Selct2Know: Domain-Specific QA Framework

• The paper introduces S2K as a framework that fuses internal model knowledge with external documents through token-level self-selection to optimize answer generation.
• It reframes domain adaptation as a knowledge selection problem, reducing reliance on expensive continued pretraining and mitigating retrieval noise.
• S2K employs a two-stage training process with Selective SFT and GRPO, achieving competitive performance on medicine, law, and finance benchmarks with far less data.

Searching arXiv for the specified Selct2Know paper and closely related work to ground the article in current research. arxiv_search(query="(He et al., 21 Aug 2025) Selct2Know Select to Know internal-external knowledge self-selection domain-specific question answering", max_results=5) arxiv_search(query="Selct2Know OR \"Select to Know\" domain-specific question answering", max_results=10) arxiv_search(query="(He et al., 21 Aug 2025)", max_results=3) Selct2Know (S2K) is a post-training framework for domain-specific question answering that treats domain adaptation as a knowledge selection problem rather than as a choice between retrieval-only augmentation and expensive continued pretraining. Introduced in "Select to Know: An Internal-External Knowledge Self-Selection Framework for Domain-Specific Question Answering" (He et al., 21 Aug 2025), S2K is designed for settings such as medicine, law, and finance, where a general LLM often retains partial but unreliable domain knowledge because domain knowledge follows a long-tail distribution in pretraining. Its core mechanism is an internal-external knowledge self-selection strategy that chooses, at the token level, between answers generated from the model’s internal parametric knowledge and answers generated from retrieved external documents, followed by selective supervised fine-tuning and a reasoning-enhanced GRPO stage.

1. Problem setting and motivation

S2K is motivated by three limitations in domain-specific question answering. First, general LLMs struggle because some domain facts, terms, and patterns have already been internalized, but many are incomplete, rare, or only partially remembered. The paper argues that this leaves useful internal knowledge underutilized: a model may be confident on some tokens in an answer but uncertain on others. Second, retrieval-augmented generation introduces latency and computation, and noisy or conflicting retrieval can push the model toward hallucinations or incorrect answers. Third, continued pretraining can internalize knowledge, but it is expensive, data-hungry, and often inflexible across domains (He et al., 21 Aug 2025).

The framework therefore rejects a binary choice between “use only retrieval” and “internalize everything through continued pretraining.” Instead, it assumes that domain QA should exploit partial internal knowledge whenever it is reliable, while using external evidence to fill gaps. The paper explicitly notes that domain-pretrained models can require enormous corpora—BioMistral is cited as needing around three billion tokens—whereas S2K aims to achieve comparable performance with orders of magnitude less data.

A central design principle is progressive knowledge acquisition. The authors argue that knowledge acquisition should be progressive, mirroring human learning: first understanding concepts, then applying them to complex reasoning. This principle governs both the synthetic data pipeline and the two-stage training procedure.

2. Internal-external knowledge self-selection

The key innovation in S2K is the internal-external knowledge self-selection strategy. For a question $Q$ and supporting document chunk $D$, the model produces two answers in parallel:

$A^E = P_M(Q,D)$

$A^I = P_M(Q)$

The fused answer is then built token by token by comparing the model’s confidence in the internal and external candidates:

$$t_i^F = \begin{cases} t_i^I,&\text{if } P_M(t_i^I \mid Q, A^F_{<i}) > P_M(t_i^E \mid Q, D, A^F_{<i})\ t_i^E,&\text{otherwise} \end{cases}$$

This formulation is explicitly autoregressive, since the previous fused tokens $A^F_{<i}$ are always part of the context. The paper’s claim is not merely that retrieval should condition generation, but that internal and external knowledge should compete during generation.

To reduce greedy token-by-token mistakes, S2K uses a window-based version. For a window of size $W$, it compares average log-probabilities and introduces a margin $C$ that makes the model more conservative about using internal knowledge:

$$t_{i:i+W}^F = \begin{cases} t_{i:i+W}^I, & \text{if } \frac{1}{W} \sum_{j=0}^{W-1} \log P_M(t_{i+j}^I \mid Q, A^F_{<i}) \geq \frac{1}{W} \sum_{j=0}^{W-1} \log P_M(t_{i+j}^E \mid Q, D, A^F_{<i}) + C\ t_{i:i+W}^E, & \text{otherwise} \end{cases}$$

Here $W$ is the window size and $D$0 is a threshold margin. The ablation study reports that increasing $D$1 reduces the proportion of selected internal tokens, with performance peaking at $D$2, and that increasing $D$3 makes selection less greedy, with performance peaking at $D$4 (He et al., 21 Aug 2025).

This mechanism is the main point of departure from standard RAG. In standard RAG, the model conditions on retrieved text; in S2K, the model actively decides whether its own memory is more trustworthy than retrieval on each generation segment. The framework also differs from continued pretraining because it does not attempt to rewrite all domain knowledge into parameters indiscriminately.

3. Data generation and progressive training

S2K assumes only an unstructured domain corpus $D$5, with no supervised QA dataset. The corpus is first cleaned and split into token-balanced chunks using NLTK to preserve coherence. For each chunk $D$6, an LLM such as DeepSeek-v3 or GPT-4o is prompted to generate a single self-contained knowledge question, referred to as question-style meta knowledge:

$D$7

Each meta question is then paired with its original chunk. The model produces both internal and external answers and fuses them with the self-selection rule. The paper argues that the resulting fused answers can combine knowledge the model already knows well, knowledge present in the document, and cases where external text is noisy but the model can correct it.

A second synthetic dataset is generated for multi-step reasoning. Using relevance-based sampling, the system retrieves the top 10 related question-chunk pairs and constructs reasoning QA examples under three explicit reasoning types: deductive reasoning, inductive reasoning, and case-based reasoning. The stated goal is to simulate realistic domain reasoning rather than only memorize facts (He et al., 21 Aug 2025).

Training proceeds in two stages. Stage 1 is Selective Supervised Fine-Tuning (Selective SFT). The paper measures token uncertainty via entropy,

$D$8

with normalization by $D$9, and defines token-wise weights as

$A^E = P_M(Q,D)$0

where $A^E = P_M(Q,D)$1 if the token prediction is correct. The loss is

$A^E = P_M(Q,D)$2

This weighting scheme gives high weight to incorrect tokens and uncertainty-proportional weight to correct tokens. The intended effect is to reduce training effort on confident, already-mastered tokens and concentrate updates on uncertain long-tail knowledge.

Stage 2 applies GRPO, described as a critic-free reinforcement learning method, on the synthetic reasoning QA data. The reward combines answer correctness and format adherence: $A^E = P_M(Q,D)$3 gives +5 for a fully correct answer and 0 otherwise, while $A^E = P_M(Q,D)$4 gives +1 for following the required "> ......ANSWER" format, 0 for general formatting errors, and $A^E = P_M(Q,D)$5 if “ANSWER” appears multiple times. The paper notes a notation issue in the source text, but states that the intended meaning is that the total reward is the sum of the accuracy reward and the format reward.

4. Ablations and operating characteristics

The ablation study attributes performance gains to the interaction of three components: fused data, Selective SFT, and GRPO. Standard SFT on external data is reported to be weaker than Selective SFT; Selective SFT on fused data is better than Selective SFT on external-only data; only GRPO helps reasoning somewhat but underperforms the full pipeline; and full S2K performs best, indicating that knowledge fusion, selective fine-tuning, and GRPO are complementary (He et al., 21 Aug 2025).

The analysis of reasoning data generation shows a large difference between random sampling and relevance-based sampling. Random sampling yields 32.6 Avg@5 / 35.0 Cons@5 / 44.6 Pass@5, whereas relevance-based sampling yields 38.6 / 43.4 / 77.1. The paper uses this result to argue that the structure of synthetic reasoning supervision materially affects downstream domain QA performance.

Reward design is also treated as consequential. The best scheme combines stronger answer reward with format reward; answer-only training hurts formatting, while format reward alone cannot fully recover correctness. This is presented as evidence that S2K’s reasoning stage is not merely a generic RL add-on but a constrained optimization toward both correctness and structured reasoning traces.

A case study is used as qualitative support for the self-selection mechanism. In the reported example, the external source confuses M1/M2 macrophage associations, while the fused answer preserves accurate facts from the original document and corrects the erroneous part. The paper presents this as evidence that token-level internal-external competition can correct wrong external knowledge while keeping useful supplementary details.

5. Experimental evaluation and efficiency

The experimental setup uses Qwen2.5-instruct-7b as the base model and BM25 for retrieval baselines. Evaluation is conducted on three domain QA benchmarks: MedQA for medicine, using the MedQA-USMLE subset; JEC-QA for law, using JEC-QA-KD from AGIEval; and FinanceIQ for finance, using the standard test set. The reported metrics are Avg@5, Cons@5, and Pass@5. Baselines include zero-shot and few-shot prompting, RAG, Self-Ask, Self-RAG, SFT, PPO, GRPO, and domain-pretrained models such as BioMistral, MMed-Llama, OpenBioLLM, Saul-7B, LawChat, Lawyer-LLaMA, finance-Llama3, xunayuan-6B-chat, and CFGPT (He et al., 21 Aug 2025).

The main table reports the following S2K results:

Domain	Benchmark	S2K
Medicine	MedQA	38.6 Avg@5 / 43.4 Cons@5 / 77.1 Pass@5
Law	JECQA	26.2 / 27.7 / 43.6
Finance	FinanceIQ	25.8 / 27.7 / 73.4

These scores are reported as outperforming standard SFT, PPO, GRPO alone, and all RAG and prompting baselines. The paper also states that S2K is the best-performing method across all three domains among the QA enhancement baselines.

The efficiency claim is a major part of the framework’s positioning. The paper states that S2K matches or rivals domain-pretrained models while using 2–3 orders of magnitude less data. A concrete example is medical adaptation with only 0.04B tokens, contrasted with hundreds of billions for full continued pretraining. This cost argument is central to the framework’s definition as a low-cost domain adaptation method rather than a domain-pretraining surrogate.

6. Disambiguation, related selection paradigms, and limitations

The acronym S2K is overloaded in the research literature. In energy disaggregation, S2K-NMF denotes Sum-to-k constrained Non-negative Matrix Factorization for non-intrusive load monitoring, a method unrelated to domain-specific QA (Rahimpour et al., 2017). In the present context, Selct2Know refers specifically to internal-external knowledge self-selection for domain-specific question answering (He et al., 21 Aug 2025).

More broadly, S2K belongs to a family of methods that reformulate difficult NLP problems as selection problems. In "Learning to Select from Multiple Options" (Du et al., 2022), selection is cast as contextualized or parallel textual entailment over candidate options; in "Generative Knowledge Selection for Knowledge-Grounded Dialogues" (Sun et al., 2023), knowledge selection is reframed as generating snippet identifiers with a sequence-to-sequence model. This suggests a broader methodological pattern: replacing monolithic classification or generation with explicit decision mechanisms over structured candidate spaces. A plausible implication is that Selct2Know extends this pattern from option-level or snippet-level selection to token-level competition between internal and external knowledge sources.

The main limitation acknowledged by the paper is that S2K is currently oriented toward static domain knowledge. It is not specifically designed for rapidly changing information or real-time updates. The authors suggest that future work could integrate RAG more directly to improve adaptability to dynamic knowledge while preserving S2K’s broad coverage (He et al., 21 Aug 2025).

The framework is presented as particularly suitable when QA data is scarce, raw corpora are available, and the model already contains some latent domain knowledge. Within those constraints, S2K’s contribution is a self-selection-based, low-cost domain adaptation pipeline that uses synthetic question generation, fused knowledge supervision, entropy-weighted fine-tuning, and GRPO to improve domain-specific QA across medicine, law, and finance.