contradiction detection

This work proposes a novel approach to open-domain claim verification that addresses the limitations of existing fact-checking systems, which often rely on a single knowledge source and thus fail to capture perspective divergence, resulting in limited coverage and poor transparency. The method leverages large language models (LLMs) to simultaneously retrieve multi-source evidence—such as from Wikipedia, PubMed, and Google—for both the original claim and its negation, integrating supporting and contradicting information. It further incorporates cross-source disagreement analysis to better model the complexity and diversity of information. By combining evidence deduplication, confidence scoring, and visualization, the approach is evaluated across four benchmark datasets using five distinct LLMs, achieving significant accuracy improvements and revealing notable differences in how various knowledge sources contribute to reasoning.

This work addresses the challenge faced by large language model (LLM) memory systems in retrieving the most up-to-date and accurate facts when confronted with temporally evolving contradictory information. To resolve this, the authors propose a deterministic memory conflict resolution method that explicitly identifies the bottleneck in the post-retrieval assembly phase. Instead of relying on LLM-based judgment, the approach employs deterministic aggregation strategies—such as max(serial) or max(timestamp)—after candidate fact extraction. Integrated with Self-Ask multi-hop reasoning and a unified retrieval chunking mechanism, the method forms an end-to-end optimized pipeline. Experimental results demonstrate significant improvements: achieving 78.0% accuracy with gpt-4o-mini and 94.8% with gpt-4o on the single-hop FactConsolidation task, and 51.5% on the multi-hop setting with gpt-4o, outperforming existing systems by 20–28 percentage points.

This work identifies and formally names a novel phenomenon in decomposed fact-checking—termed “Decomposition-Induced Context-Memory Conflict” (DI-CC)—which arises when the decomposition stage introduces model-generated beliefs that contradict the original text, thereby producing erroneous atomic claims. The study demonstrates that while DI-CC operates via the same mechanism as classical context-memory conflict, it occurs at a distinct procedural stage. Experimental evaluations employing linear probing, SelfCheckGPT-style self-consistency sampling, and context-aware decoding reveal that linear probes effectively detect DI-CC even on unseen decompositions (AUC = 0.86–0.88), whereas self-consistency methods fail (AUC = 0.51). Although context-aware decoding mitigates DI-CC, it substantially reduces parsing accuracy on texts involving complex anaphora.

Existing metrics for factuality and faithfulness struggle to evaluate how language models handle documents containing both supporting and contradictory evidence. This work proposes ConflictScore, the first formal and quantitative framework for assessing a model’s ability to recognize and articulate conflicting evidence. It decomposes model responses into atomic claims, fine-grained labels their relationships with all source documents, and introduces two complementary metrics: CS-C (Conflict Sensitivity) and CS-R (Response Reasonableness). Built upon this framework, the ConflictBench benchmark encompasses diverse conflict types. Experiments demonstrate that ConflictScore effectively identifies overconfident claims across domains and serves as a feedback signal that significantly improves model truthfulness on TruthfulQA.

In digital forensics, the atomicity and integrity of storage snapshots lack rigorous definitions that jointly guarantee both instantaneousness and causal ordering—undermining evidentiary admissibility in legal proceedings. To address this, we propose a novel atomicity definition grounded in causal consistency, overcoming the limitation of conventional time-based atomicity models. We further rectify conceptual flaws in existing integrity definitions and introduce a revised, theoretically sound yet engineering-practical integrity criterion—explicitly supporting copy-on-write (CoW) implementations. Our approach integrates causal modeling, formal snapshot semantics, CoW mechanism analysis, and formalization of forensic quality criteria, yielding a verifiable snapshot semantic framework. This work establishes the first theoretical foundation for forensic tool design that unifies causal ordering with instantaneous state capture, thereby significantly enhancing the forensic validity and judicial admissibility of live data acquisition.

This study addresses a critical issue in large language model (LLM) agents wherein tentative statements are erroneously solidified into definitive facts during memory compression, leading subsequent reasoning to rely uncritically on unverified information. The authors identify this “confidence fabrication” phenomenon as stemming from the linguistic certainty of phrasing rather than the reliability of the source itself. Through multi-turn dialogues, memory mechanism analysis, and adversarial memory injection, they demonstrate that a single misleading memory can induce erroneous judgments, yet introducing just one additional corroborating source suffices to restore accurate decision-making. Furthermore, conventional approaches such as trust labels or explicit negation instructions prove not only ineffective but potentially detrimental, exacerbating the problem rather than mitigating it.

This work addresses the persistent degradation of agent reasoning and tool use caused by erroneous memories—such as contamination, staleness, or misattribution—which existing approaches struggle to correct without discarding valid knowledge. The paper formalizes, for the first time, the post-failure memory recovery problem and introduces a dependency-guided rollback repair mechanism. By constructing a typed memory-action dependency graph, the method tracks downstream effects at runtime, selectively deactivates unreliable memories, and replays only those computations relevant to the final answer. Evaluated on a controlled benchmark of 150 cases, the approach achieves an 85.3% recovery rate—surpassing the best baseline (77.3%)—while fully eliminating error sources and preserving all benign memories. In 50 stress-test scenarios, it attains a 68.0% recovery rate and significantly outperforms baselines, achieving the highest statement invalidation F1 score of 0.669.

This work addresses memory contamination in language agents caused by the retention of outdated information in long-term memory, which degrades decision accuracy. The authors propose TEPA, a novel mechanism that introduces revocable memory lifecycle management: observations are represented as keyed precedents, and conflicts between new evidence and existing memories are dynamically detected. Upon detecting such conflicts, TEPA revokes invalidated memories, ensuring retrieval is always grounded in the most current and valid knowledge. This approach enables dynamic falsification, auditability, and reactivation of memories. Evaluated across diverse memory drift scenarios, TEPA achieves an accuracy of 0.950, substantially outperforming conventional strategies that rely solely on append-only or overwrite-based memory updates.

Current evaluations of long-term memory in large language models predominantly rely on accuracy metrics for isolated questions, which fail to capture the dynamic behavior of user-specific facts under varying conditions. This work proposes MemTrace, a novel benchmark that shifts the evaluation unit from individual questions to knowledge points and constructs multidimensional, controllable probes along three axes: memory age, question type, and evidence condition. Using this framework, we systematically evaluate 13 system configurations across four memory paradigms and reveal that high overall accuracy masks significant deficiencies in tracking factual evolution and correcting erroneous premises. These failures primarily stem not from missing retrieval but from an inability to effectively leverage available evidential context.

Current evaluation frameworks struggle to characterize the arbitration behavior of large language models (LLMs) in retrieval-augmented generation (RAG)-based fact verification, particularly when conflicts arise between parametric priors and retrieved evidence. This work proposes PAVE, a diagnostic platform that introduces the first prior-aware verifier evaluation framework. It categorizes models into four cognitive states based on the correctness and confidence of their priors and quantifies their arbitration strategies using Jensen-Shannon divergence (JSD). Experiments across seven mainstream LLMs reveal that existing verifiers exhibit highly model-dependent and unreliable arbitration behavior. The proposed lightweight JSD-based arbitration method, which requires no model modification, significantly improves fact-checking accuracy and demonstrates robust performance across diverse model families.