acceptodds
Under review as a conference paper at ICLR 2027

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

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 (_**V**erifiable **A**ctivation of Parametric **K**nowledg**E**_), 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. Across seven benchmarks and two backbone families, VAKE achieves the best average performance on both the in-distribution multi-hop and OOD benchmark groups, outperforming inference-time methods and GRPO. Complementary attribution and probing analyses on 2Wiki suggest that the inserted triples primarily contribute bridging information not derivable from the retrieved context and can surface knowledge missed by direct prompting. These results support the view that VAKE activates latent parametric knowledge rather than copying the input context or memorizing dataset-specific associations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
ML: Post-Training, Fine-Tuning & Model Alignment

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xntBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VAKE, a two-stage RL framework for improving access to parametric knowledge in LLMs. In Priming, the model inserts bridging triples into a retrieved subgraph, and a frozen answerer provides the reward. In Reasoning, the model learns to answer directly using the capability developed during Priming. Experiments across seven QA benchmarks show consistent gains over existing prompting and RL baselines.

Critical strengths
  1. VAKE introduces a creative two-stage training framework that separates explicit knowledge elicitation from direct answer reasoning.

  2. The motivation is meaningful: LLMs may contain relevant facts in their parameters but fail to recall them when answering directly.

  3. The experiments are broad and support the paper’s main claim that VAKE improves access to parametric knowledge. The authors evaluate across multiple benchmarks and model scales, and include attribution analyses and ablations.

Critical weaknesses
  1. The main results show that VAKE outperforms Self-Ask, RECITE, Unlock, and GRPO, but the paper does not compare their inference costs. Since VAKE first generates triples and then reasons over them, reporting measures such as generated token length or computation cost would help assess its overall advantage.

  2. Table 3 shows that the share of inserted triples attributed to model parameters is 95.0%, 95.8%, and 80.3% for Qwen3-0.6B, 4B, and 8B, respectively. The drop for the largest model is unexpected and seems at odds with the common expectation that larger models store more factual knowledge. The authors should explain this result.

  3. VAKE’s training depends on the frozen answerer providing a reliable reward based on the augmented subgraph. However, even when the subgraph contains the necessary facts, the answerer may hallucinate or fail to answer correctly, resulting in a misleading negative reward. The paper does not evaluate how sensitive the framework is to this issue, which is important for understanding whether VAKE can be trained reliably.

Questions for authors

See above.

Reviewer_WRsROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VAKE, a two-stage reinforcement-learning framework for factual question answering with incomplete retrieved subgraphs. During Priming, a policy generates additional relational triples and receives an answer-quality reward from a frozen downstream answerer. During Reasoning, the Priming checkpoint is further optimized to answer from the original subgraph without explicit triple insertion. Experiments cover three in-distribution multi-hop datasets and four held-out datasets using Qwen2.5 and Qwen3 backbones. The paper reports improved average answer accuracy and investigates the learned behavior through source attribution, masked-slot probing, sampling, answer perplexity, and cross-model activation patching.

Critical strengths

S1. The paper discuss a meaningful question: how to help LLMs access knowledge they already possess. Its explicit Priming framework offers an interesting way to study this problem and improve factual question answering.

S2. By separating knowledge activation from answer reasoning and using a frozen answer generator with a fixed sparse subgraph, the approach makes the contribution of generated knowledge easier to inspect.

Critical weaknesses

W1. The Reasoning stage still relies on retrieved subgraphs, while the evaluation removes answer-containing triples to create incomplete evidence. This may mainly test missing-fact completion in a controlled setting. Comparisons with competitive RAG baselines and naturally incomplete retrieval would help clarify the method’s distinctive contribution and generality.

W2. Although the paper focuses on eliciting existing parametric knowledge, the final Reasoning stage still relies on retrieved subgraphs. It remains unclear whether the learned capability transfers to answering without external graph cues, limiting the demonstrated generality of the approach.

Questions for authors

Q1. How often do the generated triples contain the final answer? It is unclear whether Priming provides useful intermediate facts beyond simply passing candidate answers to the frozen answerer.

Q2. Since VAKE relies on externally retrieved subgraphs, how does it compare with standard RAG or graph-based RAG under comparable resource budgets? Such comparisons would help establish its practical benefits.

Q3. Why is the retrieved subgraph retained during the Reasoning stage? Evaluating the final direct-answering model with and without graph input would help clarify whether the learned capability still depends on external cues.

Reviewer_zdFVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VAKE (Verifiable Activation of Parametric KnowledgE), a two-stage RL framework for multi-hop QA that externalizes parametric knowledge through explicit Priming and transfers the acquired elicitation capability to Reasoning. The Priming stage trains a policy to recall bridging triples that are missing in a deliberately sparse knowledge graph. The Reasoning stage starts from the first stage's checkpoint and runs another round of GRPO to do direct CoT reasoning from original subgraph. On Qwen2.5-7B and Qwen3-8B, VAKE achieves the best average performance on both in-distribution (2Wiki/HotpotQA/MuSiQue) and OOD (Bamboogle/NQ/TriviaQA/PopQA) benchmarks. Ablations cover the two training stages and different input conditions; the appendix adds attribution, probing, pass@k and activation-patching analyses.

Critical strengths
  1. The insert-then-answer design for priming stage is a clean intervention such that M0 and R(q) fixed leaves the inserted triples the only variable, ensuring that any improvement in the answerer’s output is directly attributable to the explicitly injected triples.
  2. The pipeline of the two training stages that separates factual recall from reasoning, using KG as the interface is a new way that turning "stored but inaccessible" into a trainable objective.
  3. The experiments' settings, general-capability check, reward ablation, pass@k are clearly documented. The appendix gives all prompts, GRPO hyperparameters, and the full retrieval configuration.
Critical weaknesses
  1. Parameter-originated does not mean parametric knowledge. Models can make up knowledge if it cannot infer knowledge from subgraphs, instead of using its own parametric knowledge.
  2. Stage 2 confounds the evidence for Stage 1. Stage 2 is meant to show that Priming improved recall, but it is itself another 500 steps of GRPO with the same reward. The VAKE♢-vs-GRPO♢ gap could come from Stage 1 or simply from training twice as long. A 1000-step GRPO♢ run is needed to show the necessary effect of stage 1.
  3. The main results use only two backbones, both from the Qwen family and of similar size (Qwen2.5-7B and Qwen3-8B). Whether the method or the attribution findings could be transferred to other model families (e.g., Llama, Mistral) or to larger/smaller scales is untested.
Questions for authors
  1. Please report GRPO♢ trained for 1000 steps on both backbones.
  2. It is unclear that how the retrieved subgraph is built for the OOD datasets. Is any OOD evaluation closed-book?
  3. The main method needs gold passages and the gold answer to build R(q). How would the method be applied when neither is available?
  4. Fig. 6 shows that the full two-stage pipeline was trained on Qwen2.5-3B and 14B, but no test or OOD results for these runs appear in any table.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.