PaperlayerView exact revision on arXiv

Source-linked AI summary

GameCommBench: A Unified Benchmark and Type-Aware Evaluation for AI-Generated Game Commentary

Qirui Zheng, Zhengteng Lin, Yunyi Xiao, Junhao Li, Keyuan Cheng, Xingbo Wang, Yongyi Wang, Lingfeng Li, Yunlong Lu, Wenxin Li

arXiv:2610.11129v1cs.AIcs.CL

TL;DR

AI-GGC is an open-ended task, yet research resources remain fragmented and commentary serves several distinct functions. The paper introduces GAMECOMMBENCH, a unified benchmark, and TACE, a type-aware evaluator, then finds uneven commentator capabilities, with live observation and strategic analysis remaining bottlenecks.

  • Problem

    AI-GGC resources are fragmented, while commentary requires live observation, strategic analysis, and recall of relevant game context.

    Source: §1 · p. 1

  • Method

    The paper unifies game contexts in GAMECOMMBENCH and proposes TACE, which evaluates descriptive, analytical, and background commentary with type-specific criteria.

    Source: §1 · pp. 1–2

  • Results

    AI commentators show non-uniform capabilities, with grounded game-state description and faithful strategic reasoning remaining key bottlenecks.

    Source: §5.2 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    GAMECOMMBENCH and TACE support more comparable and interpretable evaluation of AI-GGC.

    Source: §6 · p. 8

    The benchmark covers five games and English text commentary, not speech-based or other multimodal game-broadcast presentation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Game commentary is an open-ended generation task requiring multimodal perception, strategic reasoning, and contextual knowledge. Existing AI-Generated Game Commentary (AI-GGC) studies remain fragmented across games, modalities, and evaluation protocols, while overlap-based or holistic evaluators fail to capture the functional heterogeneity of commentary. We introduce \textsc{GameCommBench}, a unified benchmark spanning board games, sports, and esports, with commentary aligned to heterogeneous game contexts and annotated by commentary type. We further propose Type-Aware Commentary Evaluation (TACE), a structured framework for evaluating different types of commentary. We then validate TACE for reliability and human agreement, and use it to benchmark representative AI commentators. Results reveal non-uniform capability profiles, with live observation and strategic analysis emerging as major bottlenecks. Together, \textsc{GameCommBench} and TACE provide a diagnostic foundation for comparable and interpretable AI-GGC evaluation.

1 Introduction

AI-GGC combines live observation, strategic analysis, and contextual recall, but its research resources and evaluation approaches remain fragmented. The paper introduces a unified benchmark and type-specific evaluation, then uses them to diagnose commentator capabilities.

1 Introduction

  • AI-GGC requires live observation, strategic analysis, and recall of relevant game context, making it an open-ended generation task.Commentary helps audiences follow events, interpret decisions, and place moments in broader context.Source: §1 · p. 1
  • Existing resources remain tied to particular games, modalities, data formats, and commentary focuses, limiting systematic cross-setting comparison.The field lacks a shared interface for organizing heterogeneous game contexts.Source: §1 · p. 1
  • Overlap metrics penalize valid variation, while holistic evaluators miss the distinct criteria needed for descriptive, analytical, and contextual commentary.A game state can support multiple valid commentaries and reasoning processes.Source: §1 · p. 1
  • GAMECOMMBENCH unifies board games, sports, and esports, while TACE evaluates commentary with metrics tailored to its functional type.This supports systematic comparison across games and finer-grained capability diagnosis.Source: §1 · pp. 1–2
  • The authors validate TACE and benchmark representative AI commentators, finding non-uniform capabilities and persistent live-observation and strategic-analysis bottlenecks.The paper also reports repeated-run reliability, human validation of LLM-mediated components, and cross-judge consistency checks.Source: §1 · p. 2

2 Benchmark Construction

GAMECOMMBENCH brings five games across board games, sports, and esports into a shared benchmark. Its collection and processing pipeline aligns game contexts with commentary, labels commentary by function, and packages the records under a common interface.

2.1 Benchmark Scope and Domain Coverage

  • Five games across three domains were selected to expose variation in state representation, game dynamics, and commentator demands.Chess and Go represent board games, soccer and basketball represent sports, and League of Legends represents esports.Source: §2.1 · p. 2, §2 · p. 2

2.2 Data Collection and Processing Pipeline

  • Collection aligns board-game records by turn, sports clips with transcribed broadcast commentary, and timestamped LoL clips with API snapshots.Player and team background information is collected as supplementary context.Source: §2.2 · p. 2

2.3 Data Composition and Distribution

  • 2,159 chess turns and 2,857 Go turns form part of a benchmark whose scale, modality, and commentary composition vary across games.The subsets also include 2,681 soccer clips, 2,981 basketball clips, and 7,675 LoL clips; board commentary is more analytical, sports more descriptive, and esports more balanced.Source: Figure 2 · p. 3, §2.3 · p. 3

2.2 Data Collection and Processing Pipeline

  • Cleaning repairs segment-boundary alignment; commentary is labeled by semantic function and packaged with game state under a shared record schema.Unit-level labels allow mixed-function utterances to be evaluated by type, while the schema retains game-specific inputs and raw commentary.Source: §2.2 · p. 3

3 Type-Aware Commentary Evaluation

TACE separates commentary into descriptive, analytical, and background functions because a single holistic score cannot capture their distinct evaluation needs. Its three pipelines assess event coverage and support, game-state-grounded reasoning, and the accuracy and contextual relevance of background claims.

3 Type-Aware Commentary Evaluation

  • TACE evaluates descriptive, analytical, and background commentary through separate branches rather than a single holistic score.Each branch uses criteria tailored to its commentary type.Source: §3 · pp. 3–4, Figure 3 · p. 5

3.1 Descriptive Commentary Evaluation

  • Correctness measures the share of generated events supported by reference events, while completeness measures the share of reference events covered.Reference events are extracted from human descriptive units with domain-specific signals, and generated events are aligned to them.Source: §3.1 · p. 4

3.2 Analytical Commentary Evaluation

  • Analytical commentary is decomposed into reasoning units and relations, then verified against game state at both node and edge levels.Analytical Faithfulness measures the fraction of units and relations verified; Reasoning Depth measures the longest verified chain, normalized within the instance.Source: §3.2 · p. 4

3.3 Background Commentary Evaluation

  • Background segments are decomposed into claims and scored for factual accuracy using retrieved evidence and for relevance to the immediate commentary context.Factuality and Relevance are computed by averaging claim-level judgments over the segment.Source: §3.3 · p. 5

4 Validation of TACE

TACE is evaluated for run-to-run reliability and human agreement. The reported results indicate stable scores and rankings, high type-decomposition accuracy, and strong agreement between LLM-mediated steps and human annotations.

4.1 Reliability of TACE

  • TACE’s six metrics show CV values below 3.5% and Spearman correlations from 0.856 to 0.965 across metrics, indicating stable scores and rankings.The repeated-run evaluation covers 6 models, 5 games, and 20 sampled cases per model-game pair; correlations are also reported with Gemini-3-Flash replacing GPT-5.4-mini as judge.Source: §4.1 · pp. 5–6, Table 1 · p. 6

4.2 Human Validation

  • LLM-based commentary type decomposition exceeds 95% overall accuracy across games.For each game, 100 decomposed commentaries were sampled and manually labeled.Source: §4.2 · p. 6
  • Most validated LLM-mediated steps exceed 90% accuracy against human annotations across descriptive, analytical, and background components.The validation covers 90 instances and more than 3,700 intermediate items, including event extraction, matching, reasoning relations, verification, and background-claim judgments.Source: §4 · p. 5, §4.2 · p. 6

5 Experiments

Experiments use GAMECOMMBENCH to compare representative VLMs and existing game-specific systems across games and commentary types. Results show capability imbalances, with live observation and strategic analysis remaining bottlenecks, while scaling and stronger reasoning settings bring uneven gains.

5.1 Evaluation Setup

  • Six VLMs are evaluated on the full benchmark under shared generation conditions, alongside chess and soccer AI-GGC systems for comparison.Inputs include symbolic board-game states or 4-FPS video; esports also includes timestamp-aligned API metadata, and human commentary is not supplied during generation.Source: §5.1 · p. 6, §5 · p. 6

5.2 Main Benchmark Results

  • No VLM leads across all metrics: GPT-5.4 performs best on descriptive and analytical measures, while other models remain competitive on background factuality or relevance.Figure 4 reports aggregated scores across six TACE metrics and game-wise capability profiles.Source: Figure 4 · p. 7, §5.2 · p. 7
  • 0.473 average completeness versus 0.227 correctness marks a verbosity–grounding mismatch; both measures drop sharply in video games, where live observation is a bottleneck.The passage notes that board-game symbolic input partly mitigates the issue, while human references may omit objectively valid move-level events.Source: §5.2 · p. 7
  • 0.246 average analytical faithfulness and 0.196 average reasoning depth indicate difficulty producing game-grounded strategic reasoning, even though GPT-5.4 performs best on these measures.Human analytical commentary has high faithfulness (0.78) but lower depth (0.28), so the two measures capture complementary qualities.Source: §5.2 · p. 7
  • 0.903 average relevance versus 0.683 factuality characterizes background commentary; open-weight models lead on both background metrics, while factuality still has room to improve.The paper suggests retrieval-augmented or evidence-grounded generation as a possible way to provide verified background information.Source: §5.2 · pp. 7–8

5.3 Scale and Reasoning Mode

  • Larger models generally raise descriptive completeness without reliably improving correctness, improve analytical faithfulness and background factuality, but do not consistently increase reasoning depth.Background relevance is already high for most models and is less sensitive to scale.Source: §5.3 · p. 8
  • Higher thinking settings improve five of six analytical scores across three closed-source models, but reasoning-depth effects vary by model.Gemini-3.1-Pro-Preview gains substantially in both faithfulness and depth; Claude-Sonnet-4.5’s reasoning depth decreases modestly.Source: §5.3 · p. 8

6 Conclusion

The paper presents GAMECOMMBENCH and TACE as a unified benchmark and type-aware evaluation framework for AI-generated game commentary. Validation finds stable, human-aligned signals, while experiments identify uneven capabilities and bottlenecks in grounded observation and strategic reasoning.

6 Conclusion

  • GAMECOMMBENCH unifies AI-GGC evaluation across board games, sports, and esports.The benchmark includes heterogeneous game contexts and type-annotated commentary.Source: §6 · p. 8
  • TACE evaluates descriptive, analytical, and background commentary with a type-aware framework, and validation finds stable, human-aligned evaluation signals.The paper presents the benchmark and framework as supporting more comparable and interpretable AI-GGC evaluation.Source: §6 · p. 8
  • Experiments reveal non-uniform capability profiles, with grounded game-state description and faithful strategic reasoning remaining key bottlenecks.Further analyses find uneven gains across commentary capabilities from model scale and reasoning mode.Source: §6 · p. 8

Limitations

GAMECOMMBENCH covers representative AI-GGC settings but is bounded by its game, language, and modality scope, and full multimodal evaluation is costly.

Limitations

  • The benchmark covers representative settings across board games, sports, and esports, but does not exhaust game genres, languages, or broadcasting styles.Source: Limitations · p. 9
  • The current benchmark evaluates English text commentary across five games and excludes speech-based commentary, voice delivery, and other multimodal broadcasting forms.Source: Limitations · p. 9
  • Extending coverage to other languages, regional styles, game communities, or speech and multimodal settings may require new data collection and domain-specific processing; full multimodal evaluation is costly, especially with closed-source APIs.Source: Limitations · p. 9

Ethical Considerations

The paper identifies risks from errors in generated commentary and describes the public data sources and consent practices used for GAMECOMMBENCH.

Ethical Considerations

  • Inaccurate descriptions or unsupported interpretations could mislead viewers if generated commentary is used directly in public-facing systems, especially text-only live updates.Source: Ethical Considerations · p. 9
  • The benchmark uses public names of players, teams, matches, and commentators where needed, but does not collect private or sensitive personal information or intentionally include offensive content.Source: Ethical Considerations · p. 9
  • The benchmark draws on public or previously released materials without separate consent from players, teams, or commentators, and users should respect source licenses, terms, and attribution requirements.Source: Ethical Considerations · p. 9

A Detail of Data Collection and Processing Pipeline

The appendix details how heterogeneous game data and commentary are processed into a common benchmark format, with workflows adapted to each game's source modality.

A.1 Unified record schema and normalization

  • The unified JSONL schema preserves a common semantic contract across sports, board games, and esports, with fields for game state, background, and categorized commentary.Commentary records distinguish raw and processed text and normalize categories as descriptive, analytical, or background; football and basketball use stable MD5-based identifiers.Source: §A.1 · p. 10

A Detail of Data Collection and Processing Pipeline

  • The pipeline collects and aligns raw game data, applies quality control and temporal repair, separates commentary types, and packages the results in a unified format.Source: Figure 5 · p. 11

A.2 Game-specific processing workflows

  • Processing intensity varies by source: chess comments are already aligned to PGN plies, Go records need unusable comments removed, and football, basketball, and LoL require explicit temporal alignment and cleaning.Source: §A.2 · p. 11

A.2.1 Football

  • Football processing segments and classifies transcript sentences, filters them against captions, fuses retained commentary with captions, reclassifies the fused text, and converts records to the unified schema.Source: §A.2.1 · p. 11

A.2.2 Basketball

  • Basketball first gates caption-transcript pairs for relevance because captions and transcripts may describe adjacent plays, then classifies, filters, and fuses retained text and selects two high-ranked examples from batches of three.Source: §A.2.2 · p. 11

A.2.3 Chess

  • Chess processing converts grouped GameKnot pages into PGN, cleans and normalizes move comments, and needs no timestamp matching or caption-grounded LLM cleaning because comments are attached to moves.Source: §A.2.3 · pp. 11–12

A.2.4 Go

  • Go records are extracted from FoxWQ JSON into SGF, with empty, malformed, duplicated, or off-topic material skipped or removed before move-comment pairs are normalized.Source: §A.2.4 · p. 12

A.2.5 League of Legends

  • LoL workflows align timestamped subtitles with event windows or combine live-stat snapshots, retaining significant API changes as structured context such as team gold, kills, and player metrics.Source: §A.2.5 · p. 12

C Evaluation

TACE routes segmented commentary through descriptive, analytical, and background evaluation branches, with separate criteria for observable events and reasoning.

C Evaluation

  • TACE evaluation is organized into segmentation, descriptive, analytical, and background branches, with configurable run scope and normalized sport aliases.Source: C Evaluation · pp. 12–13

C.1 Segmentation for evaluation

  • Candidate commentary is segmented into its smallest independent semantic units, assigned category IDs, and routed to descriptive, analytical, or background evaluation.Source: §C.1 · p. 13

C.2 Descriptive evaluation

  • Descriptive evaluation compares generated events with a reference event list to measure completeness and correctness, setting both scores to zero when no candidate event is extracted.Source: §C.2 · p. 13

C.3 Analytical evaluation

  • Analytical evaluation extracts propositions and constructs sparse forward-edge candidates using adjacent propositions and hybrid retrieval, then audits edges and validates the graph as a DAG.Source: §C.3 · p. 13
  • Analytical faithfulness is a graph-grounded correctness ratio over judged propositions and final edges; effective depth uses the longest valid path of true nodes and edges, normalized within each sport-record group.Source: §C.3 · p. 13

C.4 Background evaluation

Background evaluation turns category-3 units into individually checkable claims, then assesses factuality against retrieved background evidence and relevance against the same commentary’s descriptive and analytical clauses.

C.4 Background evaluation

  • Category-3 units are converted into verifiable claims carrying source, type, focus, entity, and time-scope information.These fields make later factuality and relevance judgments operate on independently checkable units.Source: §C.4 · p. 13
  • Factuality uses metadata-derived background knowledge, bounded retrieval, and evidence IDs restricted to retrieved sources; claims without valid evidence are labeled false.The retriever combines structured channels and optional semantic-embedding retrieval.Source: §C.4 · pp. 13–14
  • Scores are rounded to four decimals, with None returned when no claim can be scored.Source: §C.4 · p. 14

D Human Validation

Human validation compares LLM-produced annotations with trained internal annotators’ judgments. Type-aware decomposition reaches 95.8% overall accuracy across the evaluated games.

D Human Validation

  • Trained internal annotators independently labeled commentary units or evaluation items using provided context and guidelines, and their outputs were compared with LLM results.They relied only on provided materials; no external crowdworkers or paid participant pools were recruited, and no personal information was collected.Source: D Human Validation · p. 14
  • Table 6 compares analytical-metric model rankings from the main GPT-5.4-mini-based judge and an alternative Gemini-3-Flash-based judge on a held-out subset.Source: Table 6 · p. 14
  • For decomposition validation, 100 commentary units per game were sampled and labeled descriptive, analytical, or background by human annotators.The LLM labels were compared with human annotations to compute accuracy.Source: D Human Validation · p. 14
  • 95.8% overall accuracy, ranging from 93.0% on Go to 98.0% on football and basketball, supports reliable type-aware decomposition across games.Source: D Human Validation · p. 14

E Cross-Judge Consistency Check

This section documents structured prompts and procedures for generating, classifying, and evaluating game commentary across several settings, including football-specific event extraction and analysis. The procedures emphasize type-specific content, evidence-grounded factuality and relevance, and cross-judge consistency checks for analytical nodes and edges.

E Cross-Judge Consistency Check

  • Analytical model rankings remain consistent when Gemini-3-Flash replaces GPT-5.4-mini for node and edge verification, especially for Reasoning Depth.The held-out-subset comparison uses Spearman rank correlation and leaves the rest of the TACE pipeline unchanged.Source: E Cross-Judge Consistency Check · p. 14

Prompt A.1: Football sentence segmentation and type classification

  • Football commentary is segmented into sentences and assigned one of three categories: descriptive, analytical, or background.The prompt requires complete, non-overlapping sentence segmentation and strict JSON output.Source: Prompt A.1: Football sentence… · p. 15

Prompt A.2: Basketball sentence segmentation and type classification

  • Basketball commentary is segmented into sentences and assigned one of three categories: descriptive, analytical, or background.The prompt requires complete, non-overlapping sentence segmentation and strict JSON output.Source: Prompt A.2: Basketball sentence… · pp. 15–16

Prompt A.3: Caption-grounded sentence filtering (football)

  • Football sentence filtering retains commentary only when references are clear, content is caption-relevant and consistent, and any corrections are minimal.An incomplete opening sentence or uncertain first-sentence player/team pronoun is rejected because preceding context may be truncated.Source: Prompt A.3: Caption-grounded sentence… · p. 16

Prompt A.4: Football caption-commentary fusion

  • Football caption-commentary fusion includes only relevant, action-anchored commentary that enhances a chronological narrative while remaining factually bounded by the sources.Standalone tactical analysis and detached background are excluded; the prompt also prohibits external facts and speculation.Source: Prompt A.4: Football… · p. 17

Prompt A.5: Basketball caption-commentary relevance gate

  • The basketball relevance gate accepts commentary tied to caption events, tactics, mentioned players or teams, or related game implications, and rejects unrelated or generic filler.The prompt calls for a semantic judgment based solely on the caption-commentary relationship without over-interpreting vague connections.Source: Prompt A.5: Basketball… · p. 18

Prompt A.6: Basketball caption-commentary fusion

  • Basketball caption-commentary fusion creates a unified broadcast-style narrative by retaining commentary that complements the caption and excluding unrelated or contradictory material.The narrative must use only source-provided information and avoid speculation or invented game details.Source: Prompt A.6: Basketball… · pp. 18–19
  • Basketball caption-commentary fusion instructs the model to preserve caption facts and chronology while adopting the source commentary’s style and avoiding speculation.It also calls for filtering contradictory material and selectively incorporating commentary that enhances the broadcast narrative.Source: Prompt A.6: Basketball… · p. 19

Prompt A.7: Basketball quality selection prompt

  • Basketball selection ranks commentaries by total scores across descriptive, analytical, and background dimensions, returning the two highest-scoring items.The dimensions assess real-time detail, tactical interpretation, and contextual knowledge, respectively.Source: Prompt A.7: Basketball quality… · p. 19

Prompt B.1: Football / Basketball candidate-generation prompt

  • The football and basketball generation prompt combines live event description, strategic analysis, and historical context in a broadcast-style narrative.It also requires segment-only scope, factual grounding, consistent English, and a duration-matched output.Source: Prompt B.1: Football / Basketball… · p. 20

Prompt B.2: Chess / Go candidate-generation prompt

  • The chess and Go prompt guides move-by-move commentary through observation, strategic analysis, and historical recall, while allowing these capabilities to be blended flexibly.Inputs include board or move state and recent history; the prompt also requires factual grounding and fluent, professional English output.Source: Prompt B.2: Chess / Go… · p. 21

Prompt B.3: League of Legends candidate-generation prompt

  • The League of Legends prompt likewise combines live observation, strategic analysis, and historical context, with commentary grounded in the clip and limited to its segment.It specifies authentic esports vocabulary, fluent broadcast delivery, and no invented game details.Source: Prompt B.3: League of Legends… · pp. 21–22

Prompt C.1: Evaluation segmentation and type classification (football)

  • Football evaluation segments commentary into the smallest independent semantic units and assigns each one category based on its primary intent.The rules distinguish observable description, analysis—including pure emotional exclamations—and background facts, with objective data statements classified as background.Source: Prompt C.1: Evaluation segmentation… · pp. 22–23

Prompt C.2: Reference event extraction (football)

  • Football reference-event extraction uses the clip label, processed reference commentary, and match-background XML to identify the main events and describe who was involved and the result when applicable.Source: Prompt C.2: Reference event extraction… · p. 23

Prompt C.2: Reference event extraction (football); Prompt C.3: Candidate event extraction (football)

Prompt C.4: Completeness matching

  • Completeness matching requires a confident match for each reference event with a consistent action, clearly identified subject, and matching result when applicable.The output maps each ground-truth event to a candidate event or null.Source: Prompt C.4: Completeness matching · p. 24

Prompt C.5: Correctness matching

  • Correctness matching uses evidence-conservative criteria: core action and applicable outcome must match, while unsupported critical details or conflicts warrant no match.Flexible actor references and less-specific generalizations are allowed when supported by the ground truth.Source: Prompt C.5: Correctness matching · p. 25

Prompt C.6: Analytical relation standard

  • The analytical relation standard distinguishes CAUSE, REASON, CONDITION, CONTRAST, and ELABORATION relations between propositions.These categories respectively describe production, explanation, hypothetical premises, reversed expectations, and structural unpacking.Source: Prompt C.6: Analytical relation standard · p. 25

Prompt C.7: Analytical proposition extraction

  • Analytical proposition extraction stays within each recorded source sentence, does not fuse sentences, and preserves sentence and proposition order.The full commentary serves only as context for reference resolution, and propositions should be complete units.Source: Prompt C.7: Analytical proposition… · pp. 25–26

Prompt C.8: Analytical edge extraction

  • Analytical edge extraction keeps only clear, necessary, information-bearing forward relations, rejecting loose thematic links and redundant shortcuts.It considers later propositions in batch order and treats propositions uniformly rather than relying on sentence boundaries.Source: Prompt C.8: Analytical edge extraction · p. 26

Prompt C.9: Analytical edge audit

  • Analytical edge auditing retains a candidate only when its direction and relation are supported and the edge is necessary in the final sparse DAG.Temporal adjacency, same-topic continuity, and redundant shortcuts are grounds for deletion; uncertainty also favors deletion.Source: Prompt C.9: Analytical edge audit · p. 27

Prompt C.10: Analytical node and edge truth judgment

  • Node and edge truth judgments use the reference-event list as primary evidence, with commentary only to resolve referents; edges require supported endpoints and a supported directed relation.Unsupported event facts make nodes false, while loose thematic, temporal, or narrative continuity does not establish a true edge.Source: Prompt C.10: Analytical node and edge… · pp. 27–28
  • The analytical judgment prompt requires node and edge label arrays to match the numbers of candidate propositions and final DAG edges, respectively.Its example returns Boolean labels for both arrays.Source: Prompt C.10: Analytical node and edge… · p. 28

Prompt C.11: Background claim extraction (football)

  • Football background extraction calls for medium-granularity, independently judgeable claims supported by the source sentence, without inventing unsupported claims.Context is reserved for disambiguation, entity resolution, and time-scope completion; no claim is returned when none is valid.Source: Prompt C.11: Background claim… · p. 28

Prompt C.12: Background factuality judgment

  • Background factuality judgments are binary, prioritize the local knowledge-base evidence package, and favor false when available evidence does not justify true.Common knowledge may supplement, but not replace, stronger knowledge-base evidence.Source: Prompt C.12: Background factuality… · p. 29

Prompt C.13: Background relevance judgment

  • A background claim is relevant when it meaningfully supports, explains, or contextualizes a descriptive or analytical clause; one claim may match multiple clauses.The prompt requires exact matched clause IDs and empty ID lists when there is no meaningful match.Source: Prompt C.13: Background relevance… · p. 30
Loading 2610.11129v1…