.85	87.97	10k Augmented	15.21	40.47	30.07	88.09

Table 5: Active learning performance on the TriviaQA dataset. Bold values indicate the best performance.

Method	BLEU	ROUGE-L	METEOR	BertScore(%)
Traditional Methods ( $5k$ samples selected from $10k$ pool)
Random	$21.75\pm 0.03$	$38.53\pm 0.02$	$29.65\pm 0.02$	$90.01\pm 0.01$
Entropy	$19.59\pm 1.15$	$37.91\pm 0.10$	$29.38\pm 0.07$	$89.91\pm 0.01$
Coreset	$22.04\pm 0.07$	$38.69\pm 0.05$	$29.81\pm 0.03$	$90.16\pm 0.01$
BADGE (adapted)	$19.64\pm 0.15$	$38.70\pm 0.05$	$29.73\pm 0.05$	$90.06\pm 0.01$
Our Method ( $5k$ samples)

Table 6: Performance comparison of active learning methods on the NQ_Open dataset. Our method, KA2L (

5k

Unknown ), significantly outperforms all adapted traditional AL method, approaching the performance of the Full Dataset (

10k

) with only half the data budget. Results are reported as mean ± std over

4

runs. Bold values indicate the best performance.

Method	BLEU	ROUGE-L	METEOR	BertScore(%)
Traditional Methods (5k samples selected from 10k pool)
Random	$7.19\pm 0.22$	$26.67\pm 0.05$	$18.55\pm 0.05$	$86.44\pm 0.02$
Entropy	$6.85\pm 0.09$	$26.66\pm 0.07$	$18.46\pm 0.05$	$86.40\pm 0.01$
Coreset	$8.11\pm 0.11$	$28.48\pm 0.08$	$19.86\pm 0.03$	$86.83\pm 0.01$
BADGE (adapted)	$6.68\pm 0.12$	$26.76\pm 0.06$	$18.64\pm 0.07$	$86.41\pm 0.01$
Our Method (5k samples)

Table 7: Performance comparison of active learning methods on the MedMCQA dataset. Bold values indicate the best performance.

Method	BLEU	ROUGE-L	METEOR	BertScore(%)
Traditional Methods (5k samples selected from 10k pool)
Random	$12.33\pm 0.14$	$45.25\pm 0.01$	$32.96\pm 0.02$	$88.95\pm 0.01$
Entropy	$\textbf{22.34}\pm\textbf{0.08}$	$46.25\pm 0.09$	$33.91\pm 0.06$	$89.10\pm 0.01$
Coreset	$20.37\pm 1.80$	$45.64\pm 0.07$	$33.31\pm 0.04$	$89.02\pm 0.00$
BADGE (adapted)	$20.58\pm 0.22$	$45.50\pm 0.06$	$33.15\pm 0.03$	$88.98\pm 0.01$
Our Method (5k samples)

Table 8: Performance comparison of active learning methods on the TriviaQA dataset. Bold values indicate the best performance.

Models	Ours	SE Probe	Accuracy Probe	Log-Likeli hood	Regular Entropy	P(True)	Semantic Entropy
	Ours	SE Probe	Accuracy Probe	Log-Likeli hood	Regular Entropy	P(True)	Semantic Entropy
DeepSeek-R1-Distill-Qwen-7B	0.89	0.85	0.81	0.61	0.84	0.83	0.86
glm4-9b-chat	0.83	0.81	0.72	0.56	0.76	0.82	0.80
Llama-2-7b-chat-hf	0.85	0.78	0.70	0.55	0.73	0.71	0.77
Llama-3.1-8B-Instruct	0.89	0.85	0.73	0.57	0.75	0.79	0.79
Mistral-7B-Instruct-v0.1	0.88	0.83	0.76	0.62	0.75	0.78	0.81

Table 9: Performance comparison of knowledge distribution probes (AUROC) on TriviaQA dataset. Our method achieves the highest score. Bold values indicate the best performance.

Figure 2: Classifier AUROC score across different hidden layers (embedding layer excluded) on the Trivia_QA dataset.

Dataset	$\gamma^{*}-0.20$	$\gamma^{*}-0.10$	$\gamma^{*}-0.05$	$\gamma^{*}$	$\gamma^{*}+0.05$	$\gamma^{*}+0.10$	$\gamma^{*}+0.20$
TriviaQA	0.8543	0.8602	0.8623	0.8853	0.8847	0.8784	0.8721
NQ_Open	0.8232	0.8290	0.8277	0.8307	0.8265	0.8248	0.8247
MedMCQA	0.8281	0.8267	0.8323	0.8345	0.8285	0.8238	0.8333

Table 10: AUROC scores of the knowledge distribution probe under perturbed binarization thresholds for LLaMA-3.1-8B-Instruct (Layer

31

). The optimal threshold

\gamma^{*}

is determined by minimizing MSE. The highest score in each row is highlighted in bold.

Figure 3: PCA Analysis on Mistral-7B-Instruct-v0.3, showing a macroscopic overview of knowledge distribution.

#	Question	Ground Truth	Generated Answers (10 Samples)
Low Entropy (Known Knowledge, SE $\approx$ 0)
1	how many keys are on microsoft windows standard keyboard	104	104 (10 times)
2	who played willy wonka in the original movie	Gene Wilder	Gene Wilder (8 times), gene wilder (2 times)
3	how many original harry potter books are there	7	7 (6 times), Seven (2 times), seven (2 times)
Medium Entropy (The Knowledge Boundary, SE $\approx$ 1.3 - 1.7)
4	the haagen-dazs brand of ice cream originated in which country	United States	United States (2), Poland (4), Scandinavia (1), scandinavian (1), Belgium (1), Jamaica (1)

Table 11: Case Study: Generative distributions of 10 independent samples across different SE ranges. The Medium SE category explicitly visualizes the ”Knowledge Boundary” where outputs mix correct facts with plausible hallucinations.

Dataset	Avg Usable Variants	Source BERTScore	Self-BLEU	Exact-Copy Rate
MedMCQA	3.51	0.908	42.24	¡0.01
NQ_Open	3.37	0.880	35.01	¡0.01
TriviaQA	3.19	0.913	48.13	¡0.01

Table 12: Intrinsic quantitative evaluation of generated question variants. Avg Usable Variants indicates the ratio of unique variants among 5 generations. Source BERTScore measures semantic relatedness to the original question. Self-BLEU measures the lexical diversity among the variants themselves.

Augmentation Trait	Original Source Question	Generated Question Variants
Structural Rephrasing (TriviaQA)	Ex-world champion boxer Alan Rudkin, who died in September 2010, fought at what weight?	• Ex-world champion boxer Alan Rudkin, who died in September 2010, was fighting at what weight and age in 1960s?
Entity Substitution (NQ_Open)	Which player has won the most UEFA champions league titles?	• Which player has won the most trophies in the European cup 2018? • Which team has won the most trophies in the European cup 2017?
Terminological Perturbation (MedMCQA)	Type of immunologic injury in AIHA (Auto immune hemolytic anemia)	• Type of immunological response to IgA hemolytic anemia (IgA) is

Table 13: Qualitative examples of generated question variants. The decoder introduces beneficial diversity through structural rephrasing, entity substitution, and terminology perturbation, effectively broadening the augmented data distribution.

Dataset	Domain Ratio	General	Domain	BLEU	ROUGE-L	METEOR	BERTScore	Exact-Copy Rate
NQ_Open	0%	250,000	0	5.42	22.94	39.18	85.83	¡0.01
NQ_Open	10%	225,000	25,000	6.24	23.39	41.92	86.40	¡0.01
NQ_Open	20%	200,000	50,000	7.67	26.79	45.98	87.16	¡0.01
NQ_Open	30%	175,000	75,000	6.94	25.37	44.07	86.93	¡0.01
TriviaQA	0%	250,000	0	16.12	43.56	50.33	89.57	¡0.01
TriviaQA	10%	225,000	25,000	19.16	47.07	53.18	90.22	¡0.01

Table 14: Effect of decoder-training data composition on generation metrics using Llama-3.1-8B-Instruct . The total training size is fixed at 250,000. Bold values indicate the highest metric among the evaluated ratios.

Active Learning

Active learning selects the examples, labels, features, or experiments most useful for improving a model under limited budgets.

See all 288 papers

Context Training with Active Information Seeking
May 13, 2026Zeyu Huang, Adhiguna Kuncoro, Qixuan Feng +4Knowledge Augmentation for Language Models LLM Prompting
The University of Edinburgh · Google DeepMind
Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation
Jun 9, 2026Yupu Hao, Zhuoran Jin, Huanxuan Liao +2Knowledge Augmentation for Language Models Inference-Time Scaling
1The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China
From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning
Aug 19, 2026Zuocheng Ying, Yang Yang, Yumou Wu +6Factual Knowledge in Language Models Implicit Reasoning in Language Models
ByteDance · Huazhong University of Science and Technology

May 13, 2026J/K move · Enter open · S save

Context Training with Active Information Seeking

Zeyu Huang, Adhiguna Kuncoro, Qixuan Feng, Jiajun Shen+3

The University of Edinburgh · Google DeepMind

Knowledge Augmentation for Language Models LLM Prompting

Open paper AskPDF arXiv

Most existing large language models (LLMs) are expensive to adapt after deployment, especially when a task requires newly produced information or niche domain knowledge. Recent work has shown that, by manipulating and optimizing their context, LLMs can be tailored to downstream tasks without updating their weights. However, most existing methods remain closed-loop, relying solely on the model's intrinsic knowledge. In this paper, we equip these context optimizers with Wikipedia search and browser tools for active information seeking. We show that naively adding these tools to a standard sequential context optimization pipeline can actually degrade performance compared to baselines. However, when paired with a search-based training procedure that maintains and prunes multiple candidate contexts, active information seeking delivers consistent and substantial gains. We demonstrate these improvements across diverse domains, including low-resource translation (Flores+), health scenarios (HealthBench), and reasoning-heavy tasks (LiveCodeBench and Humanity's Last Exam). Furthermore, our method proves to be data-efficient, robust across different hyperparameters, and capable of generating effective textual contexts that generalize well across different models.

Abstract

Active Learning

Explore similar work

Context Training with Active Information Seeking

Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation

From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning