Pith.

Computation and Language · arXiv:2610.36915 v1 · submitted 27 Oct 2026 · 41 references

Do We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAG

Parv Khatri, Harshul Jain, Mehak Kapoor, Tanya Sah

The paper says

“When the sources retrieved for a query disagree, the right response depends on why they disagree: a question with several valid answers, a contested topic, stale information and a false source each call for different behaviour. Cattan et al.”

more of the abstract

(2025) showed that telling a generator the conflict type improves its responses by about 24 points, but that predicting the type is hard even for frontier LLMs. We study whether typed decision models, which return probabilities over fixed options in a single forward pass instead of generating text, can do this classification fast and cheaply. Using the CONFLICTS benchmark and the original paper's prompt verbatim, we evaluate all 11 decision models served by one commercial inference gateway, from 7 providers, against Gemini models on identical inputs. Asked directly, the decision models reach 50.0–67.0% accuracy; 10 of the 11 match or exceed Gemini 2.5 Flash, the paper's model (61.3%), at 4–39x lower billed cost (2 were free) and 2.6–6.0x lower latency. The current Gemini 3.8 Flash (70.0%) is significantly more accurate than 8 of them, but sending each decision model's least confident quarter of queries to it gives a cascade statistically tied with it for 8 of the 11, at 25–37% of its cost. End to end, answers generated with the types of 10 of the 11 follow the expected behaviour as often as with the LLM's types. Decomposing the taxonomy into yes/no questions lowers accuracy for every model, and a repair fitted on one model does not transfer. Pre-specified diagnostics show that dates play no measurable role and that sensitivity to source order varies widely between models. Decision models detect a planted false source less reliably than the LLM as hard labels, but their probabilities still separate planted from clean sources. We release code, prompts and all model outputs at https://github.com/Parv17k/decision-models-conflict-typing.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 28 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. A careful, transparent empirical evaluation showing that cheap decision models can replace LLMs for conflict-type classification in RAG, with limitations properly acknowledged. This line is wrong

1 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 41 references say.

1 minor objection

The paper says, at Table 2 footnote

Pith says: minor objection 1 of 1

The note for ScaleDown Classify-1 ('§Its API accepts only a single text field, so it receives the same content as one text (the LLMs' numbered results).') could be placed in the table header for clarity, since it applies to all rows of that model.

What would settle it

Re-run the full classification comparison with a fixed, pinned snapshot of Gemini 2.5 Flash (e.g., via a frozen API version or a locally deployed model) to verify that the 61.3% baseline is stable and that the 10 decision models still match or exceed it; this would address the acknowledged uncertainty from hosted model changes.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 28 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 38 s · 22,392 tokens read, 4,327 written
  3. Editor wrote the verdict above 11 s · 23,138 tokens read, 1,208 written
  4. Referee accept: 1 minor 23 s · 23,226 tokens read, 522 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The central claim holds up: 10 of the 11 decision models match or exceed the original LLM (Gemini 2.5 Flash) in conflict-type accuracy, at 4–39× lower cost and 2.6–6.0× lower latency. The evidence is solid for what it is.

What’s new: this is the first systematic census of decision models on the conflict-typing task, and it goes well beyond a simple accuracy comparison. The paper includes a formulation study (flat vs. tree questions), confidence-gated cascades with cost analysis, pre-specified diagnostics (dates, source order, planted misinformation), and an end-to-end answer-quality test. The methodology is well controlled: same inputs, verbatim prompts, paired comparisons with bootstrap and McNemar tests, a pre-registered protocol with frozen settings, and 92% agreement between the LLM judge and human annotators. All code, prompts, and outputs are released.

The soft spots are real but fully acknowledged by the authors. CONFLICTS has only 406 test queries, giving confidence intervals of about ±5 points, so differences under ~5 points are not significant. Misinformation has only 5 natural instances (4 test), so the planted variant is a proxy. Hosted models change without notice, and the census covers one gateway on one day. The cascade results show statistical ties, not equivalence, as the authors note. The end-to-end gain depends on the judge. None of these are hidden or spun.

Who should read this: anyone deploying or researching cost-efficient classifier layers in RAG pipelines. It’s an engineering contribution, not a scientific paradigm shift, but it is thorough and honest. The paper deserves serious peer review; it will likely be a useful reference for practitioners choosing among decision models. I would cite it as evidence that these models are viable for conflict typing, subject to the benchmark’s limits.

The strongest objection

The paper is a well-executed empirical evaluation with clear methodology, pre-registered analysis, explicit limitations, and released artifacts. The central claim that most decision models match or exceed the original study's LLM under the same inputs is supported by the paired comparison. The cascade analysis is properly controlled, and the diagnostics are pre-specified. The main limitations (small CONFLICTS benchmark with 406 test queries, only 5 natural misinformation instances, confidence intervals ~±4.8 points, hosted model version drift, end-to-end judge bias) are all transparently discussed by the authors. None of these constitute an unacknowledged fatal flaw; they are appropriate caveats to an otherwise sound empirical contribution. The conclusions do not overstate beyond what the data and analysis can support.

The reader identified the small dataset and wide confidence intervals as the weakest assumption, which I agree is the primary limitation. The paper honestly acknowledges this and also lists model instability, label distribution, and judge bias. No hidden weakness or internal inconsistency was found after careful reading. The central claim is appropriately supported by paired comparisons and controlled cascades. Therefore the verdict should remain ACCEPT with no adjustment.

The referee's report

If the results hold, decision models provide a practical, cheap, and fast alternative to LLMs for the critical subtask of conflict typing in search-augmented generation, a setting where latency and cost accumulate per query. The paper's methodology is well-controlled (same inputs, verbatim prompts, paired comparisons, bootstrap and McNemar tests, pre-registered protocol, validated LLM judge with 92% agreement). The evaluation is thorough, covering accuracy, calibration, per-class F1, cascades, formulation variants, and diagnostics. The limitations are discussed openly and do not undermine the central claim. The study is reproducible (code, prompts, and outputs released).

Pith's reading of the paper (a machine's summary, not the authors')

The paper asks whether typed decision models—classifiers that return probabilities in one forward pass—can replace large language models (LLMs) for the task of identifying why retrieved search results conflict. Using the CONFLICTS benchmark, the authors test all 11 decision models served by one commercial gateway against Gemini LLMs. They find that 10 of the 11 match or exceed the original study's best LLM (Gemini 2.5 Flash) on five-class accuracy, at 4–39× lower cost and 2.6–6.0× lower latency. When a decision model's least confident quarter of queries is escalated to a current LLM (Gemini 3.8 Flash), the cascade is statistically tied with that LLM alone at 25–37% of its cost.

What carries the argument. Typed decision models, which accept unstructured state (query and search results) and typed questions (choice, yes/no) and return probabilities over fixed options in a single forward pass without generating text. The paper uses these models with a 'flat' choice question whose definition and examples are taken verbatim from the original LLM prompt. Confidence from the flat question's top probability is then used to gate a cascade: for each decision model, the fraction of queries with lowest confidence is sent to a stronger LLM (Gemini 3.8 Flash), replacing the decision model's prediction.

Core claim. The paper establishes that typed decision models, as a class, can serve as a fast, cheap substitute for LLMs in classifying the type of knowledge conflict among search results for retrieval-augmented generation. On the CONFLICTS test set (406 queries), 10 of the 11 decision models achieve accuracy statistically tied with or better than Gemini 2.5 Flash (61.3%), reaching 50.0–67.0%, while costing $0.04–$0.41 per thousand queries versus $1.66 and responding in 0.26–0.72 seconds versus 1.55 seconds. A confidence-gated cascade that escalates the decision model's most uncertain 25% of queries to Gemini 3.8 Flash (70.0% alone) yields accuracy tied with that LLM for 8 of the 11 models at 25–37% of its cost. Moreover, in an end-to-end test, answers generated using the decision model–predicted types follow the expected behaviour as often as those using the LLM's types.

The premise it rests on. The CONFLICTS benchmark contains only 406 test queries, so differences under about five percentage points are not statistically significant, and the label distribution does not represent real search traffic; the paper's conclusions about ties and small accuracy gaps rest on this limited sample.

If the paper is right (Pith's inferences, not the authors')

Conflict typing, previously requiring an expensive LLM call, can be done by a decision model at a fraction of the cost and latency, making it practical to run on every search-augmented request.

Confidence-gated cascades can recover the accuracy of a current frontier LLM at 25–37% of its cost, providing a deployable cost-quality trade-off.

End-to-end answer quality is preserved when decision model types replace LLM types in the pipeline, meaning cheaper classification does not hurt downstream response quality.

The yes/no decomposition of the taxonomy fails for all models, so the flat choice question is the reliable format for this task.

Decision models detect a planted false source more reliably in their probability scores than in their top label, suggesting that uncertainty-aware thresholds could improve misinformation detection.

The paper implies that hosted decision models could be deployed as a universal 'conflict typing' layer in commercial RAG systems, potentially replacing LLM-based classifiers at scale.

The finding that dates play no measurable role suggests that decision models (and likely LLMs) rely on content cues like temporal language, not explicit dates; this could inform prompt design for temporal queries.

Census results are limited to a single inference gateway and one day's model versions; reproducibility across providers and over time remains an open question, especially as closed-source models shift without notice.

Decision models' weak performance on the few natural 'Misinformation' instances and their better probability separation on planted cases suggests that calibration on in-domain, long-tail examples is critical before deployment.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

When the sources retrieved for a query disagree, the right response depends on why they disagree: a question with several valid answers, a contested topic, stale information and a false source each call for different behaviour. Cattan et al. (2025) showed that telling a generator the conflict type improves its responses by about 24 points, but that predicting the type is hard even for frontier LLMs. We study whether typed decision models, which return probabilities over fixed options in a single forward pass instead of generating text, can do this classification fast and cheaply. Using the CONFLICTS benchmark and the original paper's prompt verbatim, we evaluate all 11 decision models served by one commercial inference gateway, from 7 providers, against Gemini models on identical inputs. Asked directly, the decision models reach 50.0–67.0% accuracy; 10 of the 11 match or exceed Gemini 2.5 Flash, the paper's model (61.3%), at 4–39x lower billed cost (2 were free) and 2.6–6.0x lower latency. The current Gemini 3.8 Flash (70.0%) is significantly more accurate than 8 of them, but sending each decision model's least confident quarter of queries to it gives a cascade statistically tied with it for 8 of the 11, at 25–37% of its cost. End to end, answers generated with the types of 10 of the 11 follow the expected behaviour as often as with the LLM's types. Decomposing the taxonomy into yes/no questions lowers accuracy for every model, and a repair fitted on one model does not transfer. Pre-specified diagnostics show that dates play no measurable role and that sensitivity to source order varies widely between models. Decision models detect a planted false source less reliably than the LLM as hard labels, but their probabilities still separate planted from clean sources. We release code, prompts and all model outputs at https://github.com/Parv17k/decision-models-conflict-typing.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-28

The authors' main claim, as the classifier picked it out 0.55

“Asked directly, the decision models reach 50.0--67.0% accuracy; 10 of the 11 match or exceed Gemini 2.5 Flash, the paper's model (61.3%), at 4--39x lower billed cost (2 were free) and 2.6--6.0x lower latency.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.93 · human participants 0.04 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.99

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
When the sources retrieved for a query disagree, the right response depends on why they disagree: a question with several valid answers, a contested topic, stale information and a false source each call for different behaviour. background 1.00 0.07 0.03 asserted 1.00
Cattan et al. (2025) showed that telling a generator the conflict type improves its responses by about 24 points, but that predicting the type is hard even for frontier LLMs. background 0.98 0.03 0.04 asserted 1.00
We study whether typed decision models, which return probabilities over fixed options in a single forward pass instead of generating text, can do this classification fast and cheaply. contribution 0.81 0.38 0.04 unclear 0.40
Using the CONFLICTS benchmark and the original paper's prompt verbatim, we evaluate all 11 decision models served by one commercial inference gateway, from 7 providers, against Gemini models on identical inputs. method 0.99 0.95 0.03 asserted 0.93
Asked directly, the decision models reach 50.0--67.0% accuracy; 10 of the 11 match or exceed Gemini 2.5 Flash, the paper's model (61.3%), at 4--39x lower billed cost (2 were free) and 2.6--6.0x lower latency. result 1.00 0.95 0.32 asserted 1.00
The current Gemini 3.8 Flash (70.0%) is significantly more accurate than 8 of them, but sending each decision model's least confident quarter of queries to it gives a cascade statistically tied with it for 8 of the 11, at 25--37% of its cost. result 1.00 0.94 0.20 asserted 1.00
End to end, answers generated with the types of 10 of the 11 follow the expected behaviour as often as with the LLM's types. result 0.99 0.95 0.33 asserted 1.00
Decomposing the taxonomy into yes/no questions lowers accuracy for every model, and a repair fitted on one model does not transfer. result 0.98 0.93 0.06 asserted 1.00
Pre-specified diagnostics show that dates play no measurable role and that sensitivity to source order varies widely between models. result 1.00 0.93 0.03 asserted 1.00
Decision models detect a planted false source less reliably than the LLM as hard labels, but their probabilities still separate planted from clean sources. result 0.99 0.93 0.06 asserted 1.00
We release code, prompts and all model outputs at https://github.com/Parv17k/decision-models-conflict-typing. contribution 0.96 0.86 0.02 asserted 0.94

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-28: 1 request, 8,585 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (2)
Figure 1 of the paper
Figure 1. Overview. A decision model types the con
Figure 2 of the paper
Figure 2. Accuracy against cost (log scale) on the test split. Each decision model alone with the flat question (blue)

Figures are the authors', reproduced from the paper.

References (41)
  1. [1] Baseten . 2026. The top ranked decision models in the world: mammoth-1 and elephant-1 . https://www.baseten.co/blog/the-best-decision-models-in-the-world-mammoth-1-and-elephant-1/ on Pith
  2. [2] Arie Cattan, Alon Jacovi, Ori Ram, Jonathan Herzig, Roee Aharoni, Sasha Goldshtein, Eran Ofek, Idan Szpektor, and Avi Caciularu. 2025. DRAGged into conflicts: Detecting and addressing conflicting sources in search-augmented LLMs . arXiv preprint arXiv:2506.08500 on PithPith's reviewarXiv
  3. [3] Lingjiao Chen, Matei Zaharia, and James Zou. 2023. FrugalGPT : How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176 on PithPith's reviewarXiv
  4. [4] Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37--46 on Pith
  5. [5] Gheorghe Comanici et al. 2025. Gemini 2.5 : Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 on PithPith's reviewarXiv
  6. [6] Tobias Deu er, Lorenz Sparrenberg, and Rafet Sifa. 2026. Evaluating and benchmarking the System One model Jev . arXiv preprint arXiv:2609.37647 on PithPith's reviewarXiv
  7. [7] Feiyu Duan, Jiayu Lin, Jia Wang, Jun Xiang, Jialiang Wu, Xinnong Zhang, Hanqi Yan, Siyuan Wang, and Zhongyu Wei. 2026. General decision models: Benchmarking and insights beyond Jev . arXiv preprint arXiv:2610.03935 on PithPith's reviewarXiv
  8. [8] Bradley Efron and Robert J. Tibshirani. 1994. An Introduction to the Bootstrap. Chapman & Hall/CRC on Pith
  9. [9] Yonatan Geifman and Ran El-Yaniv. 2017. Selective classification for deep neural networks. In Advances in Neural Information Processing Systems on Pith
  10. [10] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1321--1330 on Pith
  11. [11] Yufang Hou, Alessandra Pascale, Javier Carnerero-Cano, Tigran Tchrakian, Radu Marinescu, Elizabeth Daly, Inkit Padhi, and Prasanna Sattigeri. 2024. WikiContradict : A benchmark for evaluating LLMs on real-world knowledge conflicts from Wikipedia . In Advances in Neural Informa... on Pith
  12. [12] Jiaju Huang, Hao Yang, Xinyu Ma, Xinglong Liang, Kunyan Cai, Junqiang Ma, Shaobin Chen, Yue Sun, and Tao Tan. 2026. Can Jev judge radiology reports? E valuating a System One model for clinical factuality. arXiv preprint arXiv:2609.27607 on PithPith's reviewarXiv
  13. [13] Keyi Li, Yihao He, and Quanyi Li. 2026 a . Beyond calibration: Do a typed-decision model's probabilities obey the probability axioms? arXiv preprint arXiv:2609.33209 on PithPith's reviewarXiv
  14. [14] Xing Li, Qingcheng Chang, Jinzhong Ning, Changfeng Xu, Shenlong Zhang, Yijia Zhang, Ling Luo, and Hongfei Lin. 2026 b . Specialized decision models vs. general-purpose LLMs : Benchmarking Jev across knowledge, reasoning, and multilingual tasks. arXiv preprint arXiv:2610.11978 on PithPith's reviewarXiv
  15. [15] Yinheng Li and Justin Wagle. 2026. LLM -as- Jev : LLMs are already Jev -style decision models -- when and how to fine-tune them. arXiv preprint arXiv:2610.02076 on PithPith's reviewarXiv
  16. [16] Yubo Li, Yidi Miao, Ramayya Krishnan, and Rema Padman. 2026 c . JEV -as-a-judge: Accept when confident, escalate when unsure. arXiv preprint arXiv:2609.26550 on PithPith's reviewarXiv
  17. [17] Liquid AI . 2026 a . d1-ultra : Introducing the world's best decision model. https://www.liquid.ai/blog/d1-ultra on Pith
  18. [18] Liquid AI . 2026 b . Introducing d1 : The most capable decision model, now with vision. https://www.liquid.ai/blog/d1-decision-model on Pith
  19. [19] Siyi Liu, Qiang Ning, Kishaloy Halder, Wei Xiao, Zheng Qi, Phu Mon Htut, Yi Zhang, Neha Anna John, Bonan Min, Yassine Benajiba, and Dan Roth. 2024. Open domain question answering with conflicting contexts. arXiv preprint arXiv:2410.12311 on PithPith's reviewarXiv
  20. [20] Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. 2021. Entity-based knowledge conflicts in question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7052--7063 on Pith
show 21 more
  1. [21] Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2):153--157 on Pith
  2. [22] Microsoft . 2026. Microsoft-Decision-1 . Microsoft Foundry Labs, https://labs.ai.azure.com/innovations/microsoft-decision-1/ on Pith
  3. [23] Shubham Mishra, Samyek Jain, Gorang Mehrishi, Shiv Tiwari, Harsh Sharma, Pratik Narang, and Dhruv Kumar. 2025. From facts to conclusions: Integrating deductive reasoning in retrieval-augmented LLMs . arXiv preprint arXiv:2512.16795 on PithPith's reviewarXiv
  4. [24] Gowthamkumar Nandakishore. 2026. Laya as a typed probabilistic assessor: An independent reproduction and a preregistered study of calibration and selective escalation. arXiv preprint arXiv:2609.33843 on PithPith's reviewarXiv
  5. [25] OpenAI . 2026. Decisions API guide. https://developers.openai.com/api/docs/guides/decisions on Pith
  6. [26] Arjun Panickssery, Samuel R. Bowman, and Shi Feng. 2024. LLM evaluators recognize and favor their own generations. In Advances in Neural Information Processing Systems on Pith
  7. [27] Perplexity AI . 2026 a . pplx-decider-flash-v1.2 model card. https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2 on Pith
  8. [28] Perplexity AI . 2026 b . pplx-decider-v1.2 model card. https://huggingface.co/perplexity-ai/pplx-decider-v1.2 on Pith
  9. [29] Delip Rao and Chris Callison-Burch. 2026. Decision models vs. LLMs as rubric judges: Cheaper, faster, and wrong in the same places. arXiv preprint arXiv:2609.29769 on PithPith's reviewarXiv
  10. [30] ScaleDown . 2026. Classify overview. ScaleDown API reference, https://docs.scaledown.ai/api-reference/classify-overview on Pith
  11. [31] TypeSafe AI . 2026 a . Introducing System One models and Jev . https://typesafe.ai/blog/introducing-system-one-models-and-jev on Pith
  12. [32] TypeSafe AI . 2026 b . Jev 1.13 model jaggedness. https://docs.typesafe.ai/model-jaggedness/jev-1.13 on Pith
  13. [33] 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 Linguisti... on Pith
  14. [34] Alexander Wan, Eric Wallace, and Dan Klein. 2024. What evidence do language models find convincing? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7468--7484 on Pith
  15. [35] Chenyu Wang, Yueyuan Li, Yingmin Liu, and Yang Shu. 2026. ConflictRAG : Detecting and resolving knowledge conflicts in retrieval augmented generation. arXiv preprint arXiv:2605.17301 on PithPith's reviewarXiv
  16. [36] Han Wang, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. 2025. Retrieval-augmented generation with conflicting evidence. In Conference on Language Modeling (COLM) on Pith
  17. [37] Yuchuan Weng, Chi Liu, Xin Chen, and Srinivasan Manoharan. 2026. Our own Jev : Typed decisions in one forward pass, from an in-house judge to open LoRA models. arXiv preprint arXiv:2610.29052 on PithPith's reviewarXiv
  18. [38] 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 on Pith
  19. [39] Zixiang Xu, Zirui Song, Chiyu Zhang, Xiuying Chen, Xi Liu, Xiyang Hu, and Yue Zhao. 2026. JevOut : Natural context can flip decision models. arXiv preprint arXiv:2609.30243 on PithPith's reviewarXiv
  20. [40] Michael J.Q. 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 on Pith
  21. [41] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM -as-a-judge with MT-Bench and Chatbot Arena . In Advances in Neural Information... on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment