Anticipation Is a Conjunction: Reference-Linked Evidence Units for Patent Novelty

A patent claim is not novel ("anticipated") only if a single earlier document discloses every one of its elements. One missing element is enough to keep the claim novel. Current language models treat novelty the way retrieval systems treat relevance, as overall closeness. That is the wrong shape for the legal question. The evidence for each element is also scattered. In a utility patent, a claim on page 2 names a part, the detailed description on page 9 assigns it reference numeral 42, and Figure 3 on page 18 draws it. When the documents are cut into fixed token windows, those three pieces land in different chunks.

We propose to rebuild the evidence unit around the patent's own cross-reference structure. Each unit bundles a reference numeral, every description sentence that mentions it, and the region of each figure where it appears. Novelty is then decided element by element and combined with the logical AND that anticipation law requires, rather than scored holistically. The bet is that this combination does two things. It removes the "everything looks non-novel" bias documented for language models. It also makes the cited prior art help rather than hurt.

Core Claim

The most telling number in the file is a negative one. On the USPTO G06F novelty set, Can AI Examine Novelty of Patents? reports these accuracies: GPT-4o reading the claim alone, 0.729; GPT-4o comparing the claim against the cited prior-art text, 0.580; few-shot Llama 3 70B, 0.624; human experts, 0.600. The same paper reports a systematic bias toward predicting "Non-Novel."

Giving the model the actual evidence makes it worse. We read this as a failure of evidence structure, not of capacity. A bag of loosely similar prior-art text gives the model many partial matches. A model that scores closeness will then tend to call everything anticipated. This is the semantic-legal relevance gap described for case retrieval in Legal Case Retrieval: A Survey of the State of the Art, where near-identical narratives can carry opposite legal outcomes.

Our claim has two parts. First, decomposing the claim into elements and requiring each to be disclosed within one reference should raise accuracy above the claim-only 0.729. It should also raise recall on truly novel claims, because the decision is a conjunction rather than a similarity score. Second, the decomposition only works if each element can find its full evidence. That requires segmenting prior art by its reference-numeral graph, linking text and drawings, rather than by token count. Figures enter as one part of that unit. Whether they add anything beyond the description text that already names the numerals is an open empirical question we test directly, not an assumption.

Relation to Prior Work

The legal side supplies the reasoning shape. An Element is Worth a Thousand Words shows that explicit legal elements improve case retrieval: the implicit Elem4LCR-I model reaches MAP 55.1 against BERT's 50.6, and ground-truth elements reach 63.4. That work depends on a hand-annotated inventory of 475 criminal-law elements. Patent claims give us elements for free, because claims are already written as enumerated limitations. That removes the annotation bottleneck the paper names.

DELTA separates outcome-determining "key facts" from narrative noise, but assumes clean Fact/Reasoning sections. Patents have no such sections, but they have numerals, which serve as an alternative anchor.

PatentEdits casts novelty as textual entailment over claim sentences, reaching 70.1 F1 on kept claims. It treats sentences independently and ignores drawings. We keep the entailment framing but apply it per element and combine the results with the legal conjunction.

On segmentation, LongAttnComp supplies the strongest text-only selector. It scores fixed 256- or 1024-token chunks with a small cross-attention module, keeps a token budget, and reorders the selected chunks into document order. It beats full context on InfiniteBench Code-Debug, 76.40% versus 74.37%. However, it degrades on LongBench v2, where evidence is diffuse. Patent evidence is exactly that kind of diffuse, cross-referenced material. The paper also notes that the best chunk size depends on the task, which motivates segment boundaries that come from the content.

MuLD shows that local context preservation does not guarantee reasoning across chunks. That constrains us: we cannot expect better chunks alone to fix the problem, which is why the conjunction step is part of the method.

On the multimodal side, IMPACT and PatentLMM align patent drawings with text. IMPACT works on design patents. PatentLMM generates figure descriptions, one figure at a time at 384×384 resolution, and explicitly leaves reasoning across a whole document as future work. Neither makes a legal decision.

Two papers constrain how we treat figures:

  • SciVer finds models lean on text as an "interpretive crutch" and ignore contradicting visuals.
  • Plot Twist shows chart titles drive answers: accuracy reaches up to 79% when the title contains the answer and drops to around 30% without it.

Patent descriptions play the role of the title. We should therefore expect a model to rely on them, and we must measure honestly whether figures matter.

MuSciClaims supplies the method for that measurement. Giving GPT-4o gold panel hints raised its Biology F1 from 0.55 to 0.75. Our reference-linked units play the role of automatic localization hints.

Research Program

The system has three parts.

Segmenter. The segmenter parses a prior-art patent into a reference graph. Description sentences are linked to the numerals they mention, using simple pattern matching of the form "noun + digits," the kind of deterministic extraction that Regex in a Time of Deep Learning shows can beat embeddings when the trigger is formal. Figure regions are linked to numerals by optical character recognition of the reference callouts, run on full-resolution drawing pages rather than downsampled images. Each numeral becomes an evidence unit containing its name, all description sentences mentioning it, and crops of every figure region bearing it. Text with no numerals falls back to sentence-aligned windows with overlap.

Element verifier. The target claim is split into its limitations, using the claim's own semicolon and "wherein" structure, with dependent claims expanded to include their parent limitations. This expansion addresses the dependency failure noted in PatentEdits. For each limitation, a retriever ranks evidence units. A verifier, either a text model or a vision-language model, labels the limitation as disclosed, not disclosed, or unclear given the top units.

Conjunctive aggregator. The aggregator predicts "Non-Novel" only if, within a single reference, every limitation is disclosed. If any limitation is undisclosed, the prediction is "Novel." A calibrated soft version multiplies per-limitation disclosure probabilities. The aggregator also returns the specific missing element, which makes errors inspectable.

The main data work is extending the Ikoma and Mitamura G06F set: 1,396 training and 398 test cases. We will fetch the full cited prior-art documents, including drawings, from public USPTO records. Where the examiner's office action cites a figure or numeral for a limitation, we record it as gold evidence localization. This gives a free analogue of MuSciClaims' gold hints.

Baselines

Naive reference points. The do-nothing reference is the majority class on the G06F test split. Next is the claim-only GPT-4o result of 0.729 accuracy reported by Can AI Examine Novelty of Patents?. This is the number to beat. A method that uses evidence but cannot beat a model that never sees the evidence has not shown that it reasons over evidence. We also report that paper's claim-plus-cited-text figures (0.580 for GPT-4o, 0.624 for few-shot Llama 3 70B) and its human figure (0.600). We re-run the claim-only and claim-plus-text settings ourselves on our backbones so all comparisons share models.

Strongest segmentation competitor. This is LongAttnComp-style compression (LongAttnComp) applied to the concatenated prior art. We try 256- and 1024-token chunks, with the same token budget our units consume and positional reordering. Its output feeds the same verifier backbone, both holistically and through our element verifier. This is the strongest competitor because it is the best published content selector for long inputs, and it isolates whether reference-linked boundaries beat learned chunk scoring. No reported result was found on Lacuna for LongAttnComp on patent novelty.

Text-only multimodal ablation. Identical units are used with the figure crops removed. This isolates the visual contribution.

Element-decomposition ablation. Reference-linked units feed a single holistic novelty judgment with no conjunction. This isolates the aggregator.

Fairness. All systems share the same open backbones: Llama 3 70B for text, and InternVL3 or Qwen2.5-VL for images. They share the same few-shot examples, the same retrieval budget, and the same hyperparameter search on a held-out slice of the training split.

What counts as beating them. The proposal succeeds if the full system exceeds 0.729 accuracy and exceeds the compression baseline by a margin larger than the bootstrap confidence interval on 398 cases. Novel-class recall must also rise, not just overall accuracy.

Why not just…?

Why not just give a long-context vision-language model the whole patent with all its drawings? This is the obvious end-to-end alternative. The evidence in the file argues against it, but not conclusively. More cited text already lowered GPT-4o from 0.729 to 0.580 in the G06F study. MuLD shows long-window models still struggle with synthesis across distant passages. PatentLMM's 384×384 downsampling suggests whole-page inputs will lose the callout numerals our linking needs. We run this as a baseline at the largest context the backbone supports. If it matches our system, the segmentation contribution is void, and we would report that.

Why not just use text, since descriptions already name every numeral? This is a real threat, and the SciVer "text crutch" finding predicts it. Utility-patent descriptions often describe drawings exhaustively, so figures may be redundant. We do not argue around this. The text-only ablation measures it directly. We further split the test set by whether the examiner cited a figure for the deciding limitation. If figures help nowhere, including that subset, the proposal becomes a text-only reference-linking method. The conjunction result would still stand on its own.

Why not just prompt the model to check each element, without special segmentation? Element-wise prompting over ordinary retrieved chunks may capture most of the gain. That would mean the legal conjunction, not the segmentation, does the work. The element-decomposition ablation and the compression-plus-element-verifier baseline separate these two effects cleanly. We expect, but have not shown, that numeral-severed chunks cause false "not disclosed" judgments. The evidence-localization measurement below tests this.

Proposed Experiments

Primary test. On the G06F test split, we report accuracy, macro-F1, and novel-class recall for all baselines and ablations. Bootstrap intervals are computed over the 398 cases. This is decisive in two ways:

  • If the conjunctive system with reference-linked units beats both the claim-only 0.729 and the compression baseline, while raising novel-class recall, the core claim holds.
  • If accuracy rises but novel-class recall does not, we have only shifted the decision threshold, not removed the bias.

Localization. On cases where examiners cited figures or numerals, we measure whether the retrieved units contain the examiner-cited evidence (recall at k). We report this separately for reference-linked units and for token chunks. This tests the mechanism, not only the outcome.

Figure perturbation. Following the MuSciClaims perturbation logic, we edit a disclosed numeral out of a figure crop while leaving the text intact, and the reverse. If the verifier's judgment never changes when only the figure changes, it is ignoring drawings, regardless of headline accuracy.

Transfer. We apply the element splitter and conjunction to PatentEdits Kept/Edited/Deleted prediction, text only, comparing against its reported 70.1 F1 on kept claims. This checks whether the conjunction transfers to a larger, differently labeled corpus.