Source-linked AI summary
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
TL;DR
LLMs often store facts they cannot reliably access, so VAKE separates explicit knowledge elicitation from answer reasoning and reports further gains when reasoning optimization follows Priming.
Problem
LLMs may encode richer factual knowledge than they can reliably elicit, while standard evaluations conflate recall failures with encoding failures.
Method
VAKE uses Priming to insert inspectable bridging triples into sparse retrieved subgraphs, then transfers the elicitation capability to direct Reasoning.
Results
GRPO after Priming yields further gains, demonstrating complementarity between knowledge activation and reasoning optimization.
Takeaways & Limitations
VAKE makes activated knowledge observable and its answer-enabling effect directly testable while separating elicitation from reasoning.
Takeaways & Limitations
VAKE assumes a frozen model whose parameters encode knowledge that may be inaccessible for a given query.
Abstract
from arXiv · showhide
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
LLMs can encode extensive factual knowledge yet fail to reliably recall it, and standard evaluations conflate failures of knowledge encoding with failures of access. VAKE addresses this bottleneck with a two-stage framework that explicitly elicits inspectable parametric knowledge before transferring that capability to direct reasoning.
- Motivation: LLMs may fail to recall factual knowledge stored in their parameters, even when target facts are present in the training corpus.This “stored but inaccessible” problem reflects a gap between richer internal representations and what decoding reliably elicits.
- Motivation: Standard accuracy evaluations conflate encoding failures with recall failures, although the two require different interventions.Scaling and knowledge injection address encoding deficits, whereas post-training methods target access to already encoded knowledge.
- VAKE framework: VAKE uses two reinforcement-learning stages: Priming inserts elicited relational triples into a sparse retrieved subgraph, while Reasoning transfers this capability to direct question answering.A separate frozen answerer evaluates the augmented graph and supplies outcome-based rewards during Priming.
- Compatibility: VAKE integrates with standard post-training pipelines as an RL-based complement while retaining compatibility with reasoning optimization.The framework separates knowledge activation from direct answer reasoning and supports combining activation with further reasoning optimization.
- VAKE framework: Relational triples make activated knowledge discrete, observable, and attributable for evaluating its content and effect.This representation enables inspection of the knowledge elicited from the model’s parameters.
2 Related Work
Related work frames factual QA as an access problem and identifies activation methods whose reasoning and retrieval entanglement prevents clean measurement. VAKE addresses this confound with incomplete retrieved subgraphs, a frozen answerer, and reinforcement-learned bridging-triple insertion.
- Access versus storage: Factual QA is increasingly framed as an access problem because greedy outputs lower-bound the knowledge encoded in model parameters.Direct prompting underestimates recall, answers vary across paraphrases, and fine-tuning new facts can harm existing recall.
- Activation methods: Existing activation methods use inference-time prompt search, RECITE, Self-Ask, and Step-Back, but entangle activation effects with reasoning or retrieval.These approaches build on chain-of-thought and generated-knowledge prompting.
- Activation methods: Training-time reinforcement-learning methods reweight existing knowledge rather than adding facts, while related variants route access through reasoning chains or side objectives.Token-level and distributional analyses similarly characterize RLVR as sharpening the base distribution rather than expanding it.
- VAKE’s distinction: VAKE removes this measurement confound by supplying an incomplete retrieved subgraph and keeping the answerer frozen during Priming.The inserted bridging triples are therefore the only free variable, making their answer-enabling effect observable and attributable.
- VAKE’s distinction: VAKE treats the subgraph as a medium for query-conditioned augmentation rather than an external structure to traverse, and learns Priming with reinforcement learning.The subgraph provides relevant cues without a sufficient answer path.
3 Method
VAKE activates parametric knowledge through a two-stage reinforcement-learning procedure: Priming makes query-conditioned bridging knowledge explicit and attributable, while Reasoning transfers this capability to direct answering. Both stages optimize answer quality with GRPO but differ in whether a frozen answerer consumes inserted triples or the policy answers directly.
- Priming: Priming inserts verifiable bridging triples into the retrieved subgraph, then uses a frozen answerer to produce the answer.Holding the answerer and retrieved subgraph fixed makes the inserted triples the only free variable affecting the intervention.
- Activation objective: VAKE generates a query-conditioned increment I from model parameters, making parametric knowledge explicit without retrieving it from an external source.The policy produces I, whose attribution and answer-enabling effect are evaluated through a controlled intervention.
- Reasoning: Reasoning initializes from the Priming policy and answers directly from the original input without inserted triples or a frozen answerer.This stage tests whether explicit activation transfers to chain-of-thought answering and whether it provides gains complementary to reasoning optimization.
- Priming: A correct answer appearing only after insertion isolates the effect of elicited knowledge, especially when decisive evidence is absent and non-derivable from the retrieved subgraph.Priming therefore requires no SFT-style labels for inserted triples; its reward depends on whether the frozen answerer uses them to produce the answer.
- Optimization: Both stages use GRPO with a shared answer-quality reward combining exact match and token-level F1, plus stage-specific format shaping.In Priming, reward reaches the policy through the frozen answerer’s answer; in Reasoning, it scores the policy’s direct output.
4 Experiments
VAKE is evaluated across in-distribution and held-out OOD QA benchmarks, where Priming and Reasoning provide complementary gains over matched baselines. Additional analyses show that the method preserves general capabilities and improves access to parameter-originated knowledge across input settings.
- Evaluation setup: VAKE is evaluated on three in-distribution multi-hop datasets and four held-out OOD datasets, with all OOD evaluations using a HotpotQA-trained checkpoint.The OOD set includes Bamboogle, NQ, TriviaQA, and PopQA.
- Priming gains: Priming alone improves the in-distribution average by 2.4 and 2.2 points over the strongest non-reasoning baseline on Qwen2.5-7B and Qwen3-8B, respectively.VAKE-P also has the largest dataset-level margins on HotpotQA, exceeding the strongest non-reasoning baseline by 2.3 and 2.2 points.
- Priming and Reasoning: Adding Reasoning raises VAKE-P’s in-distribution average by 1.9 points on Qwen2.5-7B and 1.7 on Qwen3-8B under the same bridging-triple insertion and CoT pipeline.Without inserted triples, VAKE surpasses GRPO by 6.1 and 2.6 points in the CoT-only setting.
- OOD generalization: VAKE achieves the best OOD averages, reaching 38.9 on Qwen2.5-7B and 41.4 on Qwen3-8B; on Qwen3-8B it ranks first on all four OOD datasets.On Qwen3-8B, VAKE improves over the best competing method by 1.8 points on average, with a 2.4-point gain on Bamboogle.
- Capability preservation: VAKE-P and VAKE remain on par with Base across math, instruction following, and general knowledge, with near-identical average scores.The evaluated benchmarks are GSM8K, AIME24/25, IFEval, and MMLU.
- Knowledge attribution: Over 80% of inserted triples are judged to originate from parametric knowledge rather than being inferable from the retrieved subgraph.Across backbone scales, source attribution labels triples as subgraph-inferable when derivable from the retrieved subgraph and parameter-originated otherwise.
- Ablations and input settings: VAKE outperforms Base, VAKE-P improves over Base, and both trained checkpoints improve across closed-book, triples-only, and document-augmented inputs under the same insertion pipeline.The closed-book improvement supports parametric knowledge elicitation, while ablations show VAKE performs best on all datasets and Priming benefits direct answering without triple insertion.
5 Conclusion
The paper frames factual question answering as accessing implicit parametric knowledge rather than merely storing it. VAKE addresses this through two-stage reinforcement learning that separates explicit knowledge elicitation from direct answer reasoning.
- VAKE treats factual question answering as a problem of accessing parametric knowledge that may remain implicit and inaccessible.
- VAKE uses two reinforcement-learning stages to separate explicit knowledge elicitation from direct answer reasoning.
- Priming: During Priming, the policy inserts bridging triples into a sparse retrieved subgraph, while a frozen answerer supplies outcome-based rewards.