Paper deep dive
From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning
Zuocheng Ying, Yang Yang, Yumou Wu, Chuanbo Zhu, Jiarui Wang, Ziqi Wu, Jingming Cai, Junqing Yu, Zikai Song
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/20/2026, 4:47:26 AM
Summary
The paper introduces VAKE (Verifiable Activation of Parametric KnowledgE), a two-stage reinforcement learning framework designed to activate latent parametric knowledge in Large Language Models (LLMs) that is stored but inaccessible. The framework consists of a Priming stage, where an RL policy inserts verifiable bridging triples into a sparse retrieved subgraph to help a frozen answerer generate correct answers, and a Reasoning stage, where the policy learns to perform implicit chain-of-thought reasoning to answer directly from the original input. Experiments on Qwen2.5-7B and Qwen3-8B models across seven benchmarks demonstrate that VAKE outperforms standard baselines and transfers effectively to out-of-distribution datasets, proving that it activates latent knowledge rather than copying context.
Entities (11)
Relation Signals (8)
VAKE → consistsof → Priming
confidence 95% · VAKE consists of Priming and Reasoning.
VAKE → consistsof → Reasoning
confidence 95% · VAKE consists of Priming and Reasoning.
Priming → uses → bridging triples
confidence 92% · In Priming, an RL-optimized policy elicits latent parametric knowledge as relational triples and inserts them into a sparse retrieved subgraph.
VAKE → evaluatedon → Qwen2.5-7B-Instruct
confidence 90% · We use Qwen2.5-7B-Instruct [26] and Qwen3-8B [37] with consistent hyperparameters.
VAKE → evaluatedon → Qwen3-8b
confidence 90% · We use Qwen2.5-7B-Instruct [26] and Qwen3-8B [37] with consistent hyperparameters.
Reasoning → uses → GRPO
confidence 90% · In Reasoning, the policy initialized from Priming is optimized via GRPO to answer directly from the original retrieved subgraph
VAKE → outperforms → standard baselines
confidence 88% · Experiments across seven benchmarks and models from 3B to 14B show that VAKE consistently outperforms standard baselines
VAKE → transfersto → OOD datasets
confidence 85% · including when transferring directly from HotpotQA to OOD datasets.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Although Large Language Models (LLMs) encode rich factual knowledge in their parameters, reliably recalling and verifying such knowledge remains a key bottleneck in factual question answering. Existing end-to-end methods entangle knowledge elicitation with reasoning, making it difficult to determine whether correct answers arise from parametric knowledge or the input context. To address this challenge, we propose VAKE (Verifiable Activation of Parametric KnowledgE), a two-stage reinforcement-learning framework that externalizes latent parametric knowledge through explicit Priming and transfers the acquired elicitation capability to implicit Reasoning. Given a query and an insufficient retrieved subgraph, the Priming policy explicitly inserts bridging triples as verifiable evidence, with supervision provided by rewards derived from answers generated by a separate frozen model over the augmented subgraph. Building on the policy learned during Priming, the Reasoning stage trains the model to answer from the original input, testing whether the capability acquired through explicit knowledge elicitation transfers to implicit reasoning. Experiments across seven benchmarks and models from 3B to 14B show that VAKE consistently outperforms standard baselines, including when transferring directly from HotpotQA to OOD datasets. LLM-based evaluation further shows that over 80% of the inserted triples provide factual bridging knowledge not derivable from the retrieved context, while more than half elicit knowledge inaccessible through direct prompting. These results suggest that VAKE activates latent parametric knowledge rather than copying the input context or memorizing dataset-specific associations.
Tags
Links
- Source: https://arxiv.org/abs/2608.18581v1
- Canonical: https://arxiv.org/abs/2608.18581v1
Trouble viewing inline? Open PDF directly →
Full Text
45,712 characters extracted from source content.
Expand or collapse full text
1]ByteDance 2]Huazhong University of Science and Technology [⋆]Equal contribution [†]Corresponding author Yang at , Zikai Song at From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning Zuocheng Ying Yang Yang Yumou Wu Chuanbo Zhu Jiarui Wang Ziqi Wu Jingming Cai Junqing Yu Zikai Song Affiliation: [ Affiliation: [ Email: yang.yves@bytedance.com Email: skyesong@hust.edu.cn Abstract Although Large Language Models (LLMs) encode rich factual knowledge in their parameters, reliably recalling and verifying such knowledge remains a key bottleneck in factual question answering. Existing end-to-end methods entangle knowledge elicitation with reasoning, making it difficult to determine whether correct answers arise from parametric knowledge or the input context. To address this challenge, we propose VAKE (Verifiable Activation of Parametric KnowledgE), a two-stage reinforcement-learning framework that externalizes latent parametric knowledge through explicit Priming and transfers the acquired elicitation capability to implicit Reasoning. Given a query and an insufficient retrieved subgraph, the Priming policy explicitly inserts bridging triples as verifiable evidence, with supervision provided by rewards derived from answers generated by a separate frozen model over the augmented subgraph. Building on the policy learned during Priming, the Reasoning stage trains the model to answer from the original input, testing whether the capability acquired through explicit knowledge elicitation transfers to implicit reasoning. Experiments across seven benchmarks and models from 3B to 14B show that VAKE consistently outperforms standard baselines, including when transferring directly from HotpotQA to OOD datasets. LLM-based evaluation further shows that over 80% of the inserted triples provide factual bridging knowledge not derivable from the retrieved context, while more than half elicit knowledge inaccessible through direct prompting. These results suggest that VAKE activates latent parametric knowledge rather than copying the input context or memorizing dataset-specific associations. 1 Introduction To equip large language models (LLMs) with general world knowledge and usable reasoning ability, researchers adopt a standard two-stage training pipeline. Pre-training encodes vast factual knowledge into LLMs’ parameters, while post-training methods, such as supervised fine-tuning (SFT) and reinforcement learning (RL), adapt this parametric knowledge to task-specific behaviors and outputs [36, 41, 37]. Prior literature suggests that frontier models achieve near-saturated encoding of factual knowledge from encyclopedic corpora such as Wikipedia [3, 7]. However, a critical bottleneck remains: models may fail to recall factual knowledge within their parameters, an issue defined as “stored but inaccessible” [44, 32]. Recall-oriented evaluations further expose this limitation: models may fail to answer queries correctly even when the target facts are fully present in the training corpus [44, 22]. Prior analyses suggest that internal parametric representations store richer factual information than what greedy decoding can reliably elicit, yet fact-association recall remains fragile [32, 8]. Scaling model size and training data can alleviate encoding deficits, but yields limited gains in accessing latent parametric knowledge [3, 22]. Figure 1: VAKE activates inaccessible parametric knowledge. (a) Factual information is encoded in LLM parameters. (b) Traditional methods suffer from fact recall failure: encoded facts remain inaccessible during thinking, leading to ungrounded reasoning and a wrong answer. (c) VAKE activates the relevant factual information before thinking, enabling fact recall success and the correct answer. This limitation is difficult to diagnose with standard accuracy-based evaluations, which conflate recall failures with encoding failures despite requiring different interventions. Encoding failures motivate scaling or knowledge injection, whereas recall failures call for post-training methods that improve access to knowledge already encoded in model parameters. Existing methods use instruction tuning (See Figure 1) and reinforcement learning to improve knowledge utilization [38, 21]. However, their reliance on unstructured generation entangles knowledge elicitation with answer reasoning, making the effect of activation difficult to isolate [3, 22]. Crucially, eliciting latent parametric knowledge often requires external query-relevant cues to trigger the model’s fact-association recall. To address this, we draw inspiration from the “priming effect” in cognitive science: a phenomenon where exposure to specific cues activates latent memory representations, enhancing subsequent task performance. Transferring this to LLMs, we hypothesize that introducing an independent, cue-conditioned activation phase over an insufficient retrieved subgraph can awaken the model’s dormant parametric knowledge. To this end, we propose Verifiable Activation of Parametric KnowledgE (VAKE), a two-stage reinforcement-learning framework that separates explicit knowledge elicitation from direct answer reasoning. VAKE consists of Priming and Reasoning. In Priming, an RL-optimized policy elicits latent parametric knowledge as relational triples and inserts them into a sparse retrieved subgraph. A separate, frozen answerer then evaluates this augmented graph and provides outcome-based rewards. This insert-then-answer intervention ensures that any improvement in the answerer’s output is directly attributable to the explicitly injected triples, providing a verifiable reward signal for knowledge activation. In Reasoning, the policy initialized from Priming is optimized via GRPO to answer directly from the original retrieved subgraph without explicitly injected triples. This effectively internalizes the acquired elicitation capability into implicit chain-of-thought (CoT) reasoning. Experiments show that VAKE consistently outperforms non-retrieval activation baselines, standard GRPO, and prior RL-based activation methods across model scales. Models optimized on a single multi-hop dataset transfer effectively to diverse out-of-distribution (OOD) multi-hop and single-hop datasets. Applying GRPO after Priming yields further gains, demonstrating the complementarity of knowledge activation and reasoning optimization. Therefore, VAKE can be seamlessly integrated into existing post-training pipelines as an RL-based complement and SFT-based knowledge activation. In summary, our main contributions are as follows: • A two-stage framework that separates knowledge activation from direct answer reasoning. Priming explicitly elicits latent parametric knowledge as relational triples, while Reasoning transfers the learned elicitation capability to direct question answering. • An observable and attributable representation of activated knowledge. Relational triples provide discrete and inspectable evidence for evaluating the content and effect of knowledge activation. • Compatibility with standard post-training paradigms. Empirically validated by our ablation results, VAKE serves as an RL-based complement to standard knowledge activation and seamlessly integrates into standard post-training pipelines while retaining full compatibility with reasoning optimization. 2 Related Work 2.1 Access, Not Storage Factual QA is increasingly framed as an access problem rather than a storage one, because a model’s greedy output only lower-bounds what it encodes. Three lines of evidence support this framing. External behavior shows that direct prompting underestimates recall [44, 32], that models answer inconsistently across paraphrases [12, 6], and that fine-tuning on new facts can even hurt existing recall [9]. Internal probing shows that hidden states carry truthfulness signals absent from the decoded answer [2, 1, 16, 23] and that internal knowledge measurably exceeds what the model explicitly generates [8]. Controlled provenance further shows that an injected corpus can make a fact provably parametric yet still leave the model unable to surface it [22, 24, 18]. We therefore study multi-hop QA instances limited by access to encoded knowledge, and seek an activation procedure that improves such access. 2.2 Eliciting Activation and Its Confounds Existing activation methods fall into two families, and both leave the same measurement confound in which the effect of activation cannot be cleanly separated from reasoning or retrieval. Inference-time methods produce intermediate content at query time through prompt search [28, 46], RECITE [29], Self-Ask [25] and Step-Back [45], all built on chain-of-thought and generated-knowledge prompting [33, 17, 31]. Training-time methods use reinforcement learning, where a correctness reward reweights knowledge the model holds rather than new facts [38], and token-level and distributional analyses take the same view that RLVR sharpens the base distribution rather than expanding it [21, 40, 35]. Related RL variants either route access through a reasoning chain [7, 19, 4] or combine access with side objectives such as factuality [27, 15, 34]. Because both families operate over free-form text and provide no controlled context, activation can only be inferred from output changes rather than directly attributed. We take a route that removes this confound. We provide the model with an incomplete retrieved subgraph that offers relevant cues without a sufficient answer path, and keep the answerer frozen during Priming, so the inserted bridging triples are the only free variable and their answer-enabling effect is observable and attributable. We treat the subgraph as a medium for query-conditioned augmentation rather than an external structure to traverse, and learn the Priming policy with reinforcement learning. 3 Method 3.1 Problem Formulation Let M0M_0 be a frozen model whose parameters encode knowledge that may not be directly accessible for a given query. For a question q with ground-truth answer a⋆a and an insufficient context c, our goal is to elicit the required knowledge from model parameters and express it as an explicit, query-conditioned increment ℐI. A parameterized policy πθ _θ generates this increment as ℐ∼πθ(⋅∣q,c).I _θ(· q,c). (1) We define successful activation by the following model-relative intervention: M0(q,c)≠a⋆,M0(q,c∪ℐ)=a⋆.M_0(q,c)≠ a , M_0(q,c )=a . (2) Here, the knowledge is parametric in source but becomes explicit in ℐI. Unlike retrieval augmentation, ℐI is generated by the policy rather than retrieved from an external source. Its attribution and answer-enabling effect are evaluated through the controlled intervention above. We implement this activation via two stages: Priming and Reasoning. Drawing on the cognitive notion of priming, where cues facilitate access to latent information, Priming trains the policy to elicit implicit parametric knowledge. We externalize the elicited knowledge as answer-enabling bridging triples ℐI, making activation observable and operationally attributable. Reasoning then trains the policy to perform CoT reasoning that exploits this elicitation capability and corrects hallucinations, enabling direct question answering without explicit triple insertion. 3.2 Task Definition We instantiate the parametric-knowledge activation task defined above on a triple graph =(s,r,o)G=\(s,r,o)\ parsed offline from the associated documents. In the main protocol, raw documents are used only for graph construction and are not exposed during training or standard inference; the model observes only triples. A retriever returns a query-relevant subgraph ℛ(q)⊆R(q) , which serves as the context c=ℛ(q)c=R(q). A few diagnostic experiments explicitly add raw documents to Priming, but the answerer always receives only triples. The retriever is designed to be shallow and sparse, providing query-relevant cues that can condition the elicitation of parametric knowledge rather than a complete supporting path. The concrete retrieval configuration is described in Sec. 4.1. Figure 2: Overview of our proposed VAKE. Given a sparse query-relevant subgraph ℛ(q)R(q), VAKE first performs explicit priming by learning to insert verifiable bridging triples ℐI for a frozen answerer M0M_0 (Stage I). It then transfers the learned knowledge-elicitation capability to implicit CoT reasoning through GRPO in the Reasoning stage (Stage I). 3.3 Two-Stage Activation: Priming and Reasoning The two stages operate on the same input (q,ℛ(q))(q,R(q)) and optimize the same answer-quality objective with respect to a⋆a , but differ in how the answer is produced. In the Priming stage, the policy inserts explicit bridging triples as intermediate evidence for a frozen answerer. In the Reasoning stage, the Priming-initialized policy generates the answer directly from the original input, transferring the learned knowledge-elicitation capability to direct reasoning. The two stages are optimized sequentially. Figure 2 gives an overview of this two-stage activation process. 3.3.1 stage I: Priming (insert-then-answer) In this stage, the policy πθ _θ acts as a subgraph-augmentation policy. Given the query and the retrieved subgraph, it externalizes bridging facts as a set of inserted triples ℐ∼πθ(⋅∣q,ℛ(q)),I _θ(· q,R(q)), (3) after which a frozen answerer M0M_0 produces a^=M0(q,ℛ(q)∪ℐ). a=M_0 (q,R(q) ). (4) This insert-then-answer intervention makes knowledge activation observable, attributable, and rewardable. With M0M_0 and ℛ(q)R(q) held fixed, ℐI is the only free variable; a transition from an incorrect answer without ℐI to a correct one with ℐI therefore isolates the effect of the inserted knowledge. Attribution is further supported by comparing each inserted triple against ℛ(q)R(q): when the decisive evidence is absent and non-derivable from the retrieved subgraph, the gain cannot be explained by copied context and is attributed to knowledge elicited from the policy’s parameters. Priming thus needs no SFT-style labels for inserted triples; its reward comes only from whether the downstream frozen answerer uses them to produce a⋆a . 3.3.2 stage I: Reasoning The insert-then-answer decomposition in stage I is an instrument for making activation attributable, rather than the intended final usage mode of the model. After Priming, we return the policy to its native question-answering role. Initialized from the stage I policy, θ←θ(1),θ←θ^(1), (5) the model now answers directly, a^∼πθ(⋅∣q,ℛ(q)), a _θ(· q,R(q)), (6) without an inserted set or a frozen answerer, and is optimized by GRPO with the same input setting and reward design as Priming, but with a stage-specific format-shaping term. This stage tests whether the explicit activation induced by Priming can be carried into chain-of-thought answering. This second stage also tests whether gains from Priming are complementary to reasoning gains. If Priming merely reproduces the effect of standard chain-of-thought optimization, continuing with reasoning-oriented GRPO should yield limited additional improvement. Additional gains instead indicate complementarity under sequential optimization. We therefore report Priming followed by Reasoning as a core result rather than as a mere ablation, since it examines whether Priming can be incorporated before reasoning-oriented GRPO in existing post-training pipelines. 3.4 Reinforcement Learning Objective Both stages in Section 3 are trained by reinforcement learning and optimized with the same GRPO algorithm. They share an answer-quality reward against the gold answer a⋆a , but differ in how a a is produced and which output format is required. We keep the reward outcome-driven but shape it into a small number of interpretable components rather than a single hard 0/10/1 signal. For stage s, the overall reward is: r(s)(a^,a⋆,o)=rqa(a^,a⋆)+rfmt(s)(o).r^(s)( a,a ;o)\;=\;r_qa( a,a )\;+\;r_fmt^(s)(o). (7) The answer-quality term rqar_qa is a soft blend of exact match (EM) and token-level F1 between a a and a⋆a , weighted so that exact correctness dominates while the F1 component densifies the reward landscape and mitigates the sparsity of a pure 0/10/1 signal on hard multi-hop questions. The stage-specific format-shaping term rfmt(s)r_fmt^(s) depends only on whether the trajectory o is well formed: it grants a small bonus when the output follows the required triple or reasoning format, and imposes a penalty when it is malformed, with severe malformation overriding the quality term so that unparseable trajectories cannot be rewarded. The two stages instantiate a a differently. In Priming, the trainable policy emits an inserted triple set and the answer is produced by the frozen answerer, a^=M0(q,ℛ(q)∪ℐ), a=M_0 (q,R(q) ), (8) so the reward reaches the policy only through ℐI. In Reasoning, the policy answers directly, a^∼πθ(⋅∣q,ℛ(q)), a _θ(· q,R(q)), (9) and the reward scores its own output. The concrete weights, bonuses, and penalty magnitudes in Eq. (7) are detailed in the Supplementary Material. 3.4.1 Group-Relative Policy Optimization We optimize both Priming and Reasoning with GRPO. For each question q, the old policy πθold _ _old samples a group of G trajectories oii=1G\o_i\_i=1^G. In Priming, each trajectory specifies inserted bridging triples, and the frozen answerer M0M_0 predicts a^i a_i from the subgraph after triple insertion. In Reasoning, the trajectory directly contains the reasoning trace and final answer. Each trajectory receives the stage-specific reward ri=r(s)(a^i,a⋆,oi)r_i=r^(s)( a_i,a ;o_i). Following GRPO, we compute the group-relative advantage by normalizing rewards within the sampled group: A^i=ri−mean(rjj=1G)std(rjj=1G)+δ, A_i= r_i-mean(\r_j\_j=1^G)std(\r_j\_j=1^G)+δ, (10) where δ is a small constant for numerical stability. The same sequence-level advantage is assigned to all tokens in oio_i. Let ρi,t(θ) _i,t(θ) be the token-level probability ratio between πθ _θ and πθold _ _old, and define ci,t(θ)=minρi,t(θ)A^i,clipε(ρi,t(θ))A^i.c_i,t(θ)= \! \ _i,t(θ) A_i,\;clip_ \! ( _i,t(θ) ) A_i \. (11) The GRPO objective is (θ)=[1G∑i=1G1|oi|∑t=1|oi|(ci,t(θ)−βDi,tKL)],J(θ)=E\! [ 1G _i=1^G 1|o_i| _t=1^|o_i| (c_i,t(θ)-β D_i,t^KL ) ], (12) where ε controls the clipping range, β weights the KL penalty, and Di,tKLD_i,t^KL denotes the token-level KL penalty to the stage-specific reference policy. Gradients are propagated only through πθ _θ. During Priming, M0M_0 is fixed and non-differentiable, contributing only through the reward induced by its generated answer. Thus, optimization learns to insert more useful bridging triples without adapting the answerer to compensate for weak insertions. During Reasoning, no separate answerer is used, and the policy is rewarded directly by the quality of its generated answer. 4 Experiments 4.1 Experimental Setup In-Distribution Multi-Hop QA Out-of-Distribution QA Method 2Wk† Hotp† MuS† Avg Bam† NQ⋆ Triv⋆ Pop⋆ Avg Qwen2.5-7B-Instruct Base 27.2 32.7 13.4 24.4 12.0 28.7 51.6 18.0 27.6 Base ♢ 25.2 33.2 13.6 24.0 21.6 28.3 51.2 17.8 29.7 Self-Ask 29.8 38.6 18.4 28.9 41.6 30.6 55.4 18.0 36.4 RECITE 30.7 36.3 17.6 28.2 40.8 34.8 57.4 19.8 38.2 Unlock 29.4 35.3 16.6 27.1 12.8 28.5 52.0 18.3 27.9 GRPO 32.5 35.4 16.2 28.0 12.8 28.7 51.8 18.2 27.9 GRPO ♢ 33.5 31.9 15.8 27.1 30.4 28.0 52.5 18.9 32.5 VAKE-P ♠ 33.1 40.9 20.0 31.3 32.0 30.8 54.0 18.8 33.9 VAKE ♠ ♢ 36.2 42.3 21.0 33.2 34.4 31.1 56.2 20.1 35.5 VAKE ♢ 35.5 43.4 20.4 33.1 41.6 33.2 59.5 21.2 38.9 Qwen3-8B Base 28.2 36.2 14.8 26.4 14.4 30.8 50.9 19.6 28.9 Base ♢ 26.5 33.5 18.4 26.1 40.8 33.8 60.1 22.3 39.3 Self-Ask 34.8 38.5 19.8 31.0 32.8 27.9 53.0 18.5 33.1 RECITE 33.0 40.1 19.6 30.9 33.6 35.3 58.5 21.5 37.2 Unlock 31.4 38.5 21.6 30.5 15.2 32.0 51.9 20.3 29.9 GRPO 34.2 36.8 20.2 30.4 16.8 35.7 61.4 22.5 34.1 GRPO ♢ 36.2 41.8 19.0 32.3 40.8 34.1 61.2 22.4 39.6 VAKE-P ♠ 35.0 42.3 22.4 33.2 32.0 31.3 59.7 21.8 36.2 VAKE ♠ ♢ 37.1 43.7 23.8 34.9 36.0 34.4 61.0 22.2 38.4 VAKE ♢ 37.9 44.9 20.6 34.5 43.2 37.4 62.3 22.5 41.4 Table 1: Answer accuracy (%) on ID and OOD QA benchmarks. † and ⋆ indicate multi-hop and single-hop QA datasets, respectively. The badges denote the test-time pipeline: ♠ = bridging-triple insertion and ♢ = CoT reasoning. For each backbone group, best and second-best results are highlighted. Datasets. We evaluate on two groups. (i) In-distribution: three multi-hop datasets, 2WikiMultihopQA [11], HotpotQA [39], and MuSiQue [30], used for both training and testing. (i) Out-of-distribution (OOD): four held-out datasets used only for evaluation: the multi-hop QA dataset Bamboogle [25], and three single-hop QA datasets, NQ [14], TriviaQA [13], and PopQA [20]. All OOD evaluations use the checkpoint trained on HotpotQA. Backbones. We use Qwen2.5-7B-Instruct [26] and Qwen3-8B [37] with consistent hyperparameters. VAKE-P denotes the Priming-only checkpoint, and VAKE denotes the checkpoint after both Priming and Reasoning. At test-time, VAKE uses either direct or insert-then-answer inference, while VAKE-P uses insert-then-answer inference with a frozen untrained answerer. Baselines. All baselines take the question and the same retrieved subgraph, without extra retrieval, in three families. Direct answering: Base answers directly and Base with Reasoning adds CoT, prompted for Qwen2.5-7B and native for Qwen3-8B. Inference-time methods: Self-Ask [25] and RECITE [29]. Training-time (RL) methods: Unlock [38] applies GRPO without insertion using a 72B judge, and GRPO is a standard baseline for Qwen2.5-7B under the same RL budget, evaluated with direct answering or CoT. Graph construction and retrieval. We build G per dataset with an LLM-based OpenIE pipeline that extracts entities from each passage and then extracts triples conditioned on those entities. For a question q, we anchor its mentions to nodes in G by string matching and expand a two-hop BFS neighborhood keeping the top five edges per hop by cosine similarity to q, without reranking or query rewriting. We then drop any triple containing the gold answer and cap the remaining union at K=10K=10 triples to form ℛ(q)R(q), leaving it insufficient by construction. Evaluation metrics. Unless otherwise stated, tables report judge-EM: given q, the prediction a a, the reference a⋆a , and any dataset-provided aliases, a GPT-4o judge (temperature 0) outputs a binary semantic-equivalence decision, accepting surface variants but rejecting entity substitutions or factual drift. The judge sees neither the retrieved subgraph nor the inserted triples, decoupling scoring from the intervention. The prompt and rubric are in the Supplementary Material. Implementation Details. For each in-distribution dataset, we train VAKE with about 14K training examples and 200 validation examples. Each run uses one epoch, a batch size of 28, and a learning rate of 1×10−61× 10^-6, yielding roughly 500 training steps. The policy samples 8 priming trajectories per step at temperature 0.7, with maximum prompt and response lengths of 2048 tokens. The answer-quality reward weights exact match and token-level F1 by 0.6 and 0.4, respectively; malformed outputs receive a −1.0-1.0 format penalty. 4.2 Main Results We compare VAKE with inference-time and training-time baselines on in-distribution and OOD QA benchmarks in Table 1, using matched test-time pipelines where applicable. Priming provides consistent gains With Priming alone and no reasoning at inference, VAKE-P attains the best in-distribution average on both backbones, exceeding the strongest non-reasoning baseline by 2.4 and 2.2 points on Qwen2.5-7B and Qwen3-8B. The largest dataset-level margins appear on HotpotQA, where VAKE-P exceeds the strongest non-reasoning baseline by 2.3 and 2.2 points on the two backbones, respectively. Since VAKE-P keeps the answerer frozen and changes only its input, the gains reflect the inserted bridging triples, not any adaptation of the answerer. Whether those triples supply retrievable facts or knowledge held in the model’s parameters is the question we take up in §4.4. Priming and Reasoning provide complementary gains. Under bridging-triple insertion with CoT at inference, VAKE outperforms VAKE-P across all in-distribution datasets and both backbones, raising the average by 1.9 points on Qwen2.5-7B and 1.7 on Qwen3-8B. This gain isolates the contribution of the Reasoning stage under a fixed test-time pipeline. In the contrasting CoT-only setting, with no triples inserted, VAKE surpasses GRPO, the strongest reasoning baseline, by 6.1 and 2.6 points on the two backbones. Because this gain arises without any inserted triples, it reflects reasoning improved by Priming rather than extra evidence supplied at inference. We test this complementarity more directly through controlled ablations in §4.5. VAKE generalizes across OOD tasks. VAKE achieves the best OOD average on both backbones, reaching 38.9 on Qwen2.5-7B and 41.4 on Qwen3-8B, and surpasses the strongest inference-time and RL baselines. The advantage is clearest on Qwen3-8B, where VAKE improves over the best competing method by 1.8 points on average and ranks first on all four OOD datasets. The largest gain appears on Bamboogle, the held-out multi-hop benchmark, where VAKE outperforms the strongest baseline by 2.4 points. This result suggests that the learned activation capability transfers beyond the training distribution and remains effective when accessing parametric knowledge across longer reasoning chains. 4.3 Robustness: Scaling and General Capabilities This section studies whether VAKE remains stable across backbone sizes without sacrificing broader model capabilities. For the scaling and input-condition analyses, we evaluate the VAKE-P checkpoint equipped with bridging-triple insertion. We further permit triple generation to be conditioned on raw documents, while the corresponding frozen base answerer still receives only the augmented subgraph. Method GSM8K IFEval AIME24 AIME25 MMLU AVG Base 92.04 75.99 33.33 23.33 70.84 59.11 Unlock 91.28 75.20 30.00 23.33 70.84 58.13 GRPO 92.42 76.52 23.33 30.00 70.62 58.58 VAKE-P 92.42 75.64 36.67 20.00 70.99 59.14 VAKE 91.74 76.82 30.00 26.67 70.87 59.22 Table 2: General capability evaluation on standard benchmarks. All scores are reported as percentages using each benchmark’s standard evaluation metric. The best and second-best results in each column are highlighted. Figure 3: Judge-EM of VAKE-P with bridging-triple insertion on 2Wiki across six Qwen backbones, using triples only (T) or triples with raw documents (T+D). Priming gains persist across backbone families and sizes. Figure 3 reports judge-EM on 2Wiki across six Qwen backbones under triples-only and document-augmented inputs for bridging-triple insertion. In the document-augmented setting, raw documents are available only to VAKE-P during triple generation; the corresponding frozen base answerer still predicts from the subgraph after triple insertion. VAKE-P outperforms Base in nearly all settings, except for the smallest Qwen3-0.6B model, where limited capacity and longer inputs may hurt triple-generation quality. The gains remain visible as model size increases, showing that stronger backbones still benefit from bridging-triple insertion. Adding raw documents further improves VAKE-P in several settings without exposing documents to the answerer, suggesting that extra context helps generate more effective bridging triples. VAKE preserves general capabilities. Table 2 spans three categories: math, instruction following, and general knowledge. We evaluate on GSM8K [5], AIME24/25 [42, 43], IFEval [47], and MMLU [10], reporting pass@10 on AIME24/25 and accuracy on the rest. VAKE-P and VAKE stay on par with Base across all three categories, with near-identical average scores, showing that the QA gains come at no cost to the evaluated general capabilities. 4.4 Attribution of the Activated Knowledge We next ask whether the inserted triples restate the retrieved subgraph or expose parametric knowledge.All analyses in this section are conducted on 2Wiki. Inserted triples mostly originate from parametric knowledge. For each backbone scale, we run LLM-based source attribution on the triples inserted by its own VAKE-P checkpoint. Given the retrieved subgraph and an inserted triple, the same GPT-4o judge labels the triple as subgraph-inferable if it can be derived from the retrieved subgraph, and as parameter-originated otherwise; the judge prompt is in the Supplementary Material. Across all scales, over 80% of inserted triples are judged parameter-originated, and fewer than 20% are attributed to the retrieved subgraph, as shown in Table 3. Thus, most insertions cannot be explained as restatements or derivations of the retrieved subgraph and are operationally attributed to parametric knowledge. Qwen2.5 Qwen3 Source 3B 7B 14B 0.6B 4B 8B Params 84.2 90.8 92.3 95.0 95.8 80.3 Subgraph 15.8 9.2 7.7 5.0 4.2 19.7 Table 3: Source attribution of VAKE-P-inserted triples across Qwen2.5/Qwen3 scales. Each cell reports the percentage of inserted triples attributed to model parameters or retrieved context. Probing reveals a gap between knowledge encoding and access. For the factually correct, parameter-originated triples inserted by VAKE-P (7B), we mask the head (S), relation (R), or tail (O) of each triple (s,r,o)(s,r,o) and measure reproduction accuracy. We compare VAKE-P with frozen same-family backbones from 7B to 72B (Figure 4). Frozen backbones achieve high head and relation accuracy, indicating that much of the relevant knowledge is encoded in their parameters. However, their tail accuracy stays below 27.0, whereas VAKE-P reaches 35.2 and outperforms even the 72B backbone. This contrast suggests that Priming improves access to encoded answer-bearing information. Figure 4: Probing masked-slot reproduction of VAKE-P triples on 2Wiki across VAKE-P(7B) and frozen backbones. 4.5 Ablation Studies We ablate VAKE on Qwen2.5-7B-Instruct along two axes: its training components and input conditions. Priming and Reasoning training are complementary. Table 4 ablates VAKE’s two training stages. All variants are evaluated without bridging-triple insertion; VAKE and w/o Priming use CoT reasoning. VAKE performs best on all datasets, while w/o Reasoning surpasses both Base and w/o Priming on HotpotQA and MuSiQue, showing that Priming benefits direct answering even without triple insertion. Variant 2Wiki HotpotQA MuSiQue VAKE 35.50 43.40 20.40 w/o Priming 33.50 31.90 15.80 w/o Reasoning 28.50 37.20 16.40 Base 27.20 32.70 13.40 Table 4: Method ablation on multi-hop QA. w/o Priming and w/o Reasoning remove the corresponding training stage. All variants are evaluated without bridging-triple insertion. Method Closed-book Triples Triples+Docs Base 24.20 27.60 30.70 VAKE-P 30.00 33.10 41.30 VAKE 30.10 36.20 46.60 Table 5: Information ablation on 2Wiki, with all methods evaluated via bridging-triple insertion. The learned activation capability generalizes across input settings. Table 5 compares Base, VAKE-P, and VAKE on 2Wiki across closed-book, triples-only, and document-augmented inputs. All methods use bridging-triple insertion, thereby isolating the effect of the trained checkpoints under the same test-time pipeline. VAKE-P and VAKE improve over Base in all three settings, confirming that the gains arise from training rather than triple insertion alone. The closed-book improvement further supports parametric knowledge elicitation, as bridging triples are generated from the question without retrieved context. Under document-augmented input, VAKE improves over Base by 15.9 points, showing that the learned activation capability also complements richer external context. 5 Conclusion In this paper, we frame factual question answering as a problem of accessing parametric knowledge rather than storing it, as encoded facts often remain implicit and inaccessible. We propose VAKE, a two-stage reinforcement-learning framework that separates explicit knowledge elicitation from direct answer reasoning. In Priming, the policy inserts bridging triples into a sparse retrieved subgraph while a frozen answerer provides outcome-based rewards, rendering the activated knowledge observable and its answer-enabling effect directly testable. In Reasoning, the elicitation capability learned during Priming transfers to direct chain-of-thought answering and combines with reasoning-oriented optimization. Experiments on two instruction-tuned backbones across seven multi-hop and single-hop benchmarks show that VAKE consistently outperforms inference-time and reinforcement-learning baselines and generalizes to OOD datasets. A scaling analysis across six Qwen backbones shows consistent activation gains at larger scales. Attribution analysis indicates that the inserted triples predominantly surface parametric knowledge inaccessible through direct prompting rather than restating retrieved context. VAKE can be integrated into existing post-training pipelines as an RL-based complement to QA-pair construction and SFT-based knowledge activation. Our ablations further show that Priming improves direct answering without test-time triple insertion, while the learned activation capability generalizes from closed-book to document-augmented settings. References [1] A. Azaria and T. Mitchell (2023) The internal state of an llm knows when it’s lying. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 967–976. Cited by: §2.1. [2] C. Burns, H. Ye, D. Klein, and J. Steinhardt (2022) Discovering latent knowledge in language models without supervision. arXiv preprint arXiv:2212.03827. Cited by: §2.1. [3] N. Calderon, E. Ben-David, Z. Gekhman, E. Ofek, and G. Yona (2026) Empty shelves or lost keys? recall is the bottleneck for parametric factuality. arXiv preprint arXiv:2602.14080. Cited by: §1, §1, §1. [4] X. Chen, I. Kulikov, V. Berges, B. Oğuz, R. Shao, G. Ghosh, J. Weston, and W. Yih (2025) Learning to reason for factuality. arXiv preprint arXiv:2508.05618. Cited by: §2.2. [5] K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, et al. (2021) Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168. Cited by: §4.3. [6] Y. Elazar, N. Kassner, S. Ravfogel, A. Ravichander, E. Hovy, H. Schütze, and Y. Goldberg (2021) Measuring and improving consistency in pretrained language models. Transactions of the Association for Computational Linguistics (TACL) 9, p. 1012–1031. Cited by: §2.1. [7] Z. Gekhman, R. Aharoni, E. Ofek, M. Geva, R. Reichart, and J. Herzig (2026) Thinking to recall: how reasoning unlocks parametric knowledge in llms. arXiv preprint arXiv:2603.09906. Cited by: §1, §2.2. [8] Z. Gekhman, E. Ben-David, H. Orgad, E. Ofek, Y. Belinkov, I. Szpektor, J. Herzig, and R. Reichart (2025) Inside-out: hidden factual knowledge in LLMs. In Conference on Language Modeling (COLM), Cited by: §1, §2.1. [9] Z. Gekhman, G. Yona, R. Aharoni, M. Eyal, A. Feder, R. Reichart, and J. Herzig (2024) Does fine-tuning llms on new knowledge encourage hallucinations?. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 7765–7784. Cited by: §2.1. [10] D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt (2020) Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300. Cited by: §4.3. [11] X. Ho, A. D. Nguyen, S. Sugawara, and A. Aizawa (2020) Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, p. 6609–6625. Cited by: §4.1. [12] Z. Jiang, F. F. Xu, J. Araki, and G. Neubig (2020) How can we know what language models know?. Transactions of the Association for Computational Linguistics (TACL) 8, p. 423–438. Cited by: §2.1. [13] M. Joshi, E. Choi, D. S. Weld, and L. Zettlemoyer (2017) Triviaqa: a large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 1601–1611. Cited by: §4.1. [14] T. Kwiatkowski, J. Palomaki, O. Redfield, M. Collins, A. Parikh, C. Alberti, D. Epstein, I. Polosukhin, J. Devlin, K. Lee, et al. (2019) Natural questions: a benchmark for question answering research. Transactions of the Association for Computat ional Linguistics 7, p. 453–466. Cited by: §4.1. [15] J. Li and H. T. Ng (2025) Reasoning models hallucinate more: factuality-aware reinforcement learning for large reasoning models. In Advances in Neural Information Processing Systems (NeurIPS), External Links: Link Cited by: §2.2. [16] K. Li, O. Patel, F. Viégas, H. Pfister, and M. Wattenberg (2023) Inference-time intervention: eliciting truthful answers from a language model. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 36, p. 41451–41530. Cited by: §2.1. [17] J. Liu, A. Liu, X. Lu, S. Welleck, P. West, R. Le Bras, Y. Choi, and H. Hajishirzi (2022) Generated knowledge prompting for commonsense reasoning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), Cited by: §2.2. [18] K. Liu, Z. Chen, Z. Fu, W. Zhang, R. Jiang, F. Zhou, Y. Chen, Y. Wu, and J. Ye (2025) Structure-aware domain knowledge injection for large language models. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL)", p. 29443–29464. Cited by: §2.1. [19] M. Ma and J. Hewitt (2026) Improving parametric knowledge access in reasoning language models. arXiv preprint arXiv:2602.22193. Cited by: §2.2. [20] A. Mallen, A. Asai, V. Zhong, R. Das, D. Khashabi, and H. Hajishirzi (2023) When not to trust language models: investigating effectiveness of parametric and non-parametric memories. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), p. 9802–9822. Cited by: §4.1. [21] H. Meng, K. Huang, S. Wei, C. Ma, S. Yang, X. Wang, G. Wang, B. Ding, and J. Zhou (2026) Sparse but critical: a token-level analysis of distributional shifts in rlvr fine-tuning of llms. arXiv preprint arXiv:2603.22446. Cited by: §1, §2.2. [22] S. M. Mousavi, S. Alghisi, and G. Riccardi (2026) What does loss optimization actually teach, if anything? knowledge dynamics in continual pre-training of llms. arXiv preprint arXiv:2601.03858. Cited by: §1, §1, §2.1. [23] H. Orgad, M. Toker, Z. Gekhman, R. Reichart, I. Szpektor, H. Kotek, and Y. Belinkov (2025) LLMs know more than they show: on the intrinsic representation of LLM hallucinations. In International Conference on Learning Representations (ICLR), Cited by: §2.1. [24] O. Ovadia, M. Brief, R. Lemberg, and E. Sheetrit (2025) Knowledge-instruct: effective continual pre-training from limited data using instructions. arXiv preprint arXiv:2504.05571. Cited by: §2.1. [25] O. Press, M. Zhang, S. Min, L. Schmidt, N. Smith, and M. Lewis (2023) Measuring and narrowing the compositionality gap in language models. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 5687–5711. Cited by: §2.2, §4.1, §4.1. [26] Qwen, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, et al. (2025) Qwen2.5 technical report. External Links: 2412.15115, Link Cited by: §4.1. [27] B. Ren, S. Qiao, N. Zhang, D. Zheng, and H. Chen (2026) KnowRL: exploring knowledgeable reinforcement learning for factuality. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), p. 39640–39658. Cited by: §2.2. [28] T. Shin, Y. Razeghi, R. L. L. Iv, E. Wallace, and S. Singh (2020) Autoprompt: eliciting knowledge from language models with automatically generated prompts. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 4222–4235. Cited by: §2.2. [29] Z. Sun, X. Wang, Y. Tay, Y. Yang, and D. Zhou (2023) Recitation-augmented language models. In International Conference on Learning Representations (ICLR), Cited by: §2.2, §4.1. [30] H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal (2022) MuSiQue: multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics 10, p. 539–554. Cited by: §4.1. [31] X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou (2023) Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations (ICLR), External Links: Link Cited by: §2.2. [32] Y. Wang, Y. Chen, W. Wen, Y. Sheng, L. Li, and D. D. Zeng (2024) Unveiling factual recall behaviors of large language models through knowledge neurons. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 7388–7402. Cited by: §1, §1, §2.1. [33] J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. V. Le, and D. Zhou (2022) Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 35, p. 24824–24837. Cited by: §2.2. [34] Z. Wei, X. Yang, K. Sun, J. Wang, R. Shao, J. Chen, M. Kachuee, T. Gollapudi, Y. Liao, N. Scheffer, et al. (2025) Truthrl: incentivizing truthful llms via reinforcement learning. arXiv preprint arXiv:2509.25760. Cited by: §2.2. [35] X. Wen, Z. Liu, S. Zheng, S. Ye, Z. Wu, Y. Wang, Z. Xu, X. Liang, J. Li, Z. Miao, J. Bian, and M. Yang (2026) Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base LLMs. In International Conference on Learning Representations (ICLR), External Links: Link Cited by: §2.2. [36] A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Ling, et al. (2026) DeepSeek-v4: towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348. Cited by: §1. [37] A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. (2025) Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: §1, §4.1. [38] W. Yang, H. Zang, J. Zhang, W. Shi, D. Su, J. Wang, X. Cheng, and F. Sun (2026) Beyond reasoning: reinforcement learning unlocks parametric knowledge in llms. Note: arXiv preprint arXiv:2605.07153 External Links: 2605.07153, Link Cited by: §1, §2.2, §4.1. [39] Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning (2018) HotpotQA: a dataset for diverse, explainable multi-hop question answering. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 2369–2380. Cited by: §4.1. [40] Y. Yue, Z. Chen, R. Lu, A. Zhao, Z. Wang, Y. Yue, S. Song, and G. Huang (2025) Does reinforcement learning really incentivize reasoning capacity in LLMs beyond the base model?. In Advances in Neural Information Processing Systems (NeurIPS), External Links: Link Cited by: §2.2. [41] A. Zeng, X. Lv, Z. Hou, Z. Du, Q. Zheng, B. Chen, D. Yin, C. Ge, C. Huang, C. Xie, et al. (2026) GLM-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763. Cited by: §1. [42] Y. Zhang and T. Math-AI (2024) American invitational mathematics examination (aime) 2024. Cited by: §4.3. [43] Y. Zhang and T. Math-AI (2025) American invitational mathematics examination (aime) 2025. Cited by: §4.3. [44] D. Zheng, M. Lapata, and J. Z. Pan (2024) How reliable are llms as knowledge bases? re-thinking facutality and consistency. arXiv preprint arXiv:2407.13578. Cited by: §1, §1, §2.1. [45] H. S. Zheng, S. Mishra, X. Chen, H. Cheng, E. H. Chi, Q. V. Le, and D. Zhou (2024) Take a step back: evoking reasoning via abstraction in large language models. In International Conference on Learning Representations (ICLR), Cited by: §2.2. [46] Z. Zhong, D. Friedman, and D. Chen (2021) Factual probing is [mask]: learning vs. learning to recall. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), p. 5017–5033. Cited by: §2.2. [47] J. Zhou, T. Lu, S. Mishra, S. Brahma, S. Basu, Y. Luan, D. Zhou, and L. Hou (2023) Instruction-following evaluation for large language models. arXiv preprint arXiv:2311.07911. Cited by: §4.3.