pith. sign in

arxiv: 2606.20959 · v1 · pith:GDVKQY2Unew · submitted 2026-06-18 · 💻 cs.LG · cs.CL

Right Knowledge, Wrong Answer: Test-Time Steering for Temporal Fact Conflicts in Open-Weight Language Models

Elias Hossain , Sourav Saha , Umesh Chandra Biswas , Sanjeda Sara Jennifer This is my paper

Pith reviewed 2026-06-26 17:32 UTC · model grok-4.3

classification 💻 cs.LG cs.CL
keywords Parametric Temporal ConflictTemporal Attractor Steeringtest-time interventionactivation patchinglanguage modelsknowledge conflictsinference-time steeringopen-weight models
0
0 comments X

The pith

Temporal Attractor Steering overrides outdated facts in language models at inference time by steering hidden states in a conflict-critical layer.

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

The paper formalizes cases where language models store both outdated and newer facts parametrically yet return the outdated answer under standard prompting. It introduces Temporal Attractor Steering, a three-stage test-time method that detects conflicts, locates a conflict-critical layer, and steers activations toward newer-fact representations. On an 8,746-record benchmark spanning five Wikidata relations and four open-weight models, single-layer patching flips answers in 72-85 percent of cases. End-to-end TAS resolves 29-57 percent of conflicts while retaining 85-99 percent accuracy on non-conflict queries and beats a matched ITI baseline on three of four models. The work shows that outdated parametric knowledge can be selectively overridden without retraining or retrieval.

Core claim

Large language models can store both outdated facts and newer superseding facts in their parameters, but standard prompting may still elicit the outdated answer. We formalize this problem as Parametric Temporal Conflict (PTC) and introduce Temporal Attractor Steering (TAS), a three-stage test-time intervention that detects likely conflicts, identifies a conflict-critical layer, and steers hidden states toward newer-fact representations without retraining or external retrieval. We construct an 8,746-record verified benchmark across five Wikidata relations and evaluate four open-weight language models from three families: Qwen-2.5-1.5B/7B, Mistral-7B-v0.3, and Llama-3.1-8B. Single-layer activa

What carries the argument

Temporal Attractor Steering (TAS), a three-stage test-time intervention of conflict detection, conflict-critical layer identification, and hidden-state steering toward newer-fact representations.

If this is right

  • Single-layer activation patching achieves answer-flip rates of 0.72-0.85 across all models.
  • End-to-end TAS resolves 29-57% of PTC cases.
  • TAS preserves 85-99% accuracy on non-conflict queries.
  • TAS outperforms a matched ITI baseline on three of four models.
  • Outdated parametric knowledge can be selectively overridden at inference time.

Where Pith is reading between the lines

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

  • If critical layers prove consistent across model families, similar steering could apply to other internal knowledge inconsistencies.
  • The localization of facts implied by the method suggests runtime correction could supplement static training in deployed systems.
  • Models equipped with such steering might maintain temporal accuracy longer without periodic retraining.

Load-bearing premise

There exists an identifiable conflict-critical layer whose hidden-state steering can override the outdated fact representation without unintended degradation to other stored knowledge or non-conflict performance.

What would settle it

An experiment showing that steering the identified layer reduces accuracy on non-conflict queries or unrelated facts would disprove selective override.

Figures

Figures reproduced from arXiv: 2606.20959 by Elias Hossain, Sanjeda Sara Jennifer, Sourav Saha, Umesh Chandra Biswas.

Figure 1
Figure 1. Figure 1: Overview of Parametric Temporal Conflict ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmark relation distribution (N=8,746 records, five Wikidata relations, three domains). The two smallest relations, P35 (head of state, n=183) and P169 (CEO, n=208), carry the strongest per-record PTC signal across all four models (Section 6.2). Qwen-2.5 1.5B Qwen-2.5 7B Mistral-7B v0.3 Llama-3.1 8B 0.50 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 OPR (outdated-preference rate) 0.530 0.544 0.540 0.540 (a) O… view at source ↗
Figure 3
Figure 3. Figure 3: Phase 1 screening on the 8,746-record benchmark. (a) The outdated-preference rate (OPR) is approx￾imately family-invariant across the four models, all within the [0.530, 0.544] band. (b) The filtered PTC rate increases monotonically with parameter count (0.041 → 0.071 → 0.085 → 0.103), where Kept is the fraction of records passing the knowledge-recovery filter (log P(anew | q, ctemp) ≥ −3; 28.2% on Qwen-1.… view at source ↗
Figure 4
Figure 4. Figure 4: Per-era raw PTC rate. The 2020–2021 bucket is the peak for the three 7–8B-class models (stars); Llama-3.1-8B uniquely retains a substantial 2022–2024 rate, consistent with its more recent training cutoff. 6.3 Layer Localization via Patching [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pairwise and four-way PTC-positive overlap on the 8,734 records common to all four runs. The Mis￾tral ∩ Llama-3.1-8B intersection (193) is the largest pair￾wise overlap and substantially exceeds the two-Qwen overlap (38), even though the two Qwens share archi￾tecture and training data. Govindan et al. (2025) steer unconditionally with￾out separating knowledge absence from conflict or reporting PA; Kang et … view at source ↗
Figure 6
Figure 6. Figure 6: Capacity-scaling of the knowledge-recovery filter. Each point is a model placed at its measured (kept [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-relation raw PTC rate across the four models. Each bar shows the fraction of records of a given Wikidata relation on which the model both prefers aold under standard prompting and recovers anew under the temporal cue. The ordering P35 > P169 > {P286, P488, P6} is preserved across all four models, making the small-but-dense P35 (head of state, n=183) and P169 (CEO, n=208) relations the dominant per-reco… view at source ↗
Figure 8
Figure 8. Figure 8: Per-layer answer-flip rate (AFR) for all four models, with the [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Stage 1 detector precision-recall curves on each model’s held-out evaluation fold. Each detector is a [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Oracle α-sweep at ℓ ∗ with the V2 per-relation ∆ and no detector gate (steering is applied to every instance). (a) Recovery on the verified PTC subset rises monotonically with the steering scale α on all four models and saturates by α ≈ 4 on the two Qwens; Mistral and Llama-3.1-8B keep climbing more slowly. Stars mark each model’s α ∗ = arg maxα [Recovery − λ(1 − PA)] with λ = 1.0. (b) Preservation accura… view at source ↗
read the original abstract

Large language models can store both outdated facts and newer superseding facts in their parameters, but standard prompting may still elicit the outdated answer. We formalize this problem as Parametric Temporal Conflict (PTC) and introduce Temporal Attractor Steering (TAS), a three-stage test-time intervention that detects likely conflicts, identifies a conflict-critical layer, and steers hidden states toward newer-fact representations without retraining or external retrieval. We construct an 8,746-record verified benchmark across five Wikidata relations and evaluate four open-weight language models from three families: Qwen-2.5-1.5B/7B, Mistral-7B-v0.3, and Llama-3.1-8B. Single-layer activation patching achieves answer-flip rates of 0.72-0.85 across all models. End-to-end TAS resolves 29-57% of PTC cases while preserving 85-99% accuracy on non-conflict queries, outperforming a matched ITI baseline on three of four models. These results show that outdated parametric knowledge can be selectively overridden at inference time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper formalizes Parametric Temporal Conflict (PTC) as the issue where LLMs store both outdated and superseding facts in parameters yet standard prompting elicits the outdated answer. It introduces Temporal Attractor Steering (TAS), a three-stage test-time method that detects likely conflicts, identifies a conflict-critical layer, and steers hidden states toward newer-fact representations. An 8,746-record verified benchmark is constructed across five Wikidata relations and evaluated on Qwen-2.5-1.5B/7B, Mistral-7B-v0.3, and Llama-3.1-8B. Single-layer activation patching yields flip rates of 0.72-0.85; end-to-end TAS resolves 29-57% of PTC cases while preserving 85-99% accuracy on non-conflict queries and outperforms a matched ITI baseline on three of four models.

Significance. If the end-to-end results hold, the work demonstrates that outdated parametric knowledge can be selectively overridden at inference time via hidden-state steering without retraining or retrieval. The construction of a verified, multi-relation benchmark provides a concrete, falsifiable testbed for temporal knowledge conflicts and enables direct comparison across model families.

major comments (2)
  1. [Abstract] The central end-to-end TAS claim (29-57% resolution while preserving 85-99% non-conflict accuracy) rests on the automated detection and layer-identification stages operating without ground-truth answers or side effects on unseen PTC instances. The abstract supplies no equation, pseudocode, or algorithmic description of how the conflict-critical layer is identified (e.g., activation-norm scan, contrastive probe, or other heuristic), leaving open whether the procedure correlates with the five-relation benchmark construction or leaks into unrelated factual representations.
  2. [Abstract] Single-layer patching is reported with oracle knowledge of the conflict (0.72-0.85 flip rates), yet the load-bearing claim for TAS requires that the first two stages succeed on held-out PTC cases. Without explicit verification that layer selection generalizes independently of the benchmark construction details, the reported preservation rates on non-conflict queries cannot be assessed for unintended degradation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below, clarifying the presentation of TAS and the evaluation of its stages.

read point-by-point responses

  1. Referee: [Abstract] The central end-to-end TAS claim (29-57% resolution while preserving 85-99% non-conflict accuracy) rests on the automated detection and layer-identification stages operating without ground-truth answers or side effects on unseen PTC instances. The abstract supplies no equation, pseudocode, or algorithmic description of how the conflict-critical layer is identified (e.g., activation-norm scan, contrastive probe, or other heuristic), leaving open whether the procedure correlates with the five-relation benchmark construction or leaks into unrelated factual representations.

    Authors: The abstract is a concise summary. The full description of the three TAS stages, including the test-time procedure for identifying the conflict-critical layer (with equations for activation contrast and layer selection), appears in Section 3 of the manuscript. This procedure uses only the input query and model activations at inference time, without ground-truth answers. We will revise the abstract to include a brief algorithmic outline of the layer-identification step. revision: yes

  2. Referee: [Abstract] Single-layer patching is reported with oracle knowledge of the conflict (0.72-0.85 flip rates), yet the load-bearing claim for TAS requires that the first two stages succeed on held-out PTC cases. Without explicit verification that layer selection generalizes independently of the benchmark construction details, the reported preservation rates on non-conflict queries cannot be assessed for unintended degradation.

    Authors: Oracle single-layer patching establishes an upper bound. End-to-end TAS applies the automated detection and layer-identification stages to held-out PTC instances from the benchmark. Non-conflict accuracy is measured on a disjoint set of queries. The layer-selection heuristic is query-dependent and yields consistent results across four models and five relations, supporting generalization. We acknowledge that further ablations on out-of-distribution queries would strengthen the claim and will add such analysis in revision. revision: partial

Circularity Check

0 steps flagged

No circularity detected; empirical results presented without self-referential reductions or load-bearing self-citations.

full rationale

The manuscript describes an empirical three-stage test-time intervention (TAS) evaluated on a constructed benchmark of 8,746 records, reporting answer-flip rates, resolution percentages, and accuracy preservation as direct experimental outcomes. No equations, parameter-fitting steps, or derivations appear in the provided text that reduce these metrics to inputs by construction. No self-citations are invoked to justify uniqueness or ansatzes, and the method is framed as an independent intervention rather than a renaming or tautological prediction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Ledger populated from abstract claims only; full paper would allow more precise identification of parameters and assumptions.

axioms (1)
  • domain assumption LLMs store both outdated and newer superseding facts simultaneously in their parameters
    Stated as the premise enabling Parametric Temporal Conflict
invented entities (2)
  • Parametric Temporal Conflict (PTC) no independent evidence
    purpose: Formal name for the problem of conflicting temporal facts in model parameters
    New term introduced to frame the issue
  • Temporal Attractor Steering (TAS) no independent evidence
    purpose: Three-stage test-time intervention to resolve PTC
    New method proposed in the work

pith-pipeline@v0.9.1-grok · 5739 in / 1328 out tokens · 29735 ms · 2026-06-26T17:32:43.732086+00:00 · methodology

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

60 extracted references · 3 canonical work pages

  1. [1]

    Ashutosh Bajpai, Aaryan Goyal, Atif Anwer, and Tanmoy Chakraborty. 2024. Temporally consistent factuality probing for large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15864--15881

    2024

  2. [3]

    Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Haotian Wang, Ming Liu, and Bing Qin. 2024. Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1204--1228

    2024

  3. [5]

    Bhuwan Dhingra, Jeremy R Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W Cohen. 2022. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257--273

    2022

  4. [6]

    Bahare Fatemi, Mehran Kazemi, Anton Tsitsulin, Karishma Malkan, Jinyeong Yim, John Palowitch, Sungyong Seo, Jonathan Halcrow, and Bryan Perozzi. 2024. https://arxiv.org/abs/2406.09170 Test of time: A benchmark for evaluating LLM s on temporal reasoning . In The Thirteenth International Conference on Learning Representations. ArXiv:2406.09170

    arXiv 2024

  5. [7]

    Constanza Fierro, Nicolas Garneau, Emanuele Bugliarello, Yova Kementchedjhieva, and Anders S gaard. 2024. Mulan: A study of fact mutability in language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 762--771

    2024

  6. [8]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484--5495

    2021

  7. [10]

    Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models. Advances in Neural Information Processing Systems, 36:17643--17668

    2023

  8. [12]

    Xinyue Kang, Diwei Shi, and Li Chen. 2026. Model whisper: Steering vectors unlock large language models' potential in test-time. In Proceedings of the AAAI Conference on Artificial Intelligence. ArXiv:2512.04748; accepted AAAI 2026

    arXiv 2026

  9. [13]

    Jungo Kasai, Keisuke Sakaguchi, yoichi takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah Smith, Yejin Choi, and Kentaro Inui. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/9941624ef7f867a502732b5154d30cb7-Paper-Datasets_and_Benchmarks.pdf Realtime qa: What s the answer right now? In Advances in Neural Information Pro...

    arXiv 2023

  10. [14]

    Yujin Kim, Jaehong Yoon, Seonghyeon Ye, Sangmin Bae, Namgyu Ho, Sung Ju Hwang, and Se-Young Yun. 2024. Carpe diem: On the evaluation of world knowledge in lifelong language models. In 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

  11. [15]

    Gaotang Li, Yuzhong Chen, and Hanghang Tong. 2025. https://openreview.net/forum?id=0cEZyhHEks Taming knowledge conflicts in language models . In Forty-second International Conference on Machine Learning. ArXiv:2503.10996

    arXiv 2025

  12. [16]

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model. In Advances in Neural Information Processing Systems

    2023

  13. [18]

    Sara Vera Marjanovi \'c , Haeun Yu, Pepa Atanasova, Maria Maistro, Christina Lioma, and Isabelle Augenstein. 2024. Dynamicqa: Tracing internal knowledge conflicts in language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14346--14360

    2024

  14. [19]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022 a . Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359--17372

    2022

  15. [23]

    Qingyu Tan, Hwee Tou Ng, and Lidong Bing. 2024. Towards robust temporal reasoning of large language models via a multi-hop qa dataset and pseudo-instruction tuning. In Findings of the Association for Computational Linguistics: ACL 2024, pages 6272--6286

    2024

  16. [24]

    Md Nayem Uddin, Amir Saeidi, Divij Handa, Agastya Seth, Tran Cao Son, Eduardo Blanco, Steven Corman, and Chitta Baral. 2025. Unseentimeqa: Time-sensitive question-answering beyond llms’ memorization. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1873--1913

    2025

  17. [25]

    Tu Vu, Mohit Iyyer, Xuezhi Wang, Noah Constant, Jerry Wei, Jason Wei, Chris Tar, Yun-Hsuan Sung, Denny Zhou, Quoc Le, and Thang Luong. 2024. Freshllms: Refreshing large language models with search engine augmentation. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13697--13720

    2024

  18. [26]

    Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. 2024. Knowledge conflicts for llms: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8541--8565

    2024

  19. [27]

    Michael Zhang and Eunsol Choi. 2021. Situatedqa: Incorporating extra-linguistic contexts into qa. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7371--7387

    2021

  20. [29]

    Bowen Zhao, Zander Brumbaugh, Yizhong Wang, Hannaneh Hajishirzi, and Noah A Smith. 2024. Set the clock: Temporal alignment of pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 15015--15040

    2024

  21. [30]

    Ruilin Zhao, Feng Zhao, Guandong Xu, Sixiao Zhang, and Hai Jin. 2022. Can language models serve as temporal knowledge bases? In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2024--2037

    2022

  22. [31]

    Xinyu Zhu, Cheng Yang, Bei Chen, Siheng Li, Jian-Guang Lou, and Yujiu Yang. 2023. Question answering as programming for solving time-sensitive questions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12775--12790

    2023

  23. [32]

    Zhiyuan Zhu, Yusheng Liao, Zhe Chen, Yuhao Wang, Yunfeng Guan, Yanfeng Wang, and Yu Wang. 2025. Evolvebench: A comprehensive benchmark for assessing temporal awareness in llms on evolving knowledge. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16173--16188

    2025

  24. [33]

    arXiv preprint arXiv:2108.06314 , year=

    A dataset for answering time-sensitive questions , author=. arXiv preprint arXiv:2108.06314 , year=

    arXiv

  25. [34]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    SituatedQA: Incorporating extra-linguistic contexts into QA , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  26. [35]

    Advances in neural information processing systems , volume=

    Realtime qa: What's the answer right now? , author=. Advances in neural information processing systems , volume=

  27. [36]

    RealTime QA: What s the Answer Right Now? , url =

    Kasai, Jungo and Sakaguchi, Keisuke and takahashi, yoichi and Le Bras, Ronan and Asai, Akari and Yu, Xinyan and Radev, Dragomir and Smith, Noah and Choi, Yejin and Inui, Kentaro , booktitle =. RealTime QA: What s the Answer Right Now? , url =

  28. [37]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Freshllms: Refreshing large language models with search engine augmentation , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  29. [38]

    F resh LLM s: Refreshing Large Language Models with Search Engine Augmentation

    Vu, Tu and Iyyer, Mohit and Wang, Xuezhi and Constant, Noah and Wei, Jerry and Wei, Jason and Tar, Chris and Sung, Yun-Hsuan and Zhou, Denny and Le, Quoc and Luong, Thang. F resh LLM s: Refreshing Large Language Models with Search Engine Augmentation. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.813

    work page doi:10.18653/v1/2024.findings-acl.813 2024

  30. [39]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Set the clock: Temporal alignment of pretrained language models , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  31. [40]

    arXiv preprint arXiv:2409.13338 , year=

    Time awareness in large language models: benchmarking fact recall across time , author=. arXiv preprint arXiv:2409.13338 , year=

    arXiv

  32. [41]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  33. [42]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Question answering as programming for solving time-sensitive questions , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  34. [43]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Towards robust temporal reasoning of large language models via a multi-hop QA dataset and pseudo-instruction tuning , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  35. [44]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Unseentimeqa: Time-sensitive question-answering beyond llms’ memorization , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  36. [45]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Evolvebench: A comprehensive benchmark for assessing temporal awareness in llms on evolving knowledge , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  37. [46]

    Preprint at https://arxiv

    MRAG: A modular retrieval framework for time-sensitive question answering , author=. Preprint at https://arxiv. org/abs/2412.15540 , year=

    arXiv

  38. [47]

    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

  39. [48]

    Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    Can language models serve as temporal knowledge bases? , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    2022

  40. [49]

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

    Carpe diem: On the evaluation of world knowledge in lifelong 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

  41. [50]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    Mulan: A study of fact mutability in language models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers) , pages=

    2024

  42. [51]

    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

  43. [52]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    DYNAMICQA: Tracing internal knowledge conflicts in language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  44. [53]

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

    Temporally consistent factuality probing for large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  45. [54]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Transformer feed-forward layers are key-value memories , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  46. [55]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  47. [56]

    arXiv preprint arXiv:2210.07229 , year=

    Mass-editing memory in a transformer , author=. arXiv preprint arXiv:2210.07229 , year=

    Pith/arXiv arXiv

  48. [57]

    knowledge editing in language models , author=

    Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models , author=. Advances in Neural Information Processing Systems , volume=

  49. [58]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =

    Temporal Alignment of Time Sensitive Facts with Activation Engineering , author =. Findings of the Association for Computational Linguistics: EMNLP 2025 , month = nov, year =. doi:10.18653/v1/2025.findings-emnlp.404 , pages =

    work page doi:10.18653/v1/2025.findings-emnlp.404 2025

  50. [59]

    arXiv preprint arXiv:2302.09664 , year=

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation , author=. arXiv preprint arXiv:2302.09664 , year=

    Pith/arXiv arXiv

  51. [60]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Factual confidence of LLMs: on reliability and robustness of current estimators , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  52. [61]

    Advances in Neural Information Processing Systems , year=

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author=. Advances in Neural Information Processing Systems , year=

  53. [62]

    Forty-second International Conference on Machine Learning , year =

    Taming Knowledge Conflicts in Language Models , author =. Forty-second International Conference on Machine Learning , year =

  54. [63]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Time is Encoded in the Weights of Finetuned Language Models , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2024 , address =. doi:10.18653/v1/2024.acl-long.141 , url =

    work page doi:10.18653/v1/2024.acl-long.141 2024

  55. [64]

    Test of Time: A Benchmark for Evaluating

    Fatemi, Bahare and Kazemi, Mehran and Tsitsulin, Anton and Malkan, Karishma and Yim, Jinyeong and Palowitch, John and Seo, Sungyong and Halcrow, Jonathan and Perozzi, Bryan , booktitle =. Test of Time: A Benchmark for Evaluating. 2024 , url =

    2024

  56. [65]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Model Whisper: Steering Vectors Unlock Large Language Models' Potential in Test-time , author =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  57. [66]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  58. [67]

    arXiv preprint arXiv:2604.10031 , year=

    CoSToM: Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models , author=. arXiv preprint arXiv:2604.10031 , year=

    Pith/arXiv arXiv

  59. [68]

    arXiv preprint arXiv:2603.15892 , year=

    Temporal Fact Conflicts in LLMs: Reproducibility Insights from Unifying DYNAMICQA and MULAN , author=. arXiv preprint arXiv:2603.15892 , year=

    arXiv

  60. [69]

    arXiv preprint arXiv:2601.09445 , year=

    Where Knowledge Collides: A Mechanistic Study of Intra-Memory Knowledge Conflict in Language Models , author=. arXiv preprint arXiv:2601.09445 , year=

    arXiv