Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

Zorik Gekhman; Roee Aharoni; Eran Ofek; Mor Geva; Roi Reichart; Jonathan Herzig

cs.CLMar 10, 2026

Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

Authors: Zorik Gekhman, Roee Aharoni, Eran Ofek, Mor Geva, Roi Reichart, Jonathan Herzig

Organizations: Google Research · Technion - Israel Institute of Technology · Tel Aviv University

Abstract

While reasoning in LLMs plays a natural role in math, code generation, and multi-hop factual questions, its effect on simple, single-hop factual questions remains unclear. Such questions do not require step-by-step logical decomposition, making the utility of reasoning highly counterintuitive. Nevertheless, we find that enabling reasoning substantially expands the capability boundary of the model's parametric knowledge recall, unlocking correct answers that are otherwise effectively unreachable. Why does reasoning aid parametric knowledge recall when there are no complex reasoning steps to be done? To answer this, we design a series of hypothesis-driven controlled experiments, and identify two key driving mechanisms: (1) a computational buffer effect, where the model uses the generated reasoning tokens to perform latent computation independent of their semantic content; and (2) factual priming, where generating topically related facts acts as a semantic bridge that facilitates correct answer retrieval. Importantly, this latter generative self-retrieval mechanism carries inherent risks: we demonstrate that hallucinating intermediate facts during reasoning increases the likelihood of hallucinations in the final answer. Finally, we show that our insights can be harnessed to directly improve model accuracy by prioritizing reasoning trajectories that contain hallucination-free factual statements.

Figures & tables

Figure 1: Pass@ $k$ curves across two closed-book QA benchmarks and three LLMs, comparing the same models with reasoning OFF vs ON .

Figure 2: $\Omega$ in all models and datasets (§ 2 ). Models organized from the most (left) to the least effective (right) in terms of pass@ $1$ .

Figure 3: Reasoning Effectiveness on different question types in SimpleQA-Verified , with 95% confidence intervals.

Figure 4: Computation buffer effect on Gemini-2.5-Flash (§ 5.1 ). ON Single Dummy overrides the thinking trace with a short dummy sequence. ON Dummy does the same, but repeats the short dummy sequence to match the length of the original trace.

Figure 5: Reasoning effectiveness (Equation 1 ) as a function of the input length in tokens when conditioning on dummy reasoning trace (see § 5.1 ). ON Dummy X overrides the reasoning trace with a short dummy sequence which is repeated such that the input length will be X.

Figure 6: Factual priming effect on Gemini-2.5-Flash (§ 5.2 ), conditioning on facts recalled during reasoning, with reasoning either OFF ( OFF Facts ) or ON ( ON Facts ).

Figure 7: Within-question comparison of correct final-answer rates in clean (x-axis) vs. hallucinated (y-axis) reasoning traces. Each question is one point; red examples lie above the no-effect diagonal, green below.

Strategy	SimpleQA-Verified	EntityQuestions
Regular	27.9	56.9
Only Facts	30.2 (+8.2%)	58.4 (+2.6%)
Only Correct Facts	31.3 (+12.2%)	59.8 (+5.1%)

Table 1: Expected accuracy under test-time selection criteria based on factual recall and factual correctness.

Figure 8: Case study for the effectiveness of the computational buffer effect.

Figure 9: Case study for the effectiveness of factual priming.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Relation	Question Template
P176	Which company is [X] produced by?
P264	What music label is [X] represented by?
P50	Who is the author of [X]?
P26	Who is [X] married to?

Appendix

Table 2: Overview of the relations that we use from the EntityQuestions dataset and their corresponding question templates.

Figure 14: Factual priming effect on Gemini-2.5-Flash (§ 5.2 ), conditioning on facts recalled during reasoning, with reasoning OFF . OFF Facts is our main baseline used throughout the paper, while OFF Facts Conservative is provided as additional reference with conservative filtering, where each mention of the answer is removed (see § A.4.2 .

Figure 17: The effect of increasing compute. dummy_X overrides the thinking trace with a short dummy sequence which is repeated such that the total input length will be X.

Factual Knowledge in Language Models

Factual knowledge in language models concerns what facts they store, how reliably they retrieve them, and how that knowledge can be tested or changed.

See all 186 papers

Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs
May 8, 2026Wanli Yang, Hongyu Zang, Junwei Zhang +5RL for Language Models Factual Knowledge in Language Models
University of Chinese Academy of Sciences · State Key Laboratory of AI Safety, Institute of Computing Technology, CAS
MemoReason: Evaluating the Effect of Parametric Memory on Contextual Reasoning in LLMs
Sep 28, 2026Zineddine Tighidet, Andrea Mogini, Jiali Mei +2LLM Evaluation Memorization in Language Models
BNP Paribas Paris, France · Sorbonne Université Paris, France · Criteo AI Lab
Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs
Nov 8, 2025Renfei Zhang, Manasa Kaniselvan, Rylan Schaeffer abd Niloofar MireshghallahRL for Language Model Reasoning Factual Knowledge in Language Models
Carnegie Mellon University · MIT · Stanford University

May 8, 2026J/K move · Enter open · S save

Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs

Wanli Yang, Hongyu Zang, Junwei Zhang, Wenjie Shi+4

University of Chinese Academy of Sciences · State Key Laboratory of AI Safety, Institute of Computing Technology, CAS

RL for Language Models Factual Knowledge in Language Models

Open paper AskPDF arXiv

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

Abstract

Factual Knowledge in Language Models

Explore similar work

Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs

MemoReason: Evaluating the Effect of Parametric Memory on Contextual Reasoning in LLMs

Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs