Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

This paper claims that a contextual bandit choosing rewrite strategies from a 17-feature linguistic fingerprint reduces LLM hallucination, beating both no-rewrite and static prompting baselines.

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 ↗

A contextual bandit that chooses among five query-rewrite strategies, conditioned on 17 linguistic features, reduces LLM hallucination on QA benchmarks and beats static prompting and no-rewrite baselines.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The core idea—contextual bandits over linguistic features for query rewriting—is fresh and the empirical setup is mostly careful, but the headline 87.5% win rate is measured against a proxy whose dominant term is a same-family LLM judge validated on only 100 human labels, so treat the claims as conditional. the 5 major comments →

arxiv 2508.16697 v1 pith:GXDKXP7X submitted 2025-08-22 cs.CL cs.AIcs.LG

QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting

Nicole Cho , William Watson , Alec Koppel , Sumitra Ganesh , Manuela Veloso This is my paper
classification cs.CL cs.AIcs.LG
keywords hallucination mitigationquery rewritingcontextual banditsThompson samplinglinguistic featuresLLM-as-a-judgemulti-armed banditsquestion answering
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.

The reading

Hallucinations in large language models are usually addressed after the model has already produced a wrong answer. QueryBandits instead intervenes before generation: it treats the choice of a query-rewrite strategy as an online bandit problem. Each incoming query is encoded as a 17-dimensional binary linguistic feature vector, and a contextual bandit selects one of five rewrite strategies (paraphrase, simplify, disambiguate, expand, or clarify terms) to maximize a reward proxy meant to capture hallucination propensity. Across 13 QA benchmarks and 1,050 lexically perturbed queries per dataset, the best variant—Thompson Sampling with the full feature vector—wins against the no-rewrite baseline on 87.5% of trials and substantially outperforms static "paraphrase" and "expand" prompts. The paper also finds that some static rewrite strategies do worse than doing nothing, and that no single rewrite strategy is best for all queries; the per-arm learned feature weights reveal which linguistic cues favor which rewrite. The significance is that hallucination mitigation can be steered through forward-pass query shaping alone, without retraining or gradient updates.

Core claim

The central empirical claim is that contextual query rewriting reduces hallucination as measured by the paper's reward model. On 13 QA benchmarks, with 1,050 semantically invariant but lexically perturbed queries per dataset, QueryBandits-Thompson Sampling achieves an 87.5% win rate against the no-rewrite baseline and an exploration-adjusted reward of 819.04, compared with 729.20 for no rewriting. It outperforms static paraphrase by 42.6 percentage points and static expand by 60.3 percentage points. Ablating the 17-dimensional feature input drops Thompson Sampling to an 81.7% win rate and 754.66 reward, showing the features carry signal. Per-arm regression weights show that no single rewrite

What carries the argument

The machinery is a contextual multi-armed bandit with a 17-dimensional binary linguistic feature vector as context, five rewrite arms (paraphrase, simplify, disambiguate, expand, clarify terms), and a reward rt = 0.6·s_llm + 0.3·s_fuzz + 0.1·s_bleu, where s_llm is a binary LLM-judge consistency score, s_fuzz is fuzzy token-set similarity, and s_bleu is BLEU-1 unigram overlap. Each arm has a linear model mapping features to expected reward; Thompson sampling draws a posterior parameter and selects the arm with the highest sampled score. This lets the algorithm allocate exploration where feature uncertainty is high and exploitation where features reliably predict reward.

Load-bearing premise

The whole result rests on the assumption that the reward proxy—0.6 times an LLM-judge consistency score, 0.3 times fuzzy string similarity, and 0.1 times BLEU-1—is a faithful measure of hallucination, even though the judge comes from the same model family that produces the answers and the proxy was validated on only 100 human-labeled pairs.

What would settle it

Run the same QueryBandits loop on the same 1,050 perturbed queries per dataset, but score the selected rewrites with a held-out set of human factual-accuracy labels or with a judge from a different model family than the generator, instead of the paper's reward proxy. If the 87.5% win rate does not survive, or if the arms the bandit favors no longer correspond to higher factual accuracy, the reward model, not the rewriting, was driving the reported gains.

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

If this is right

  • If the central claim holds, hallucination mitigation can be done pre-generation by rewriting queries, without retraining, gradient updates, or post-hoc output filtering.
  • Static one-size-fits-all rewrites are not merely suboptimal—some, like blanket expand or clarify-terms prompts, can accrue more regret than doing nothing, so deployers should treat them as risky interventions.
  • Per-query feature vectors make rewrite policy interpretable: learned per-arm weights identify which linguistic properties, such as domain specialization, pragmatic cues, or superlatives, push a query toward one rewrite arm or another.
  • Benchmark evaluations of rewriting should use lexically perturbed but semantically invariant queries; on raw benchmark phrasings the bandit converges to no-rewriting, indicating benchmark memorization can mask the effect.
  • Contextual bandits with entropy-adjusted rewards offer a workflow for latency-sensitive or resource-constrained LLM interfaces: adapt rewrite choices online using only forward passes.

Where Pith is reading between the lines

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

  • Editorial extension: because the reward's dominant term is an LLM judge from the same model family as the generator, some of the measured win rate may reflect the model rewarding its own preferred phrasing rather than objective factuality; a human-labeled or cross-model judge evaluation would separate these.
  • The per-arm feature weights are associational, not causal; one could test them by intervening on a single feature, such as adding or removing a superlative while holding the rest of the query fixed, and observing arm-reward changes.
  • The 17 features are treated as independent, but the paper itself flags that feature interactions are missing; pairwise or higher-order feature terms are a natural extension and could change which arm is optimal.
  • A policy trained on one LLM's judge and generator may transfer imperfectly to another; cross-model replication of the QueryBandits loop is a cheap test of whether the feature-arm associations generalize.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 5 minor

Summary. The paper proposes QueryBandits, a contextual multi-armed bandit framework that selects one of five query-rewrite strategies (paraphrase, simplify, disambiguate, expand, clarify terms) on the basis of a 17-dimensional binary linguistic feature vector, with the goal of reducing LLM hallucination. The reward is a convex combination of an LLM-as-judge consistency score, a fuzzy string similarity score, and BLEU-1 (Eq. 1), with weights selected on a 100-pair human-labeled validation set. Experiments are run on 13 QA datasets with roughly 1,050 lexically perturbed queries per dataset. The paper reports that Thompson Sampling with context achieves an 87.5% win rate over a no-rewrite baseline, outperforms static prompting, and that ablating the feature vector degrades performance. The authors also report per-arm regression weights that suggest no single rewrite strategy is optimal for all queries.

Significance. If the empirical claims held, the paper would make a useful contribution: it frames query rewriting as an online decision problem, provides an interpretable feature-arm analysis, and reports the interesting observation that some static rewrite strategies perform worse than no rewriting. The pipeline and algorithm descriptions are clear, and the authors are explicit about some limitations. However, the central claim that QueryBandits reduce hallucination is currently supported only through a proxy reward whose external validation is thin, and the evaluation is restricted to a specially selected set of perturbed queries. The headline numbers therefore are not yet evidence of hallucination reduction on natural inputs. The contribution is potentially interesting but needs substantially stronger validation.

major comments (5)
  1. [Reward Model (Eq. 1), Appendix A.1] The reward proxy is used both to define hallucination and to train/evaluate the bandit. The dominant term s_llm is a binary judgment by a GPT-4o-based assessor, while the model being evaluated is gpt-4o-2024-08-06; the paper itself acknowledges in A.1 that the reward model may reflect LLM bias. The only external validation is a held-out set of 100 human-labeled pairs; no per-dataset breakdown or exact AUC is given. Since every win rate and regret number is computed from this rt, the 87.5% win rate measures optimization of the proxy, not hallucination reduction. A larger, independent human-labeled evaluation on the actual test outputs, or a factual-correctness measure not derived from the same model family, is required to support the central claim.
  2. [Section 4, Dataset and Query Construction] The evaluation is conditional on an adversarially selected subset. Queries are kept only if the original query is answered correctly and one to three of its five semantically invariant perturbations produce incorrect answers; the incoming query is then a perturbed version. Moreover, Figure 7 shows that on original benchmark queries the contextual bandit converges to no-rewrite, which the authors attribute to memorization. Thus the 87.5% win rate is a win rate on deliberately unstable perturbations, not on natural queries. The paper should report performance on unselected or random query distributions, or clearly restrict the claim to the perturbation-selected setting.
  3. [Evaluation Metrics and Table 3] No confidence intervals, standard errors, or significance tests are reported for any metric. The differences among the top three contextual bandits are 819.04 vs. 818.79 vs. 818.60, i.e., differences of 0.25 and 0.19 reward units; without variance estimates the ordering and the claim that Thompson Sampling is best are unsupported. In addition, the WinRate definition states N=100 test queries, whereas the experiments use approximately 1,050 queries per dataset; it is unclear what N is for the aggregated 87.5% figure. Per-dataset results and uncertainty quantification are needed.
  4. [Title and Section 3, Choice of Algorithms; Appendix A.2] The title claims 'no-regret rewriting,' but no regret bound is stated or proved for the contextual algorithms. The text says regret bounds are deferred to Appendix A.2, but Appendix A.2 contains only update equations and algorithm summaries. Furthermore, the exploration-adjusted reward Radj = Σ(rt + λ Ht) with λ=0.1 adds an entropy bonus that favors exploratory policies; comparing this quantity against a zero-entropy no-rewrite baseline conflates exploration with correctness. The authors should report raw average reward separately from the exploration-adjusted metric, and either supply regret bounds or soften the 'no-regret' terminology.
  5. [Section 4, Experimental Configuration] There is an arithmetic inconsistency in the stated scale: 'Total Pulls = M × D × |QD| = 253,440, with |QD| ≈1050, M = 15, and D = 16.' Table 2 lists 13 datasets, not 16, and 13 × 15 × 1050 = 204,750. The inconsistency affects the reported scale of the evaluation and should be corrected.
minor comments (5)
  1. [Page 2, after the abstract] There is a large unformatted text block beginning 'Can you determine the number of integers between 6 and 74/5?' that appears to be a leaked model output. It should be removed; the intended running example is already described in Figure 1.
  2. [Section 3, Reward Model] The definition says s_llm ∈ {0,1} is a binary consistency judgment, but the later ROC-AUC analysis of rt treats the reward as continuous. Please clarify whether the LLM judge is binary or score-valued.
  3. [Section 3, Evaluation Metrics] The WinRate formula uses N=100 test queries, while the experiments report |QD| ≈ 1050 queries per dataset. This discrepancy should be resolved and the exact N for the aggregated win rate stated.
  4. [Appendix A.2] The phrase 'for full scoring, update equations and regret bounds' promises regret bounds, but none appear. Either add the bounds (even as known results for LinUCB/Thompson sampling) or remove the promise.
  5. [Figure 2a] The Pareto-frontier plot is described in the text, but no exact ROC-AUC values are reported. Please include the actual AUC numbers for the chosen weights and for comparison points on the frontier.

Circularity Check

1 steps flagged

Headline win rate is measured on a reward proxy that also defines hallucination, making the central claim partially definitional.

specific steps
  1. self definitional [Section 1 (Reward Model), Eq. (1); Section 3 (Metric 3); Section 4 (Hypothesis 1)]
    ""we define hallucination in terms of a reward model rt = α · sllm + β · sfuzz + γ · sbleu where hallucinatory responses are those associated with low rewards" (Section 1); "Win Rate vs. Baseline ... compute the fraction of trials for which a policy’s reward rpolicy_t strictly exceeds the no-rewrite baseline’s reward rbase_t" (Section 3); "Our top contextual learner—Thompson Sampling with the 17-dimensional feature vector—achieves an 87.5% win rate ... signifying that contextual query rewriting can reduce hallucination" (Section 4)."

    Hallucination is defined as low reward via Eq. (1). The bandit's objective is to maximize that same reward, and the headline win rate is the fraction of pulls where the policy's Eq.-1 reward exceeds the no-rewrite baseline's Eq.-1 reward. Therefore 'rewriting can reduce hallucination' is, by construction, equivalent to 'rewriting increases the proxy reward.' The paper does validate the proxy against 100 human-labeled pairs (ROC-AUC, Fig. 2a), which provides an external anchor, but the reported 87.5% win rate and all regret/reward numbers are computed on the proxy itself, not on human labels. The later assertion 'our reward directly measures output correctness' (Hypothesis 3) repeats the same definitional identification. The central empirical claim thus reduces to reward optimization, a par

full rationale

The main circularity is definitional: the paper defines hallucination as low reward in Eq. (1), then trains the bandit to maximize that reward and reports win rates computed as reward comparisons. This makes the headline 'hallucination reduction' claim partially true by construction. However, the reward proxy is not purely self-referential: the weights (α, β, γ) = (0.6, 0.3, 0.1) are selected inside a 1% Pareto frontier on a held-out set of 100 human-labeled pairs, and the proxy's ROC-AUC against those labels is reported (Fig. 2a). This external anchor prevents a score of 8 or 10. The paper's own limitation statement in A.1 acknowledges that 'our reward model leverages LLM-as-judge, which may reflect the LLM's bias,' further confirming that the proxy, not ground truth, drives the reported metrics. Self-citations appear (e.g., [16], [112]) but are used as background motivation, not as load-bearing justification for the central claim; no uniqueness theorem or ansatz is imported from prior work. No other pattern of circularity is present. Because the central 'prediction' of hallucination reduction reduces by construction to reward optimization, with only a small external validation anchor, a score of 6 is appropriate: partial circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claim rests on a fixed reward proxy with fitted weights, a linear per-arm model, and LLM-based feature extraction. The proxy is validated on a small human set, but the evaluation metric and the training objective are the same, which limits the independence of the evidence. The hyperparameters are not disclosed, and the query selection procedure is a deliberate filtering step that shapes all reported results.

free parameters (3)
  • Reward weights (α, β, γ) = (0.6, 0.3, 0.1)
    Chosen via grid search to lie inside the 1% Pareto frontier of ROC-AUC on a held-out human-labeled set of 100 pairs; the weights define the hallucination proxy and therefore the bandit objective.
  • Entropy bonus λ in exploration-adjusted reward = 0.1
    Fixed by the authors to weight exploration; this metric is used for headline comparisons but is not the primary win-rate metric.
  • Bandit hyperparameters (learning rates, exploration coefficients, regularization constants) = not reported
    Tuned via grid search on a held-out validation set but values are not listed, hampering reproducibility.
axioms (4)
  • domain assumption Hallucination is a per-query phenomenon, so a bandit (rather than a full MDP) is the appropriate abstraction (Remark 1).
    The paper argues that attention over a fixed context makes each query independent; this is a modeling assumption.
  • domain assumption The 17 binary linguistic features can be reliably extracted from a query by an LLM.
    Feature extraction is delegated to GPT-4o; no inter-annotator agreement or external validation is provided.
  • ad hoc to paper The reward is linear in the per-arm feature vector, x^T θ_k, for each arm.
    All contextual bandit variants fit a linear model per arm (e.g., LinUCB, Thompson sampling with linear posterior).
  • domain assumption Lexically perturbed queries preserve semantic meaning.
    The query construction selects perturbations that are semantically invariant, measured by n-gram and LLM-judge metrics.
invented entities (2)
  • 17-dimensional binary linguistic feature vector no independent evidence
    purpose: Conditioning context for the bandit to choose a rewrite arm
    The features are hypothesized to correlate with hallucination risk and are extracted by an LLM; they are not externally validated against a ground-truth feature set.
  • Hallucination reward proxy rt = α sllm + β sfuzz + γ sbleu no independent evidence
    purpose: Training and evaluation signal for the bandit
    The proxy is a composite of LLM-judge, fuzzy-match, and BLEU scores; it is validated only on 100 human-labeled pairs, so it has limited external grounding.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting." pith.science (2026). https://pith.science/paper/GXDKXP7X

@misc{pith2026250816697,
  author       = {Pith},
  title        = {Pith review of: QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GXDKXP7X}},
  note         = {Machine review of arXiv:2508.16697}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Advanced reasoning capabilities in Large Language Models (LLMs) have caused higher hallucination prevalence; yet most mitigation work focuses on after-the-fact filtering rather than shaping the queries that trigger them. We introduce QueryBandits, a bandit framework that designs rewrite strategies to maximize a reward model, that encapsulates hallucination propensity based upon the sensitivities of 17 linguistic features of the input query-and therefore, proactively steer LLMs away from generating hallucinations. Across 13 diverse QA benchmarks and 1,050 lexically perturbed queries per dataset, our top contextual QueryBandit (Thompson Sampling) achieves an 87.5% win rate over a no-rewrite baseline and also outperforms zero-shot static prompting ("paraphrase" or "expand") by 42.6% and 60.3% respectively. Therefore, we empirically substantiate the effectiveness of QueryBandits in mitigating hallucination via the intervention that takes the form of a query rewrite. Interestingly, certain static prompting strategies, which constitute a considerable number of current query rewriting literature, have a higher cumulative regret than the no-rewrite baseline, signifying that static rewrites can worsen hallucination. Moreover, we discover that the converged per-arm regression feature weight vectors substantiate that there is no single rewrite strategy optimal for all queries. In this context, guided rewriting via exploiting semantic features with QueryBandits can induce significant shifts in output behavior through forward-pass mechanisms, bypassing the need for retraining or gradient-based adaptation.

Figures

Figures reproduced from arXiv: 2508.16697 by Alec Koppel, Manuela Veloso, Nicole Cho, Sumitra Ganesh, William Watson.

Figure 1
Figure 1. Figure 1: QueryBandits and Its Success in Mitigating Hallucination. The original query xt induces a hallucinatory output: the LLM calculates 8 integers between 6 and 74/5. QueryBandits, by leveraging the feature vector, selects the EXPAND rewrite strategy. The rewritten query x ′ t generates an accurate output of 9 integers. Noticeably, the feature vectors are different in the rewrite x ′ t - subordination (more com… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Our chosen (α, β, γ) lies deep in the 1% optimal frontier. (b) Breakdown of per-dataset arm performance: different datasets consistently favor different rewrite strategies ▶ Fuzzy-Match Sensitivity (β): Small increases in β rapidly exit the Pareto region, showing that the fuzzy-match term must be tuned carefully to avoid degrading overall accuracy. ▶ BLEU-Only Pitfall (γ): As γ increases, ROC–AUC stead… view at source ↗
Figure 3
Figure 3. Figure 3: Cumulative reward averaged across all datasets over the no-rewrite baseline for each algorithm (sorted by final performance), highlighting the superior gains achieved by contextual bandits compared to non-contextual learners and static prompt-based rewrites [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Contextual Feature Contribution Strength. These are the averaged θ weights (di￾rect contributions) of each feature to the expected reward under each arm. Positive weights indicate features that boost that arm’s reward; negative weights indicate features that penalize it. Hypothesis 4: Is there an association between query features and reward? Each rewrite arm seems to exhibit different sensitivities toward… view at source ↗
Figure 6
Figure 6. Figure 6: For Non-Contextual bandits, almost every dataset is dominated by a single arm with the highest global reward (typically 40%-60% of the trials). The remaining 40-60% is split among the other four arms as noise, the non-contextual policy has no way to "know" when within a dataset a different arm might do better. In contrast, Contextual bandits show a more even mix: the top arm is only ∼25-30%, with two or th… view at source ↗
Figure 7
Figure 7. Figure 7: Impact of the No-Rewrite Arm. Note that these experiments are conducted on the original query "as-is" in the benchmark dataset, with no perturbations. Upon enabling the NO REWRITE option, our contextual bandit rapidly converges to this arm, which then achieves the highest reward on several datasets. We attribute this behavior to the LLM’s tendency to memorize benchmark questions. Given a query feature vect… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of Feature Variance between (a) our contextual bandits and (b) its non [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of Inter-Arm Context Distances (Symmetric KL) between (a) our contextual [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of Raw feature-level regression coefficients between (a) our contextual [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of Min-Max Normalized feature-level regression coefficients between (a) our [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Reward Uplift by Contextual Feature and Strategy. Feature Uplift measures how much the presence of a binary feature changes the expected reward for a given rewrite arm, formally ∆(fi , a) = E[rt | arm = a, fi = 1] − E[rt | arm = a, fi = 0]. (a) Under the contextual bandit, the strongest positive uplifts come from Answerability (≈ +17 uniformly) and Grounding (+15–18), while Ambiguity (≈ –15 to –18) and Su… view at source ↗
Figure 13
Figure 13. Figure 13: Pairwise Normalized Coefficient Differences for Contextual Bandits. Each cell shows the min–max–normalized difference in regression weight for a given linguistic feature (rows) between two rewrite arms (columns), e.g. “Paraphrase vs Disambiguate,” “Simplify vs Expand,” etc. Cells labeled “Win” (blue) indicate the feature favors the first arm in the matchup, while “Loss” (red) indicates it favors the secon… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Diagnosing and Repairing Factual Errors in RAG under Budget Constraints

    cs.AI 2026-06 unverdicted novelty 5.0

    D2R-RAG diagnoses RAG failures via observable signals and applies adaptive repairs under latency and VRAM limits, reporting improved reliability and accuracy-efficiency trade-offs on FEVER and HotpotQA.

Reference graph

Works this paper leans on

130 extracted references · 21 canonical work pages · cited by 1 Pith paper · 11 internal anchors

  1. [1]

    Evaluating correctness and faithfulness of instruction-following models for question answering, 2024

    Vaibhav Adlakha, Parishad BehnamGhader, Xing Han Lu, Nicholas Meade, and Siva Reddy. Evaluating correctness and faithfulness of instruction-following models for question answering, 2024. URL https: //arxiv.org/abs/2307.16877

    Pith/arXiv arXiv 2024

  2. [2]

    How does the pre-training objective affect what large language models learn about linguistic properties?

    Ahmed Alajrami and Nikolaos Aletras. How does the pre-training objective affect what large language models learn about linguistic properties?, 2022. URL https://arxiv.org/abs/2203.10415

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  3. [3]

    Emmanuel Ameisen, Jack Lindsey, Adam Pearce, Wes Gurnee, Nicholas L. Turner, Brian Chen, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, Andy Jones, Andrew Persic, Zhenyi Qi, T. Ben Thompson, Sam Zimmerman, Kelley Rivo...

    2025

  4. [5]

    PolyLM: Learning about Polysemy through Language Modeling

    Alan Ansell, Felipe Bravo-Marquez, and Bernhard Pfahringer. Polylm: Learning about polysemy through language modeling, 2021. URL https://arxiv.org/abs/2101.10448

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [6]

    Dhananjay Ashok and Zachary C. Lipton. Promptner: Prompting for named entity recognition, 2023. URL https://arxiv.org/abs/2305.15444

    Pith/arXiv arXiv 2023

  6. [7]

    Schapire

    Peter Auer, Nicolò Cesa-Bianchi, Yoav Freund, and Robert E. Schapire. The nonstochastic multiarmed bandit problem. SIAM Journal on Computing, 32(1):48–77, 2002. doi: 10.1137/S0097539701398375. URL https://doi.org/10.1137/S0097539701398375

    doi:10.1137/s0097539701398375 2002

  7. [8]

    rapidfuzz/rapidfuzz: Release 3.8.1, April 2024

    Max Bachmann. rapidfuzz/rapidfuzz: Release 3.8.1, April 2024. URL https://doi.org/10.5281/ zenodo.10938887

    2024

  8. [9]

    Harnessing gpt-3.5-turbo for rhetorical role prediction in legal cases, 2023

    Anas Belfathi, Nicolas Hernandez, and Laura Monceaux. Harnessing gpt-3.5-turbo for rhetorical role prediction in legal cases, 2023. URL https://arxiv.org/abs/2310.17413

    Pith/arXiv arXiv 2023

  9. [10]

    Piqa: Reasoning about physical commonsense in natural language

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language. In Thirty-Fourth AAAI Conference on Artificial Intelligence, 2020

    2020

  10. [11]

    Prompting language models for linguistic structure,

    Terra Blevins, Hila Gonen, and Luke Zettlemoyer. Prompting language models for linguistic structure,

  11. [12]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwi...

    Pith/arXiv arXiv 2020

  12. [13]

    Re-evaluating the role of Bleu in machine translation research

    Chris Callison-Burch, Miles Osborne, and Philipp Koehn. Re-evaluating the role of Bleu in machine translation research. In Diana McCarthy and Shuly Wintner, editors, 11th Conference of the European Chapter of the Association for Computational Linguistics , pages 249–256, Trento, Italy, April 2006. Association for Computational Linguistics. URL https://acl...

    2006

  13. [14]

    PreCo: A large-scale dataset in preschool vocabulary for coreference resolution

    Hong Chen, Zhenhua Fan, Hao Lu, Alan Yuille, and Shu Rong. PreCo: A large-scale dataset in preschool vocabulary for coreference resolution. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii, editors, Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 172–181, Brussels, Belgium, October-Novembe...

    work page doi:10.18653/v1/d18-1016 2018

  14. [15]

    PRompt optimization in multi-step tasks (PROMST): Integrating human feedback and heuristic-based sampling

    Yongchao Chen, Jacob Arkin, Yilun Hao, Yang Zhang, Nicholas Roy, and Chuchu Fan. PRompt optimization in multi-step tasks (PROMST): Integrating human feedback and heuristic-based sampling. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3859–3920, Mi...

    2024

  15. [16]

    Multiq&a: An analysis in measuring robustness via automated crowdsourcing of question perturbations and answers, 2025

    Nicole Cho and William Watson. Multiq&a: An analysis in measuring robustness via automated crowdsourcing of question perturbations and answers, 2025. URL https://arxiv.org/abs/2502. 03711

    2025

  16. [17]

    Fishnet: Financial intelligence from sub-querying, harmonizing, neural-conditioning, expert swarms, and task planning

    Nicole Cho, Nishan Srishankar, Lucas Cecchi, and William Watson. Fishnet: Financial intelligence from sub-querying, harmonizing, neural-conditioning, expert swarms, and task planning. In Proceedings of the 5th ACM International Conference on AI in Finance, ICAIF ’24, page 591–599. ACM, November 2024. doi: 10.1145/3677052.3698597. URL http://dx.doi.org/10....

    arXiv 2024

  17. [18]

    Brown, Miljan Martic, Shane Legg, and Dario Amodei

    Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences, 2023. URL https://arxiv.org/abs/1706.03741

    Pith/arXiv arXiv 2023

  18. [19]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. In NAACL, 2019

    2019

  19. [20]

    CLARK and RICHARD J

    HERBERT H. CLARK and RICHARD J. GERRIG. Understanding old words with new meanings, 1983. URL https://web.stanford.edu/~clark/1980s/Clark.Gerrig.oldwords.83.pdf

    1983

  20. [21]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1, 2018

    Pith/arXiv arXiv 2018

  21. [22]

    Charles L. A. Clarke, Nick Craswell, and Ian Soboroff. Overview of the TREC 2009 web track. In Ellen M. V oorhees and Lori P. Buckland, editors, Proceedings of The Eighteenth Text REtrieval Conference, TREC 2009, Gaithersburg, Maryland, USA, November 17-20, 2009 , volume 500-278 of NIST Special Publication . National Institute of Standards and Technology ...

    2009

  22. [23]

    Selectively answering ambiguous questions

    Jeremy Cole, Michael Zhang, Daniel Gillick, Julian Eisenschlos, Bhuwan Dhingra, and Jacob Eisenstein. Selectively answering ambiguous questions. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 530–543, Singapore, December 2023. Association for Computation...

    doi:10.18653/v1/2023 2023

  23. [24]

    hallucinations

    Dechert LLP. Ai expert challenged for relying on ai "hallucinations", Decem- ber 2024. URL https://www.dechert.com/knowledge/re-torts/2024/12/ ai-expert-challenged-for-relying-on-ai--hallucinations-.html . Accessed: 2025- 05-12

    2024

  24. [25]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    Pith/arXiv arXiv 2025

  25. [26]

    Prompt- ing and evaluating large language models for proactive dialogues: Clarification, target-guided, and non-collaboration

    Yang Deng, Lizi Liao, Liang Chen, Hongru Wang, Wenqiang Lei, and Tat-Seng Chua. Prompt- ing and evaluating large language models for proactive dialogues: Clarification, target-guided, and non-collaboration. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Findings of the Asso- ciation for Computational Linguistics: EMNLP 2023 , pages 10602–10621, Si...

    2023

  26. [27]

    Rephrase and respond: Let large language models ask better questions for themselves, 2024

    Yihe Deng, Weitong Zhang, Zixiang Chen, and Quanquan Gu. Rephrase and respond: Let large language models ask better questions for themselves, 2024. URL https://arxiv.org/abs/2311.04205

    Pith/arXiv arXiv 2024

  27. [28]

    Esin Durmus, Karina Nguyen, Thomas I. Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, Liane Lovitt, Sam McCandlish, Orowa Sikder, Alex Tamkin, Janel Thamkul, Jared Kaplan, Jack Clark, and Deep Ganguli. Towards measuring the representation of subjective global opinions in language mod...

    Pith/arXiv arXiv 2024

  28. [29]

    doi: 10.18653/v1/2023.findings-emnlp.711

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.711. URL https://aclanthology.org/2023.findings-emnlp.711/

    doi:10.18653/v1/2023.findings-emnlp.711 2023

  29. [30]

    Gptscore: Evaluate as you desire, 2023

    Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire, 2023. URL https://arxiv.org/abs/2302.04166

    Pith/arXiv arXiv 2023

  30. [31]

    Retrieval-augmented generation for large language models: A survey, 2024

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey, 2024. URL https://arxiv.org/abs/2312.10997

    Pith/arXiv arXiv 2024

  31. [32]

    On the comparative and absolute readings of superlatives, 2000

    Donka F Farkas and Katalin É Kiss. On the comparative and absolute readings of superlatives, 2000

    2000

  32. [33]

    E. J. Gumbel. The return period of flood flows. URL doi:10.1214/aoms/1177731747

    arXiv

  33. [34]

    Polysemy—Evidence from linguistics, behavioral science, and contextualized language models

    Janosch Haber and Massimo Poesio. Polysemy—Evidence from linguistics, behavioral science, and contextualized language models. Computational Linguistics, 50(1):351–417, March 2024. doi: 10.1162/ coli_a_00500. URL https://aclanthology.org/2024.cl-1.10/

    2024

  34. [35]

    The kl-ucb algorithm for bounded stochastic bandits and beyond,

    Aurélien Garivier and Olivier Cappé. The kl-ucb algorithm for bounded stochastic bandits and beyond,

  35. [36]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  36. [37]

    Leveraging Affirmative Interpretations from Negation Improves Natural Language Understanding

    Md Mosharaf Hossain and Eduardo Blanco. Leveraging affirmative interpretations from negation improves natural language understanding, 2022. URL https://arxiv.org/abs/2210.14486

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  37. [38]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55, January 2025

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55, January 2025. ISSN 1558-2868. doi: 10.1145/3703155....

    doi:10.1145/3703155 2025

  38. [39]

    Approximation to bayes risk in repeated play, 1957

    James Hannan. Approximation to bayes risk in repeated play, 1957

    1957

  39. [40]

    Towards mitigating LLM hallucination via self reflection

    Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. Towards mitigating LLM hallucination via self reflection. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1827–1843, Singapore, December

    2023

  40. [41]

    Towards mitigating hallucination in large language models via self-reflection, 2023

    Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. Towards mitigating hallucination in large language models via self-reflection, 2023. URL https://arxiv.org/abs/2310. 06271

    2023

  41. [42]

    Followbench: A multi-level fine-grained constraints following benchmark for large language models, 2024

    Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, and Wei Wang. Followbench: A multi-level fine-grained constraints following benchmark for large language models, 2024. URL https://arxiv.org/abs/2310.20410

    Pith/arXiv arXiv 2024

  42. [43]

    Adaptive-RAG: Learning to adapt retrieval-augmented large language models through question complexity

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong Park. Adaptive-RAG: Learning to adapt retrieval-augmented large language models through question complexity. In Kevin Duh, Helena Gomez, and Steven Bethard, editors, Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2024

  43. [44]

    triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension

    Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. arXiv e-prints, art. arXiv:1705.03551, 2017

    Pith/arXiv arXiv 2017

  44. [45]

    doi: 10.18653/v1/2023.findings-emnlp.123

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.123. URL https://aclanthology.org/2023.findings-emnlp.123/

    doi:10.18653/v1/2023.findings-emnlp.123 2023

  45. [46]

    Scope ambiguities in large language models

    Gaurav Kamath, Sebastian Schuster, Sowmya Vajjala, and Siva Reddy. Scope ambiguities in large language models. Transactions of the Association for Computational Linguistics , 12:738–754, 2024. ISSN 2307-387X. doi: 10.1162/tacl_a_00670. URL http://dx.doi.org/10.1162/tacl_a_00670

    work page doi:10.1162/tacl_a_00670 2024

  46. [47]

    Presupposition: What went wrong?, 2016

    Lauri Karttunen. Presupposition: What went wrong?, 2016

    2016

  47. [48]

    Matt Gardner Johannes Welbl, Nelson F. Liu. Crowdsourcing multiple choice science questions. 2017

    2017

  48. [49]

    Enriching rare word representations in neural language models by embedding matrix augmentation

    Yerbolat Khassanov, Zhiping Zeng, Van Tung Pham, Haihua Xu, and Eng Siong Chng. Enriching rare word representations in neural language models by embedding matrix augmentation. In Interspeech 2019, interspeech2019, page 3505–3509. ISCA, September 2019. doi: 10.21437/interspeech.2019-1858. URL http://dx.doi.org/10.21437/Interspeech.2019-1858

    work page doi:10.21437/interspeech.2019-1858 2019

  49. [50]

    Efficient algorithms for online decision problems.Journal of Computer and System Sciences, 71(3):291–307, 2005

    Adam Kalai and Santosh Vempala. Efficient algorithms for online decision problems.Journal of Computer and System Sciences, 71(3):291–307, 2005. ISSN 0022-0000. doi: https://doi.org/10.1016/j.jcss.2004. 10.016. URL https://www.sciencedirect.com/science/article/pii/S0022000004001394. Learning Theory 2003

    work page doi:10.1016/j.jcss.2004 2005

  50. [51]

    T. L. Lai and H. Robbins. Asymptotically efficient adaptive allocation rules

  51. [52]

    Bandit algorithms

    Tor Lattimore and Csaba Szepesvári. Bandit algorithms. Cambridge University Press, 2020

    2020

  52. [53]

    Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly

    Nora Kassner and Hinrich Schütze. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811– 7818, Online, July 2020. Association for Computationa...

    doi:10.18653/v1/2020.acl-main.698 2020

  53. [54]

    Factuality enhanced language models for open-ended text generation, 2023

    Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale Fung, Mohammad Shoeybi, and Bryan Catanzaro. Factuality enhanced language models for open-ended text generation, 2023. URL https: //arxiv.org/abs/2206.04624. 13

    Pith/arXiv arXiv 2023

  54. [55]

    Aligning language models to explicitly handle ambiguity

    Hyuhng Joon Kim, Youna Kim, Cheonbok Park, Junyeob Kim, Choonghyun Park, Kang Min Yoo, Sang-goo Lee, and Taeuk Kim. Aligning language models to explicitly handle ambiguity. In Yaser Al- Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1989–2007, Miami, Florida,...

    doi:10.18653/v1/2024.emnlp-main.119 2024

  55. [56]

    Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021. URL https://arxiv.org/ abs/2005.11401

    Pith/arXiv arXiv 2021

  56. [57]

    Long-context llms struggle with long in-context learning, 2024

    Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. Long-context llms struggle with long in-context learning, 2024. URL https://arxiv.org/abs/2404.02060

    Pith/arXiv arXiv 2024

  57. [58]

    Rlaif vs

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback, 2024. URL https://arxiv.org/abs/2309.00267

    Pith/arXiv arXiv 2024

  58. [59]

    A comparison of most-to-least and least-to-most prompting on the acquisition of solitary play skills, 2008

    Myrna E Libby, Julie S Weiss, Stacie Bancroft, and William H Ahearn. A comparison of most-to-least and least-to-most prompting on the acquisition of solitary play skills, 2008. URL https://pubmed. ncbi.nlm.nih.gov/22477678/

    work page arXiv 2008

  59. [60]

    Levinson

    Stephen C. Levinson. Pragmatics. pages 181–184. Cambridge Textbooks in Lin- guistics, 1983. URL https://www.cambridge.org/highereducation/books/pragmatics/ 6D0011901AE9E92CBC1F5F21D7C598C3#contents

    1983

  60. [61]

    ORANGE: a method for evaluating automatic evaluation metrics for machine translation

    Chin-Yew Lin and Franz Josef Och. ORANGE: a method for evaluating automatic evaluation metrics for machine translation. In COLING 2004: Proceedings of the 20th International Conference on Computational Linguistics, pages 501–507, Geneva, Switzerland, aug 23–aug 27 2004. COLING. URL https://www.aclweb.org/anthology/C04-1072

    2004

  61. [62]

    TruthfulQA: Measuring how models mimic human falsehoods

    Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 3214–3252, Dublin, Ireland, May 2022. Association for Computa- tional Linguistics. doi: 10.18653/v1/2022.acl-long.229. URL https://...

    doi:10.18653/v1/2022.acl-long.229 2022

  62. [63]

    Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models, 2024

    Ziniu Li, Tian Xu, Yushun Zhang, Zhihang Lin, Yang Yu, Ruoyu Sun, and Zhi-Quan Luo. Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models, 2024. URL https://arxiv.org/abs/2310.10505

    Pith/arXiv arXiv 2024

  63. [64]

    Query rewriting via large language models, 2024

    Jie Liu and Barzan Mozafari. Query rewriting via large language models, 2024. URL https://arxiv. org/abs/2403.09060

    arXiv 2024

  64. [65]

    ROUGE: A package for automatic evaluation of summaries

    Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain, July 2004. Association for Computational Linguistics. URL https://aclanthology.org/W04-1013/

    2004

  65. [66]

    G-eval: Nlg evaluation using gpt-4 with better human alignment, 2023

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment, 2023. URL https://arxiv.org/abs/2303. 16634

    2023

  66. [67]

    Subjective topic meets LLMs: Unleashing comprehensive, reflective and creative thinking through the negation of negation

    Fangrui Lv, Kaixiong Gong, Jian Liang, Xinyu Pang, and Changshui Zhang. Subjective topic meets LLMs: Unleashing comprehensive, reflective and creative thinking through the negation of negation. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages 12318–...

    work page doi:10.18653/v1/2024.emnlp-main.686 2024

  67. [68]

    We‘re afraid language models aren‘t modeling ambiguity

    Alisa Liu, Zhaofeng Wu, Julian Michael, Alane Suhr, Peter West, Alexander Koller, Swabha Swayamdipta, Noah Smith, and Yejin Choi. We‘re afraid language models aren‘t modeling ambiguity. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 790–807, Singapore, De...

    doi:10.18653/v1/2023.emnlp-main.51 2023

  68. [69]

    Self-refine: Iterative refinement with self-feedback, 2023

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback, 2023. URL https://arxiv.org/abs/2303.17651. 14

    Pith/arXiv arXiv 2023

  69. [70]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts, 2023. URL https://arxiv.org/ abs/2307.03172

    Pith/arXiv arXiv 2023

  70. [71]

    A Survey of Algorithms and Analysis for Adaptive Online Learning

    H. Brendan McMahan. A survey of algorithms and analysis for adaptive online learning, 2015. URL https://arxiv.org/abs/1403.3465

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  71. [72]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Conference on Empirical Methods in Natural Language Processing, 2018

    2018

  72. [73]

    Query rewriting for retrieval- augmented large language models, 2023

    Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. Query rewriting for retrieval- augmented large language models, 2023. URL https://arxiv.org/abs/2305.14283

    Pith/arXiv arXiv 2023

  73. [74]

    Efficient and robust algorithms for adversarial linear contextual bandits

    Gergely Neu and Julia Olkhovskaya. Efficient and robust algorithms for adversarial linear contextual bandits. In Jacob Abernethy and Shivani Agarwal, editors, Proceedings of Thirty Third Conference on Learning Theory, volume 125 of Proceedings of Machine Learning Research, pages 3049–3068. PMLR, 09–12 Jul 2020. URL https://proceedings.mlr.press/v125/neu20b.html

    2020

  74. [75]

    RaFe: Ranking feedback improves query rewriting for RAG

    Shengyu Mao, Yong Jiang, Boli Chen, Xiao Li, Peng Wang, Xinyu Wang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. RaFe: Ranking feedback improves query rewriting for RAG. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 884–901, Miami, Florida, USA, Novemb...

    doi:10.18653/v1/2024.findings-emnlp.49 2024

  75. [76]

    Openai o3 and o4-mini system card, 2025

    OpenAI. Openai o3 and o4-mini system card, 2025. URL https://openai.com/index/ o3-o4-mini-system-card/

    2025

  76. [77]

    (more) efficient reinforcement learning via posterior sampling

    Ian Osband, Daniel Russo, and Benjamin thompson. (more) efficient reinforcement learning via posterior sampling. Advances in Neural Information Processing Systems, 26, 2013

    2013

  77. [78]

    Controlled decoding from language models, 2024

    Sidharth Mudgal, Jong Lee, Harish Ganapathy, YaGuang Li, Tao Wang, Yanping Huang, Zhifeng Chen, Heng-Tze Cheng, Michael Collins, Trevor Strohman, Jilin Chen, Alex Beutel, and Ahmad Beirami. Controlled decoding from language models, 2024. URL https://arxiv.org/abs/2310.17022

    Pith/arXiv arXiv 2024

  78. [79]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei jing Zhu. Bleu: a method for automatic evaluation of machine translation. pages 311–318, 2002

    2002

  79. [80]

    Task-oriented query reformulation with reinforcement learn- ing

    Rodrigo Nogueira and Kyunghyun Cho. Task-oriented query reformulation with reinforcement learn- ing. In Martha Palmer, Rebecca Hwa, and Sebastian Riedel, editors, Proceedings of the 2017 Con- ference on Empirical Methods in Natural Language Processing , pages 574–583, Copenhagen, Den- mark, September 2017. Association for Computational Linguistics. doi: 1...

    work page doi:10.18653/v1/d17-1061 2017

  80. [81]

    Superlatives in Context: Modeling the Implicit Semantics of Superlatives

    Valentina Pyatkin, Bonnie Webber, Ido Dagan, and Reut Tsarfaty. Superlatives in context: Modeling the implicit semantics of superlatives, 2024. URL https://arxiv.org/abs/2405.20967

    work page internal anchor Pith review Pith/arXiv arXiv 2024

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.