Pith. sign in

REVIEW 2 major objections 4 minor 49 references

Fine-grained supersession typing adds no measurable accuracy for current-state queries; the apparent edge of a deprecation-aware ledger over flat retrieval is almost entirely presentation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 21:34 UTC pith:PIZ4AKPC

load-bearing objection Solid empirical correction of how structured-memory systems get evaluated; the render-matched control isolates presentation from mechanism, and the near-zero fine-grained mechanism residual holds up under the checks that matter. the 2 major comments →

arxiv 2607.16019 v1 pith:PIZ4AKPC submitted 2026-07-17 cs.LG

Presentation, Not Mechanism: A Render Confound in Deprecation-Aware Memory Evaluation

Zhaoyang Jiang , Zhizhong Fu , Zicheng Li , Yunsoo Kim , Jiacong Mi , Xuanqi Peng , Fei Teng , Honghan Wu This is my paper
classification cs.LG
keywords evidence-state revisiondeprecation-aware memoryrender confoundrender-matched controlquery sufficiencyreverted-revertretention vs relation typingLLM-as-judge evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that a fine-grained, deprecation-aware memory — one that tracks relation types, supersessions, and conflict status — gains almost nothing over flat retrieval once the way evidence is presented is held fixed. On reverted-revert questions, where a value is changed and later restored, the fine ledger appears to beat a flat baseline by +0.182, but a render-matched control (the same layout with deprecation disabled) accounts for +0.159 of the gap, leaving a mechanism residual that is statistically indistinguishable from zero. What actually pays for current-value questions is a coarse binary live/dead invalidation marker (+0.087), and for provenance questions it is retaining invalidated evidence, not richer typing. If the paper is right, memory evaluations must separate layout from mechanism, and systems should deploy the coarsest retained state that covers their query mixture — a finding that would re-attribute many apparently architecture-driven gains in agent memory.

Core claim

On reverted-revert questions — a value changes then is restored — a fine-grained ledger seems to beat a flat baseline by +0.182. A render-matched control (same layout, deprecation disabled) shows +0.159 is presentation; the fine mechanism residual of +0.025 (CI [-0.005, +0.057]) is indistinguishable from zero. Coarse edge invalidation contributes a real +0.087, so for current-value queries a binary live/dead marker — not fine relation typing — is the load-bearing state. For provenance queries, discarding invalidated edges collapses accuracy; retaining them restores it whether the store is coarse or fine. The query-sufficiency principle: for current-value queries, the binary deprecation parti

What carries the argument

The central instrument is the render-matched control: a transformation that takes a structured memory's exact layout and switches off the deprecation logic (forcing every relation to same-state), so layout is held fixed while mechanism is removed. Comparing the full system, the control, and the flat baseline under the same judge decomposes any apparent gain into a render term and a mechanism residual. The paper's theoretical backbone is the query-sufficiency lattice: for current-value (snapshot) queries, the binary live/dead partition of the evidence is a minimal sufficient statistic; the five-way relation typing carries zero additional information (b_type = 0), and retention of invalidated

Load-bearing premise

The decomposition rests on the paper's own LLM-labelled, LLM-judged benchmark whose phenomenon labels show only moderate human agreement — the labeller over-labels the reverted-revert stratum, half of which humans call monotonic — and there is no independent render-matched replication in the high-clutter prose regime where the render effect is claimed to be largest.

What would settle it

Run the same render-matched decomposition on a human-gold benchmark of real revision streams (software issues or encyclopedia history) with exact-match scoring and no LLM judge; if the fine-grained mechanism residual over the render-only control is clearly positive (e.g., >+0.05) on current-value questions, the paper's central claim fails. Concretely, the authors' own canonical 25-item human-validated subset is the smallest place to look: a larger canonical sample with tighter labels would test whether the near-zero residual is an artifact of the noisy stratum.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Memory-architecture evaluations should hold the rendered layout fixed when comparing a structured store to a flat baseline; otherwise reading ease will be counted as mechanism.
  • For current-state questions, systems need not build fine-grained supersession typing; a coarse invalidation marker that says which evidence is live versus dead captures the decision-relevant state.
  • Provenance questions require retaining invalidated evidence: a store that discards superseded edges cannot answer 'what was in force just before?', and retention alone — without relation labels — restores most of the capability.
  • The 'best memory' question is ill-posed without a stream prior and query mixture; the paper's selection rule is to deploy the coarsest retained state that covers the queries asked.
  • The render-matched protocol is specified as a reusable transformation so future comparisons can separate layout from mechanism.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: If the render confound is as large as measured here, results from earlier structured-memory comparisons that changed mechanism and presentation together may need re-checking; the paper itself notes this as a field gap rather than a settled conclusion.
  • Inference: A testable extension: sweep layout clarity (ordering, field separation, token count) while keeping the mechanism fixed on a high-clutter stream, and measure how much of the flat-vs-structured gap follows the layout curve; this would tell whether the render term is reducible to known position-bias effects.
  • Inference: The retention-over-typing result suggests a concrete design rule for privacy-constrained memory: budget for keeping the binary invalidation history (or a bounded summary) even when raw text is dropped; the paper's rate-distortion framing makes this a rate-allocation problem.
  • Inference: The query-sufficiency lattice could be inverted into a diagnostic: given a task's query mixture, one could pre-compute which statistics are necessary, and only build the memory tier that covers them; the paper demonstrates this on three query classes but the general routing rule remains untested.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper studies Evidence-State Revision (ESR): retrieval over self-revising records where a system must decide which claim is currently in force and when to abstain. It compares a d-blind GraphRAG+abstain baseline, a coarse-d edge-invalidation store (Graphiti-style), and a fine-d RevisionLedger with typed relations, on 2,907 ESR-Bench questions. A render-matched control (RevisionLedger's layout with deprecation disabled) decomposes the apparent +0.182 reverted-revert advantage into a +0.159 render term and a +0.025 fine-d mechanism residual, which is statistically indistinguishable from zero across two judge families; coarse-d invalidation contributes +0.087. Provenance experiments show that discarding invalidated edges is fatal and retention (not typing) restores answerability, replicating on TempLAMA with Wikidata gold. The paper concludes that memory evaluations should hold render fixed and that systems should deploy the coarsest retained state that covers the query mixture.

Significance. If the results hold, the paper makes a valuable methodological contribution: it demonstrates a concrete render confound in structured-memory evaluation and provides a reusable render-matched protocol. Strengths include the arithmetic decomposition, paired bootstrap intervals, two judge families, a human-canonical subset, controlled noise sweeps that falsify a prior sign-flip prediction, and an external TempLAMA replication with exact-match scoring and no LLM judge. The paper is also unusually explicit about its limitations (label noise κ=0.459, single architecture family, no independent render-matched replication in high-clutter prose, and the falsification of P4's predictive claim). These strengths justify publication once the coarse-d render-matching point is resolved.

major comments (2)
  1. [§4.1, Table 2] The 'Coarse-d mech.' term is defined as Graphiti − control and is described as a mechanism effect, but the control is explicitly render-matched to RevisionLedger's layout ('the fine-d ledger's exact structured layout'), not to Graphiti's layout. If Graphiti's serialisation differs from the control's — different columns, ordering, or status rendering — the +0.087 [0.046,0.130] term conflates render with mechanism. This is load-bearing for the abstract's claim that 'coarse invalidation is the only mechanism that pays' and for the practical recommendation to deploy a coarse retained state. Please state explicitly whether Graphiti was rendered with the same template as the control, or add a render-matched coarse control and re-derive Table 2. If it was not render-matched, the conclusion should be softened accordingly.
  2. [Abstract and §4.1] The fine-d mechanism residual is reported as 'indistinguishable from zero' with 95% CI [-0.005,+0.057]. This is statistically correct, but the upper bound leaves room for a small typed-mechanism benefit. The abstract's 'adds no measurable accuracy' and 'no measurable supersession benefit' are stronger than the interval supports. I suggest reporting the residual as 'not statistically distinguishable from zero' and, if the stronger claim is intended, providing an equivalence test or a pre-specified margin. This is a presentation issue rather than a flaw, but it affects how readers interpret the headline.
minor comments (4)
  1. [§4.3] 'DyKnown=11' appears to be a typo for 'DyKnow n=11'.
  2. [§3 vs. §4.1/§F] Section 3 says all main experiments use DeepSeek-V4-flash, while §4.1 and §F identify DeepSeek-V3 as the primary judge for the main analyses and render decomposition. Please harmonize the judge specification across these sections.
  3. [Appendix P4] The P4 priority order D≻U≻Fr≻S≻O is a modelling choice rather than a derived result. The paper correctly demotes P4 after its predictive claim is falsified, but it would help to flag the priority order explicitly as an assumption in the appendix statement.
  4. [Table 2] The table would be easier to read if it also reported the raw accuracies of the four systems, not only the three derived terms. This would let readers verify the arithmetic decomposition directly.

Circularity Check

0 steps flagged

No significant circularity; central result is a measured ablation, not a fitted prediction.

full rationale

The headline render/mechanism decomposition is arithmetic on measured accuracies, not a fitted or definitional claim: RL−GRW = (control−GRW)+(RL−control), with +0.184 ≈ +0.159 + +0.025 under DeepSeek-V3, and the render-matched control is explicitly constructed by disabling deprecation while holding layout fixed, so the mechanism residual is operationally defined rather than equivalent to the input. The theoretical statements that might appear definitional are explicitly flagged as such: Proposition 2b is described as 'clarifying rather than deep' and its btype=0 follows from the definition of a snapshot value; Proposition 5 is explicitly labeled retrodictive; Proposition 4's only predictive sign claim is reported as falsified, with the paper stating the granularity ranking rests on the render-matched decomposition and Corollary 1. There are no load-bearing self-citations, no imported uniqueness theorems from the authors' prior work, and no ansatz smuggled in via citation. The retention finding is externally replicated on TempLAMA with Wikidata gold, exact-match scoring, and no LLM judge. Limitations about the self-built benchmark and label noise are disclosed external-validity caveats, not internal circularity. No step reduces by construction to its own inputs in a way that would justify a nonzero circularity score.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim is a measured ablation (render-matched control), so the headline numbers are not produced by fitted constants; the only hand-chosen item is the diagnostic channel-priority order, which the paper does not lean on. The theory is conditional (oracle extraction for P2a/P2b; stated premises for P5), and the authors disclose that Proposition 5 is retrodictive and that their one pre-committed quantitative prediction (C3) was falsified. The evidential load is borne by the controlled decomposition, two judge families, the human-canonical subset, and the TempLAMA replication of the retention result. RevisionLedger is an implemented system and the binary partition Pi is a mathematical derivative of the deprecation marker — neither is a postulated entity needing independent evidence.

free parameters (1)
  • P4 decision-channel priority order (D≻U≻Fr≻S≻O) = hand-chosen
    Routing rule for the diagnostic channel decomposition (Appendix A, Proposition 4); the paper states it is a priority assignment and that the headline S-channel test is not a reassignment of the channel partition. Diagnostic only — the granularity conclusion explicitly rests on P2b and Corollary 1, not on this order.
axioms (4)
  • standard math Data-processing inequality and Fano's inequality bound the pipeline bottleneck (Proposition 5's ceiling argument)
    Used in Proposition 5 for the per-stage accuracy ceiling; standard information-theoretic results (Cover 1999; Wan et al. 2025).
  • domain assumption Oracle assumptions of the sufficiency results: exact atom extraction, exact co-keying, complete/sound deprecation marker d, single-valued keys, functional acyclic deprecation applied in transaction order
    Stated in Proposition 2a (Appendix A). The paper concedes the real extractor misses ~1 in 7 gold events, so P2a/P2b are oracle-level statements; the empirical claims carry the evidential load.
  • domain assumption Render monotonicity: answer-model decoding capacity is non-increasing in render length/complexity
    Premise (R) of Proposition 5, supported by long-context position bias (Liu et al. 2024b) and distractor sensitivity (Cuconasu et al. 2024). Not a theorem; the paper labels it an assumption.
  • domain assumption The reverted-revert stratum realizes P1's ambiguous fibre with a near-symmetric prior
    Empirical anchor for the 50% ceiling: 38 canonical vs 38 non-canonical reverts among the human-validated LLM-flagged subset (§3, Appendix A). If the realized prior on the stratum were skewed, the ceiling would be vacuous in practice.

pith-pipeline@v1.3.0-alltime-deepseek · 35310 in / 15614 out tokens · 154663 ms · 2026-08-01T21:34:53.010415+00:00 · methodology

0 comments
read the original abstract

AI systems increasingly retrieve from records that revise themselves: issue threads, encyclopedic histories, policy logs, and long conversations. The challenge is not only finding relevant evidence, but deciding which claims remain in force, which were superseded, and when to abstain. Structured memories promise to solve this with typed edges, temporal updates, and conflict status, yet evaluations often change mechanism and prompt presentation together. We study this as Evidence-State Revision, comparing flat retrieval, coarse edge invalidation, and fine-grained RevisionLedger on 2,907 high-agreement questions from GitHub, multi-repo issue histories, Wikipedia, and DyKnow-style temporal streams. A render-matched control (same layout, deprecation disabled) reveals the central confound: when a value is changed and later restored, RevisionLedger appears to beat a flat baseline by +0.182, but almost all the gain comes from easier presentation; the fine-grained mechanism residual is indistinguishable from zero (+0.021 to +0.025 across two judge families). After presentation is controlled, coarse invalidation is the only mechanism that pays for current-state queries, beating the fine ledger by 0.084; the same query-sufficiency principle says provenance mainly needs retained invalidated evidence, not richer typing. Memory evaluations should hold render fixed, and deprecation-aware systems should deploy the coarsest retained state that covers their queries.

Figures

Figures reproduced from arXiv: 2607.16019 by Fei Teng, Honghan Wu, Jiacong Mi, Xuanqi Peng, Yunsoo Kim, Zhaoyang Jiang, Zhizhong Fu, Zicheng Li.

Figure 1
Figure 1. Figure 1: The render confound. In a reverted-revert stream (30s → 60s → 30s), RevisionLedger answers the current value while the flat baseline answers uncertain (an abstention). A render-matched control keeps the ledger layout but disables deprecation labels; it still answers 30s. Thus the apparent ledger gain can come from readable, time-ordered presentation rather than fine-grained supersession typing; §4.1 quanti… view at source ↗
Figure 2
Figure 2. Figure 2: Conflict locality determines memory need. Intra-passage contradictions (L1) can be settled by reading; cross-passage conflicts (L2) require provenance and valid time to associate sources and abstain when unresolved; temporal entity-state revision (L3) requires deprecation. Tiers are a priority ordering: the highest applicable tier’s state subsumes the lower ones through its conflict-status field. The L3 mi… view at source ↗
Figure 3
Figure 3. Figure 3: Provenance needs retention, not typing. Accuracy on provenance queries (“what value held just before?”) from retained state only, pooled over five answer-LLM families (DeepSeek judge, n=4,895). A coarse store that discards closed edges cannot answer the prior-value query; retaining them, whether coarse or fine, restores answerability. The fine-d bar is therefore an apparent edge over discard, not evidence … view at source ↗
Figure 4
Figure 4. Figure 4: A binary live/dead marker is enough for current-value queries (P1/P2b). On the three-event chain e1:ϕ, e2:¬ϕ, e3 superseding e2, a d-blind store cannot represent which earlier event was deprecated, so it cannot distinguish this resolved revert from an unresolved conflict and must give the same answer to both surfaces (P1, necessity). A store that records only binary live/dead status marks e2 dead and retur… view at source ↗
Figure 5
Figure 5. Figure 5: No tested noise setting gives fine-d a win. Fine-d − coarse-d accuracy on strict reverted-revert under two corruption axes (relation-label noise; atom￾extraction dropout), five answer families, n=445/level, with 95% CI error bars. Negative values mean coarse-d wins; the point estimates remain below zero at every tested level and no setting yields a significant fine-d win (some CIs cross 0, i.e. ties). The … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 1 canonical work pages

  1. [1]

    and G \"a rdenfors, Peter and Makinson, David

    Alchourr \'o n, Carlos E. and G \"a rdenfors, Peter and Makinson, David. On the Logic of Theory Change: Partial Meet Contraction and Revision Functions. Readings in Formal Epistemology: Sourcebook. 2016. doi:10.1007/978-3-319-20451-2_13

    work page doi:10.1007/978-3-319-20451-2_13 2016

  2. [2]

    Katsuno, Hirofumi and Mendelzon, Alberto O. , year=. On the difference between updating a knowledge base and revising it , booktitle=

  3. [3]

    Journal of Logic and computation , volume=

    Merging information under constraints: a logical framework , author=. Journal of Logic and computation , volume=. 2002 , publisher=

    2002

  4. [4]

    1999 , publisher=

    Developing time-oriented database applications in SQL , author=. 1999 , publisher=

    1999

  5. [5]

    arXiv preprint arXiv:2404.16130 , year=

    From local to global: A graph rag approach to query-focused summarization , author=. arXiv preprint arXiv:2404.16130 , year=

    Pith/arXiv arXiv

  6. [6]

    , author=

    MemGPT: towards LLMs as operating systems. , author=. 2023 , journal=

    2023

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building production-ready ai agents with scalable long-term memory , author=. arXiv preprint arXiv:2504.19413 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2501.13956 , year=

    Zep: a temporal knowledge graph architecture for agent memory , author=. arXiv preprint arXiv:2501.13956 , year=

    Pith/arXiv arXiv

  10. [10]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Knowledge conflicts for llms: A survey , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  11. [11]

    Advances in Neural Information Processing Systems , volume=

    Wikicontradict: A benchmark for evaluating llms on real-world knowledge conflicts from wikipedia , author=. Advances in Neural Information Processing Systems , volume=

  12. [12]

    the moon is made of marshmallows

    Faitheval: Can your language model stay faithful to context, even if" the moon is made of marshmallows" , author=. International Conference on Learning Representations , volume=

  13. [13]

    arXiv preprint arXiv:2410.10813 , year=

    Longmemeval: Benchmarking chat assistants on long-term interactive memory , author=. arXiv preprint arXiv:2410.10813 , year=

    Pith/arXiv arXiv

  14. [14]

    International Conference on Machine Learning , pages=

    Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  15. [15]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    ContraDoc: Understanding self-contradictions in documents with large language models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  16. [16]

    arXiv preprint arXiv:2412.19437 , year=

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    Pith/arXiv arXiv

  17. [17]

    Transactions of the Association for Computational Linguistics , volume=

    Time-aware language models as temporal knowledge bases , author=. Transactions of the Association for Computational Linguistics , volume=. 2022 , publisher=

    2022

  18. [18]

    1999 , publisher=

    Elements of information theory , author=. 1999 , publisher=

    1999

  19. [19]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

  20. [20]

    International conference on learning representations , volume=

    Conformal risk control , author=. International conference on learning representations , volume=

  21. [21]

    arXiv preprint arXiv:2512.12844 , year=

    Selective Conformal Risk Control , author=. arXiv preprint arXiv:2512.12844 , year=

    Pith/arXiv arXiv

  22. [22]

    International conference on artificial intelligence and statistics , pages=

    AUC-based selective classification , author=. International conference on artificial intelligence and statistics , pages=. 2023 , organization=

    2023

  23. [23]

    IRE Transactions on Electronic Computers , number=

    An optimum character recognition system using decision functions , author=. IRE Transactions on Electronic Computers , number=. 1957 , publisher=

    1957

  24. [24]

    Journal of Machine Learning Research , volume=

    Optimal strategies for reject option classifiers , author=. Journal of Machine Learning Research , volume=

  25. [25]

    International Conference on Learning Representations , volume=

    Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts , author=. International Conference on Learning Representations , volume=

  26. [26]

    2003 , publisher=

    Temporal data & the relational model , author=. 2003 , publisher=

    2003

  27. [27]

    Advances in neural information processing systems , volume=

    Hipporag: Neurobiologically inspired long-term memory for large language models , author=. Advances in neural information processing systems , volume=

  28. [28]

    , author=

    On the Foundations of Noise-free Selective Classification. , author=. Journal of Machine Learning Research , volume=

  29. [29]

    International Conference on Machine Learning , pages=

    Classification with rejection based on cost-sensitive classification , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  30. [30]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    The power of noise: Redefining retrieval for rag systems , author=. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  31. [31]

    arXiv preprint arXiv:2408.12076 , year=

    Conflictbank: A benchmark for evaluating the influence of knowledge conflicts in llm , author=. arXiv preprint arXiv:2408.12076 , year=

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2403.09040 , year=

    Ragged: Towards informed design of scalable and stable rag systems , author=. arXiv preprint arXiv:2403.09040 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2605.00356 , year=

    MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agents , author=. arXiv preprint arXiv:2605.00356 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2307.02738 , year=

    Recallm: An adaptable memory mechanism with temporal understanding for large language models , author=. arXiv preprint arXiv:2307.02738 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2510.13590 , year=

    RAG Meets Temporal Graphs: Time-Sensitive Modeling and Retrieval for Evolving Knowledge , author=. arXiv preprint arXiv:2510.13590 , year=

    arXiv

  36. [36]

    arXiv preprint arXiv:2603.17244 , year=

    Graph-native cognitive memory for AI agents: Formal belief revision semantics for versioned memory architectures , author=. arXiv preprint arXiv:2603.17244 , year=

    arXiv

  37. [37]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Stronger baselines for retrieval-augmented generation with long-context language models , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Abstentionbench: Reasoning llms fail on unanswerable questions , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    Transactions of the Association for Computational Linguistics , volume=

    Know your limits: A survey of abstention in large language models , author=. Transactions of the Association for Computational Linguistics , volume=. 2025 , publisher=

    2025

  40. [40]

    2001 , publisher=

    Causality: models, reasoning, and inference , author=. 2001 , publisher=

    2001

  41. [41]

    Advances in neural information processing systems , volume=

    Investigating gender bias in language models using causal mediation analysis , author=. Advances in neural information processing systems , volume=

  42. [42]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  43. [43]

    arXiv preprint arXiv:2509.21199 , year=

    A fano-style accuracy upper bound for llm single-pass reasoning in multi-hop qa , author=. arXiv preprint arXiv:2509.21199 , year=

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint physics/0004057 , year=

    The information bottleneck method , author=. arXiv preprint physics/0004057 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2604.18478 , year=

    WorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time Reconciliation , author=. arXiv preprint arXiv:2604.18478 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2604.22085 , year=

    Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents , author=. arXiv preprint arXiv:2604.22085 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2604.18362 , year=

    ArbGraph: Conflict-Aware Evidence Arbitration for Reliable Long-Form Retrieval-Augmented Generation , author=. arXiv preprint arXiv:2604.18362 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2604.02431 , year=

    SelRoute: Query-Type-Aware Routing for Long-Term Conversational Memory Retrieval , author=. arXiv preprint arXiv:2604.02431 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2512.12856 , year=

    Forgetful but faithful: A cognitive memory architecture and benchmark for privacy-aware generative agents , author=. arXiv preprint arXiv:2512.12856 , year=

    arXiv