PaperlayerView exact revision on arXiv

Source-linked AI summary

Missing vs. Unused Knowledge Hypothesis for Language Model Bottlenecks in Patent Understanding

Siyang Wu, Honglin Bao, Nadav Kunievsky, James A. Evans

arXiv:2505.12452v4cs.CLcs.CYcs.DLcs.IR

TL;DR

LLMs can know relevant facts yet fail to apply them when distinguishing semantically similar, objectively different patents. The paper diagnoses this gap with clarification questions and tests whether self-generated or externally retrieved answers improve performance. It finds that failures mainly reflect difficulty deploying existing knowledge, with smaller and larger models offering complementary strengths in constructing usable cues.

  • Problem

    Evidence on whether LLMs can apply stored knowledge to holistic conceptual distinctions is limited, despite its importance for knowledge-intensive tasks.

    Source: §1 · pp. 1–2

  • Method

    The paper tests patent differentiation and compares baseline judgments with judgments conditioned on model-generated clarification questions answered by the model or with external scientific information.

    Source: §1 · pp. 1–2

  • Results

    Most errors stem from poor retrieval of existing knowledge rather than true knowledge gaps, and smaller models produce simpler, more transferable cues while larger models produce richer but more brittle cues.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings point to complementary roles across model scales: smaller models can propose a scaffold while larger models do the heavier reasoning.

    Source: §1 · p. 2

    The semantic threshold models use to distinguish patents remains underdefined, including whether judgments rest on functionality, application, architecture, or terminology.

    Source: Limitations · p. 9

Abstract

from arXiv · show

While large language models (LLMs) excel at factual recall, the real challenge lies in knowledge application. A gap persists between their ability to answer complex questions and their effectiveness in performing tasks that require that knowledge. We investigate this gap using a patent classification problem that requires deep conceptual understanding to distinguish semantically similar but objectively different patents written in dense, strategic technical language. We find that LLMs often struggle with this distinction. To diagnose the source of these failures, we introduce a framework that decomposes model errors into two categories: missing knowledge and unused knowledge. Our method prompts models to generate clarifying questions and compares three settings -- raw performance, self-answered questions that activate internal knowledge, and externally provided answers that supply missing knowledge (if any). We show that most errors stem from failures to deploy existing knowledge rather than from true knowledge gaps. We also examine how models differ in constructing task-specific question-answer databases. Smaller models tend to generate simpler questions that they, and other models, can retrieve and use effectively, whereas larger models produce more complex questions that are less effective, suggesting complementary strengths across model scales. Together, our findings highlight that shifting evaluation from static fact recall to dynamic knowledge application offers a more informative view of model capabilities.

1 Introduction

The introduction frames understanding as deploying stored knowledge to distinguish concepts, and presents patent differentiation as a demanding test of that ability. Its diagnostic results point mainly to failures to access existing knowledge, alongside complementary cue-building strengths across model scales.

1 Introduction

  • Understanding is treated as dynamic deployment of knowledge, beyond recalling facts or executing procedures, to support transfer and generalization.The testbed emphasizes mechanisms, purposes, and constraints in patents.Source: §1 · pp. 1–2
  • The patent task asks models to distinguish linguistically similar but functionally distinct inventions by detecting subtle differences in purpose and mechanism.The patents are dense and strategically obfuscated, making surface heuristics ineffective.Source: §1 · p. 2
  • Frontier models show some competence, while weaker models consistently miss key technological nuances.The paper uses this result to motivate diagnosing whether errors arise from missing or inaccessible knowledge.Source: §1 · p. 2
  • Most understanding failures reflect poor access to existing knowledge rather than a lack of knowledge, and smaller and larger models show complementary strengths in building usable cues.Smaller models tend to produce simpler, more transferable cues; larger models produce richer but more brittle cues.Source: §1 · p. 2
  • The paper introduces a large-scale patent differentiation dataset and a framework for separating missing knowledge from unused knowledge.The dataset contains 1.3M+ information/computer patents and 177K+ biomedical patents granted after 2015.Source: §1 · p. 2

2 Hard-to-Distinguish Patents

The paper uses a large patent dataset because distinct inventions can be textually similar, difficult for models to parse, and independently distinguished through expert review.

2 Hard-to-Distinguish Patents

  • The dataset contains 1,319,184 information and computer technology patents and 177,903 biomedical patents.The task is designed to be less affected by pretraining, as discussed in Appendix G.Source: §2 · p. 2
  • Patents often form dense clusters of near-duplicates whose small but meaningful differences test recognition of functional distinctions.The most-similar distinct-patent pairs have similarities concentrated at high scores; Figure 1b depicts their distribution.Source: §2 · p. 3, Figure 1 · p. 3
  • LLaMA-3.1-8B-Instruct has higher perplexity on patent descriptions than scientific papers, indicating greater difficulty predicting patent text.The comparison uses 4,000 computer science papers and 4,000 matched patent descriptions; perplexity is not a direct measure of understanding.Source: §2 · p. 3, Figure 1 · p. 3
  • Granted patents are treated as conceptually distinct because USPTO examiners review novelty, non-obviousness, and utility using prior-art searches and comparison.This institutional review provides an expert-evaluated basis for the task’s different-patent ground truth.Source: §2 · p. 3

3 Evaluating LLMs’ Understanding of the Technology Domain

The evaluation tests whether models can consistently distinguish close but distinct patents, rather than relying on a single prompt framing. Results reveal response biases and confidence patterns that complicate apparent success.

3 Evaluating LLMs’ Understanding of the Technology Domain

  • The task presents hard-to-distinguish patents with different granted IDs and asks whether they describe the same invention, using confidence-weighted votes across three generations.The final label follows whichever of the two labels has the higher total confidence.Source: §3 · p. 3
  • 0.95 correlation separates confidence-weighted voting from majority voting, while high confidence correlates with judge accuracy at 0.92.Results on samples of 1,000, 5,000, and 10,000 pairs show correlation over 0.98.Source: §3 · p. 4
  • Most tested models often identify different patents as distinct under the initial prompt, with Qwen 2.5 7B as the exception.Figure 2’s top row asks whether the patents are different; the inverse prompt exposes the limits of this apparent success.Source: §3 · p. 4, Figure 2 · p. 4
  • GPT-4o mini shows the strongest capability, followed by Llama 3.3 70B and Llama 3.1 8B, based on consistency across the two framings.GPT-4o mini has the lowest “yes” response rate when asked whether the distinct patents are the same.Source: §3 · p. 4, Figure 2 · p. 4

4 Experimental setups

The experiments diagnose patent-differentiation failures by eliciting background questions and comparing self-generated answers with answers supported by retrieved scientific information.

4 Experimental setups

  • Models are asked what background knowledge would help distinguish a patent pair, then tested with self-QA and scientific retrieval conditions.The comparison assesses whether the model can use its internal knowledge or benefits from external information.Source: §4 · p. 5
  • For each of two patents, the model generates six questions: three about surface-level concepts and three about deeper integration, improvement, and invention intent.The question order is randomized, with little impact on results.Source: §4 · p. 5
  • Scientific-RAG answers use retrieved arXiv computer science paper chunks, while Self-QA answers are generated by the model without external retrieval.In the Scientific-RAG condition, retrieved knowledge is used to answer the generated questions before the model judges the patent pair again.Source: §4 · p. 5

5 Results

The results indicate that prompting can help models deploy knowledge they already possess, while question effectiveness across model scales depends on how well the questions fit the answerer. The experiments compare baseline performance, self-questioning, scientific answers, and question sets from models of different sizes.

5.1 Missing vs. Unused Knowledge

  • At baseline, Llama 3.1 8B incorrectly classifies approximately 70% of patent pairs as identical, and adding question–answer pairs improves performance.Questions alone provide a modest improvement; adding self-generated answers significantly improves performance.Source: §5.1 · pp. 5–6
  • Scientific QA, self-QA, and question-only support understanding in that order, while Figure 3b finds no significant difference between scientific and self-produced answers.The baseline-to-self-answer improvement is more substantive than the improvement from self-generated to scientific answers.Source: §5.1 · pp. 5–6, Figure 3 · p. 6
  • The authors argue that models likely encountered much of the retrieved scientific information, supporting knowledge deployment rather than wholly new acquisition as the primary explanation for improvement.The paper reports that approximately 86% of retrieved papers that generated scientific answers were published before LLaMA 3.2’s release.Source: §5.1 · p. 6
  • Scientific answers average 62.1 ± 1.8 words, compared with 38.2 ± 0.6 words for self-generated answers, despite the answer types occupying largely overlapping embedding regions.The self-generated responses also contain a distinct outlier region that the authors suspect reflects hallucinated content.Source: §5.1 · p. 6
  • The findings suggest that relevant knowledge is already present but compressed and difficult to access without activation cues.The paper argues that representation and dynamic use can be as important as knowledge access, so retrieval may be insufficient when deployment is the bottleneck.Source: §5.1 · p. 7

5.2 It’s Not Only What to Ask, but Who Asks It

  • The 3B model performs best with 1B-generated questions, while the 8B model benefits from questions generated by both 1B and 8B models.Smaller models’ questions are compatible with larger answerers, but the reverse is not consistently true; compatibility is measured by lower misjudgment rates.Source: §5.2 · p. 7, Figure 4 · p. 8
  • Smaller models tend to ask more fundamental, shorter, and more open-ended questions, which elicit more varied answers and can prompt answerers to draw on information beyond the patent text.The accompanying analyses report higher answer perplexity and lower similarity to patent text when answerers respond to smaller models’ questions.Source: §5.2 · pp. 7–8, Figure 4 · p. 8

6 Final Remarks

The paper reframes many LLM failures as limits on deploying knowledge rather than acquiring it. Its patent-differentiation task also points to complementary roles for smaller and larger models.

6 Final Remarks

  • The paper frames many failures on knowledge-intensive tasks as limitations in knowledge deployment rather than knowledge acquisition.In the patent differentiation task, models often possess relevant knowledge but fail to activate it; external retrieval yields diminishing returns when utilization is the primary bottleneck.Source: §6 · p. 8
  • The patent differentiation task probes dynamic knowledge use in distinguishing concepts that are difficult to separate.The paper presents this task as a way to assess LLMs’ ability to apply domain knowledge.Source: §6 · p. 8
  • Complementarities across model scales suggest hybrid strategies pairing smaller models’ structural scaffolding with larger models’ deeper reasoning.The paper connects this approach to automated assessment of innovation.Source: §6 · p. 8

Limitations

The authors identify bounds in the retrieval setup and in what successful patent differentiation establishes. They point to adaptive retrieval and a clearer account of the distinctions models make as future work.

Limitations

  • The current retrieval setup uses scientific sources from arXiv papers.The authors suggest evolving retrieval beyond fixed sources and single-pass self-questioning.Source: Limitations · p. 9
  • The authors propose adaptive query refinement and model-in-the-loop retrieval as directions beyond the current setup.The suggested agentic approach would let models refine questions, decide whether to retrieve, assess retrieved information, and determine how much to seek.Source: Limitations · p. 9
  • The semantic basis for successful patent differentiation remains underdefined, including whether models distinguish function, application, architecture, or terminology.The authors also identify variation across patent classes and categories as a future research question.Source: Limitations · p. 9

Ethics Statement

The paper treats understanding as grasping relations between concepts, with differentiation and generalization as complementary processes. It also reviews limitations of current benchmarks and approaches for accessing or shaping model knowledge.

Ethics Statement

  • The authors report reviewing AI-assisted coding and grammar checking, with code reviewed by the first author and no external ethics review required.The experiments analyzed patent texts and were assessed as posing no risk to individuals across demographic or geographic groups.Source: Ethics Statement · p. 9

A Understanding Through Differentiation

  • Understanding requires navigating relationships among concepts, not merely recalling facts or performing procedures.The passage identifies causal links, correlations, hierarchies, and analogical parallels as examples of underlying structures.Source: A Understanding Through Differentiation · p. 12
  • Differentiation contrasts concepts, while generalization identifies shared properties and treats distinct instances as equivalent for a purpose.The paper presents these as complementary processes for constructing relational knowledge.Source: A Understanding Through Differentiation · p. 12

B Related Literature

  • The related-work review cautions that benchmark scores may reflect memorization or surface patterns rather than comprehension, and that evaluation-format changes can alter results.It also notes that aggregate metrics can obscure where models succeed or fail.Source: B Related Literature · p. 13
  • Retrieval-augmented generation supplies retrieved documents as explicit evidence instead of relying only on compressed parametric memory.The review describes retrieval as a response to knowledge that may be stored in models but not reliably accessible on demand.Source: B Related Literature · p. 14
  • The paper’s evaluation is motivated by contrastive learning’s aim of separating semantically related and unrelated inputs, but does not itself train contrastively.The review links consistent discrimination between distinct concepts to a deeper form of understanding.Source: B Related Literature · p. 14

C Judging Close Patents VS. Rewritten Patents

Few-shot chain-of-thought helps LLaMA-3.1-8B-instruct recognize paraphrased patent texts as equivalent, but has little effect on distinguishing semantically close, distinct inventions. The authors suggest the paraphrase task permits surface-level strategies that do not resolve the harder conceptual distinction.

C Judging Close Patents VS. Rewritten Patents

  • The Figure 5 right panel shows lower misjudgment for rephrased patents, while the left panel shows difficulty with close-but-different pairs.The plotted “yes” responses are errors in both panels, though the questions differ.Source: Figure 5 · p. 14
  • Few-shot CoT substantially improves recognition of paraphrased patents as equivalent but has minimal effect on distinguishing closely related, different inventions.Raw zero-shot judgments performed poorly on both merging and splitting tasks.Source: C Judging Close Patents VS. Rewritten… · p. 15
  • The authors suggest paraphrase success may rely on semantic overlap and pattern matching, whereas distinguishing similar but substantively different patents requires conceptual depth.Their rewritten texts retain key technical terms while changing sentence structure, making the merging task comparatively easy.Source: C Judging Close Patents VS. Rewritten… · p. 15
  • The rewrite approach does not fully obscure surface similarity; the authors note that functionally equivalent rewrites avoiding superficial alignment would be difficult to design.They say such rewrites might bring the model closer to the harder near-match task.Source: C Judging Close Patents VS. Rewritten… · p. 15

D The Patent Dataset

The dataset pairs technically similar patent texts that differ in meaningful ways, and the appendix describes prompts for generating domain questions and judging patent pairs. The prompts distinguish question levels and use structured answers for judgment and question answering.

D The Patent Dataset

  • The computer-science collection contains 1,319,184 patent pairs across five CPC classes and 41 subclasses, while the biomedical collection contains 177,903 patents.The biomedical patents are drawn from USPTO classes C01, C07–C14, and C21–C30.Source: D The Patent Dataset · p. 15
  • Among matched pairs, 82.55% share a CPC class and 68.10% share a subclass, while 10.50% were granted in the same year.The passage presents the pairs as technically similar despite differences in granted patent identity.Source: D The Patent Dataset · p. 15

E Prompts

  • The question-generation prompt organizes domain-knowledge questions into remembering, understanding, and applying levels, while separate prompts target basic recall and component interactions.The prompts request questions useful for understanding a patent without directly referencing its text.Source: E Prompts · pp. 16–17
  • The judgment prompt asks whether two patent texts describe the same patent and returns a binary label plus a confidence score from 0 to 10.The appendix also includes prompts for answering questions with retrieved scientific context and for self-talk answers.Source: E Prompts · pp. 17–18

F Results for Majority Voting Schema

The majority-voting results report that self-talk improves Llama 3.1 8B instruct’s baseline judgment but remains below scientific answers, and that larger models’ questions are less useful to answer models. The appendix also addresses pretraining exposure and the form of generated questions.

F Results for Majority Voting Schema

  • The majority-voting figures distinguish false “same” responses from correct “different” responses on pairs known to be different.Figure 6 encodes “same” as wrong; Figure 7 encodes “yes” to “different” as correct.Source: Figure 6 · p. 19, Figure 7 · p. 19
  • Self-talk improves Llama 3.1 8B instruct’s baseline majority-vote judgment but remains below scientific answers (q1 to q5 p=0.000; q6 p=0.006).The figure evaluates false “same” judgments on patents that are actually different.Source: Figure 8 · p. 19
  • Questions proposed by bigger models are less useful for Llama 3B/8B answer models judging actually different patents.Figure 9 frames the comparison as judgments of whether the patents are the same.Source: Figure 9 · p. 19

G Pretraining Exposure

  • Because the patent-pairing mechanism is extremely unlikely to have been encountered during pretraining, concatenating all unordered pairs of roughly 1.5 million patent texts would require approximately 1.2 × 10^15 tokens.This is multiple orders of magnitude larger than reported corpus sizes: ∼3 × 10^11 tokens for GPT models and ∼2 × 10^12 for Llama-70B.Source: G Pretraining Exposure · p. 20

H Example Questions

  • Questions included “what is” or “what are” in 68% of Llama 1B’s and 44% of Llama 405B’s generated questions, indicating many factual, definitional queries.Figure 10 provides example questions.Source: H Example Questions · p. 20, Figure 10 · p. 20

I Ablation Studies

Adding scientific QA pairs improves patent-judgment accuracy beyond placebo answers. Random same-paper text also helps initially, but its benefit does not grow with more questions.

I Ablation Studies

  • As more science QAs are added, judge accuracy significantly increases and becomes better than placebo answers. Confidence in correct and incorrect answers decreases, with a significantly greater drop for incorrect answers.Source: I Ablation Studies · p. 21
  • The baseline measures judge accuracy and confidence without self-questioning or scientific retrieval; the experiment adds one to six scientific QA pairs per patent.Source: I Ablation Studies · p. 20
  • Placebo text improves judgment initially, but unlike actual scientific answers, its effect does not scale with the number of questions.The placebo is a random same-paper excerpt matched in length to the actual answers.Source: I Ablation Studies · p. 21
  • As QA pairs are added, confidence in both correct and incorrect answers decreases, with a significantly greater drop for incorrect answers.Source: I Ablation Studies · p. 21
Loading 2505.12452v4…