Source-linked AI summary
EvoWiki: Incremental State Overwriting and Traceable Question Answering for Cross-Meeting Knowledge Evolution
Dongsheng Chen, Tianyu Wang, Wenhui Que
TL;DR
Cross-meeting QA must track revised states, preserve their history, and identify the version valid at query time, but existing methods inadequately cover these requirements. EvoWiki uses incremental state overwriting and deterministic Wiki-only reading, improving macro-average Judge Accuracy over the strongest baselines by 9.72 and 10.00 percentage points across six datasets and two readers.
Problem
Existing tasks inadequately cover role binding, version replacement, and multi-hop evidence composition across meetings while requiring identification of the state valid at query time.
Method
EvoWiki incrementally builds versioned, provenance-linked Wiki states with write-time disambiguation and reads valid states deterministically without raw-meeting retrieval.
Results
Across six datasets and two readers, EvoWiki surpasses the strongest baselines by 9.72 and 10.00 percentage points in macro-average Judge Accuracy.
Takeaways & Limitations
EvoWiki supports robust current-state reading and traceable, factually faithful answers with verifiable evidence.
Takeaways & Limitations
Complete-Wiki reading has linear input-context cost, while retaining all historical versions requires O(U) storage.
Abstract
from arXiv · showhide
In long-term collaboration spanning multiple meetings, factual states such as decisions and risks are continually revised, overturned, and replaced. Existing long-context methods typically stack the entire history, while many RAG and structured-memory methods organize knowledge as static or append-only facts and rely on semantic relevance at read time. Without explicit modeling of knowledge lifecycles, these approaches may retain conflicting old and new states simultaneously or discard history, leading to stale retrieval and answers that are difficult to verify. We present EvoWiki, an incremental question-answering architecture for dynamic long-form text. EvoWiki decouples offline incremental construction (BUILD) from online structured reading (READ). BUILD captures the intra-meeting micro-evolution from proposal to decision and uses entity version chains and a fine-grained State-Overwrite Protocol to explicitly distinguish current valid states from superseded history while preserving meeting-level provenance anchors. READ bypasses relevance-based Top-k retrieval and performs deterministic entity addressing, temporal resolution, and cross-entity multi-hop aggregation over the complete Wiki to produce grounded and traceable answers. We further introduce CrossMeet, a high-fidelity bilingual benchmark designed to simulate long-term state evolution, covering factual consistency, temporal reasoning, and cross-meeting multi-hop reasoning. Across six datasets and two reader models, EvoWiki improves macro-average Judge Accuracy over the strongest baselines by 9.72 and 10.00 percentage points, respectively. Human evaluation shows that EvoWiki is more robust and factually faithful under frequent state flips, validating valid-state-oriented reading as a reliable approach to cross-meeting knowledge evolution.
Introduction
Existing approaches inadequately support reliable evidence use, validity-aware replacement, and multi-hop reasoning across evolving meeting knowledge. EvoWiki addresses these gaps through versioned, provenance-preserving state overwriting and deterministic valid-state reading, producing traceable and factually faithful answers.
- Limitations of existing approaches: Long-context models and conventional RAG remain insufficient for reliable evidence use, validity-aware retrieval, and cross-meeting multi-hop reasoning.Long-context windows do not guarantee reliable reasoning, while semantic relevance can treat obsolete and current statements similarly.
- EvoWiki: EvoWiki captures proposal–discussion–decision evolution while maintaining entity versions and provenance through write-time coreference resolution, routing, and state overwriting.Its Build process supports explicit evolution tracking before reading.
- Evaluation: Across six benchmarks and two readers, experiments, state-flip analysis, and human evaluation show reliable valid-state reading with low hallucination risk and factually faithful, verifiable answers.The approach enables traceable reasoning grounded in valid states and verifiable evidence.
Related Work
Related work spans long-context and retrieval-based methods, structured and temporal memory, and meeting-focused understanding and QA. EvoWiki is positioned against these approaches by emphasizing explicit temporal state handling and meeting-level provenance.
- Long-Context and RAG: Long-context models and RAG provide complementary performance–cost profiles, but nominal context-window length does not ensure reliable evidence use.Prior work also studies conversational evaluation, temporal or structured agent memory, meeting understanding, QA, and generation.
- Retrieval and Structured Memory: Classical RAG uses sparse, dense, or hybrid retrieval, while newer systems improve retrieval control or organize memory through wikis, summaries, entity graphs, and associative structures.Examples include Self-RAG, Corrective RAG, Adaptive-RAG, WiCER, RAPTOR, GraphRAG/LightRAG, and HippoRAG 2/LogicRAG.
- Temporal Knowledge and Editing: Temporal retrieval addresses time-sensitive and arriving knowledge, while parametric editing studies fact localization, multi-hop consistency, and lifelong updates but struggles to preserve meeting-level provenance.Existing evaluation also examines factual retrieval and reasoning; the supplied passage positions EvoWiki as an alternative beginning with this provenance challenge.
Methodology
EvoWiki incrementally builds an entity-centered Wiki that explicitly overwrites obsolete states while preserving version histories and provenance. Its deterministic Wiki-only reader resolves valid states at query time and supports traceable temporal and multi-hop answers without source-text Top-k retrieval.
- Incremental construction: EvoWiki processes meetings chronologically, identifying the state valid at query time while retaining historical evidence for state evolution.Each update accesses only the prior Wiki and current meeting, never future meetings.
- Incremental construction: Offline Build aligns time, parses proposal-to-decision micro-evolution, resolves coreference, extracts information, routes entities, and overwrites states.Temporal parsing prevents candidate plans from being written as final states, while write-time coreference consolidates aliases into entity version chains.
- State overwriting: For each entity–attribute pair, the overwrite protocol closes superseded validity intervals, labels old states Overturned, appends new Active states, and maintains at most one active version.Explicit replacement edges preserve superseded versions and provenance for historical queries and audits.
- Structured reading: Online Read serializes the complete Wiki as its sole context, deterministically addresses valid states, and performs temporal or cross-entity multi-hop aggregation without raw-meeting Top-k retrieval.Supporting Wiki-state identifiers remain metadata for evidence-trail recovery, while inference is isolated from raw meetings.
- Complexity and traceability: O(U) storage retains all historical versions, while each query uses O(SW ) serialized-Wiki context and requires no external-corpus retrieval call.The main Build cost occurs offline and can be amortized across subsequent queries.
Experimental Setup
The experiments evaluate EvoWiki on the bilingual CrossMeet benchmark and four public benchmarks under controlled construction, reading, and judging protocols. CrossMeet is validated for answer grounding, meeting-state evolution, metadata reliability, and question-taxonomy agreement.
- Dataset and validation: CrossMeet contains 100 projects, 500 consecutive meetings, and 2,000 QA pairs per language, with average contexts of 26,861 English and 29,480 Chinese tokens.Question-type classification agreement is Fleiss’ κ=0.937 for English and 0.909 for Chinese.
- Dataset and validation: 98.3% of reference answers are supported by supplied meeting evidence, while all validation dimensions exceed a 95% pass rate with agreement of at least 0.84.Human validation sampled 300 instances equally from English and Chinese and assessed answers, cited evidence, and complete evolution paths.
- Evaluation benchmarks: Experiments additionally use MeetingQA, ELITR-Bench, LongMemEval, and MuSiQue, covering meeting-transcript QA, long-meeting retrieval and summarization, session-based memory updates, and static multi-hop reasoning.MeetingQA and ELITR-Bench use human meeting transcripts, whereas LongMemEval and MuSiQue evaluate long-term updates and static compositional reasoning, respectively.
- Compared methods: Nine baselines represent full-context reading, sparse/dense/hybrid retrieval, hierarchical summarization, graph retrieval, associative memory, and query-time logic reasoning.The compared systems include Direct LLM, three VanillaRAG variants, RAPTOR, GraphRAG, LightRAG, HippoRAG 2, and LogicRAG.
- Experimental controls: All methods start from the same raw meetings, while EvoWiki’s reader accesses only its complete Build-produced Wiki under identical questions, prompts, and generation settings.Structured baselines use their recommended offline construction pipelines, and conventional RAG is not given EvoWiki’s intermediate structures.
- Evaluation metric: Judge Accuracy is the primary metric, scored by Gemini 3.1 Pro as 0 for incorrect, 0.5 for correct but incomplete, and 1 for fully correct answers.Each method runs five times with matched online decoding and fixed offline knowledge structures; scores are equally macro-averaged across six datasets.
Main Results
EvoWiki delivers the strongest Judge Accuracy across six datasets and readers, with gains attributed to explicit knowledge-state organization. Ablations show that state overwriting, coreference resolution, entity-centered organization, and structural tags each materially contribute to performance.
- Overall performance: EvoWiki achieves the highest Judge Accuracy on all six datasets, averaging 60.09 with DeepSeek-V4-Flash and 63.02 with Qwen3.5-A17B.These averages exceed LogicRAG by 9.72 and 10.00 points, respectively.
- Overall performance: On LongMemEval, EvoWiki scores 61.60 and 66.00, leading the strongest baselines by 13.10 and 14.10 points.The results support state overwriting for mitigating long-running knowledge conflicts and evidence-localization difficulty.
- Overall performance: Consistent gains across readers indicate that EvoWiki’s improvement stems from knowledge-state organization rather than a particular reader.Its lead on the other five datasets further indicates that the advantage extends beyond LongMemEval.
- Component ablation: Ablations reduce average accuracy from 60.09 to 51.46 without state overwriting, 50.93 without coreference resolution, 53.73 without entity-centered organization, and 49.79 without structural tags.The corresponding drops are 8.63, 9.16, 6.36, and 10.30 percentage points.
Analysis and Discussion
EvoWiki improves efficiency, judge accuracy robustness, evidence-position invariance, and final-state recovery, while its remaining errors concentrate in Build extraction and Read version selection. Human evaluation attributes its advantage to faithfulness and completeness, and low CrossMeet abstention indicates that Wiki-only reading usually recovers valid states.
- Efficiency: 45.0% lower latency and 35.0% lower Read-token usage than Direct LLM demonstrate EvoWiki’s efficiency advantage.Table 7 measures non-streaming end-to-end response time and mean Read tokens per query after offline structures are cached.
- Judge agreement: All three judges rank EvoWiki first, with agreement measured against Gemini 3.1 Pro at κ = 0.901 for Claude Opus 4.7 and κ = 0.887 for DeepSeek-V4-Pro.The evaluation uses 1,500 anonymized responses generated from the same predictions and applies an identical ordinal rubric.
- Failure analysis: Of 212 EvoWiki errors, 54.2% are Build failures and 45.8% are Read failures, chiefly involving extraction omission, overwrite failure, stale-state reading, and incomplete answers.No entity-linking or Wiki statelocation errors are observed; Read failures more often involve obsolete-version selection or insufficient answer coverage.
- State-flip robustness: As state flips increase from 0–1 to four, EvoWiki declines only from 71.03% to 66.90%, a drop of 4.13 percentage points.On the four-flip subset, EvoWiki leads Direct LLM, VanillaRAG, GraphRAG, and LogicRAG by 5.70, 24.10, and additional reported margins.
- Evidence-position robustness: Across LongMemEval evidence-position bins, EvoWiki stays between 63.58% and 70.10%, with a 6.52-point spread and a 20.34-point lead over the runner-up in the middle bin.Entity- and state-organized Wiki reading reduces dependence on evidence position in the original meeting sequence.
- Abstention and human evaluation: EvoWiki’s abstention rates are 0.2% on CrossMeet-EN and 1.1% on CrossMeet-ZH, indicating that Wiki-only Read usually provides sufficient valid states.Every CrossMeet question is answerable, so abstention is erroneous; on the 470 answerable LongMemEval instances, EvoWiki’s abstention rate is 22.8%.
Conclusion
EvoWiki combines incremental state construction with deterministic Wiki-only reading to support current-state, traceable cross-meeting QA, alongside the bilingual CrossMeet benchmark. It outperforms strong baselines across six datasets and two readers, while remaining limited by Build extraction completeness under noisy or informal speech.
- Contributions: EvoWiki combines incremental Build and Wiki-only Read with entity version chains, write-time coreference resolution, state overwriting, and meeting-level provenance.This design maintains a current view while preserving traceable history for dynamic cross-meeting QA.
- Contributions: CrossMeet is a bilingual benchmark covering factual consistency, temporal reasoning, and multi-hop question answering.
- Results: 60.09 and 63.02 macro-average Judge Accuracy surpass the strongest baselines by 9.72 and 10.00 points across six datasets and two readers.State-flip, evidence-position, and human analyses support robust current-state reading and traceability.
- Limitations and Future Work: Wiki-only Read remains constrained by Build extraction completeness, especially under ASR noise and informal speech.Future work will explore uncertainty-aware writing, selective source verification, and broader language coverage.