Pith. sign in

REVIEW 4 major objections 6 minor 40 references

OlaMind claims that distilling expert agent reasoning, then basic-to-hard SFT/RL with judge, rule, and hallucination rewards, yields more human-like, safer customer-service models whose gains survive online A/B tests.

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

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

T0 review · deepseek-v4-flash

2026-08-04 08:08 UTC pith:U62XPS3H

load-bearing objection The online A/B results are the real news, but the offline hallucination and risk numbers are scored by the same judges used as training rewards, so they need independent validation before being trusted. the 4 major comments →

arxiv 2510.22143 v3 pith:U62XPS3H submitted 2025-10-25 cs.CL

Benchmarking and Learning Real-World Customer Service Dialogue

Tianhong Gao , Jundong Shen , Jiapeng Wang , Bei Shi , Ying Ju , Junfeng Yao , Huiyu Yu This is my paper
classification cs.CL
keywords customer service dialogueretrieval-augmented generationreinforcement learningsupervised fine-tuninghallucination detectionLLM-as-a-judgehuman-likenesschain-of-thought distillation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the gap between offline leaderboard performance and real-world deployable behavior in industrial customer service can be closed by a specific training loop. The loop first distills the hidden reasoning processes and response strategies of expert human agents into a model (Learn-to-Think), then runs two rounds of rejection-sampled supervised fine-tuning and reinforcement learning, from easy to hard data, with rewards drawn from LLM judges, rule checks, and a dedicated hallucination detector (Learn-to-Respond). The paper reports that the final model cuts hallucination rate to 8.6% and critical business risk to 6.0% in offline evaluation, and in live A/B tests raises intelligent resolution by roughly 29% and 18% in two scenarios while lowering human takeover by 6-7%. If true, this is evidence that imitating expert thinking plus a staged, reward-guided refinement is the lever, not just a bigger base model, for making customer-service AI human-like and safe.

Core claim

The paper's central claim is that OlaMind-Hard-RL, the product of the complete Learn-to-Think plus Learn-to-Respond pipeline, outperforms every compared general-purpose LLM on the paper's four offline customer-service metrics: a 4.19/5 human-likeness score, a 41.3% dialogue GSB win rate, a 6.0% critical business risk rate, and an 8.6% hallucination rate, against a direct-distillation baseline's 3.21/31.2%/33.1%. The same model then shows +28.92% / +18.42% relative intelligent resolution rate and -6.08% / -7.12% human takeover rate in online A/B tests in community-support and livestream-interaction RAG scenarios. The paper also claims the recipe transfers: applying the same training stages to

What carries the argument

The load-bearing mechanism is the two-stage curriculum the paper calls Learn-to-Think and Learn-to-Respond. Learn-to-Think uses a strong LLM to mine expert human customer-service representatives' dialogues into structured reasoning processes and response strategies, training an intermediate thinker model that produces chain-of-thought for every reply. Learn-to-Respond then alternates rejection sampling and refinement with cold-start SFT and group relative policy optimization (GRPO) in two stages: a basic stage with LLM-as-a-judge rewards (human-likeness, dialogue GSB, risk), and a hard stage that adds rule-based rewards (format, length with soft penalty, rule-match) and a dedicated hallucina

Load-bearing premise

The reported offline numbers assume the same LLM judges that served as training rewards — human-likeness, dialogue GSB, risk, and the hallucination detector — are valid, well-calibrated proxies for human judgment that the policy cannot exploit; the paper reports no independent, held-out human agreement for these metrics.

What would settle it

Take the released 3,000 sample test sets and have fresh, blinded human annotators score OlaMind-Hard-RL and the baseline on human-likeness (1-5), dialogue GSB, business risk, and hallucination. If the human-rated human-likeness score is below 3.5 or the human-confirmed hallucination rate is above about 15%, the central claim that the method is genuinely more human-like and hallucination-safe would be undercut, because the reported gains could be artifacts of the judge being the same model that was optimized.

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

If this is right

  • Industrial customer service can optimize for subjective quality and failure-mode safety, not just task completion, because the paper shows these are trainable targets.
  • A latency-friendly small model trained with the recipe can beat much larger general-purpose models on domain dialogue metrics, lowering serving cost.
  • Hallucination-driven business risk can be reduced by an order of magnitude with a dedicated reward signal, independent of retrieval-side changes.
  • Offline gains from this benchmark-to-optimization loop carry over to production: both test scenarios show higher intelligent resolution and lower human takeover rates.
  • The recipe's benefits transfer across model families, so the pipeline itself, not a specific base model, is the contribution.

Where Pith is reading between the lines

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

  • Because the LLM judges used for offline evaluation are the same as the RL rewards, the reported 4.19/41.3%/6.0%/8.6% numbers may reflect reward optimization as much as genuinely safer, more human dialogue; a blinded human re-evaluation of the 3,000 released samples would settle how much of the gap is real.
  • The steep drop in hallucination from OlaMind-Think (52.5%) to final RL (8.6%), while SFT alone leaves it at 23.7-38.9%, suggests the safety gains come mainly from reward design rather than from the distilled reasoning traces; a natural follow-up would ablate the hallucination reward's effect on out-of-distribution questions.
  • The length-reward with soft penalty is a tunable safety valve; one could test adaptive length targets per query complexity to push resolution rates further without sacrificing conciseness.
  • The paper's basic-to-hard and judge-to-rule ordering may generalize to other high-stakes RAG dialogues where reward hacking and unverifiable rewards are the bottleneck, such as technical support or clinical advice, though the paper itself only claims customer service.

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

4 major / 6 minor

Summary. The paper presents OlaMind, a multi-stage training framework for retrieval-augmented customer-service dialogue. It begins with a Learn-to-Think stage that distills human-like reasoning processes and response strategies from expert service records, followed by a Learn-to-Respond stage of cold-start SFT and basic-to-hard GRPO reinforcement learning. The RL reward combines rule-based components, LLM-as-a-judge rewards (human-likeness, dialogue GSB, and risk), and a dedicated hallucination-judge reward. Offline evaluations on 1,000-dialogue test sets report a human-likeness score of 4.19, dialogue GSB of 41.3%, critical business risk rate of 6.0%, and hallucination rate of 8.6%. Online A/B tests in community-support and livestream-interaction scenarios report relative IRR increases of +28.92%/+18.42% and HTR reductions of -6.08%/-7.12%. The framework is also applied to Qwen3-8B and Qwen3-14B, showing consistent trends.

Significance. If the results hold, the paper provides an industrially valuable recipe for building human-like and hallucination-safe customer-service agents: a complete pipeline from expert-behavior distillation to multi-reward RL, with user-facing online evidence. The online A/B test is a genuine strength, as IRR and HTR are direct business metrics that do not depend on LLM judges. The explicit hallucination-judge system and the staged basic-to-hard training are useful contributions. However, the central safety claims — the 8.6% hallucination rate and 6.0% risk rate — are evaluated with the same judge functions used as RL rewards, creating a serious circularity concern. The online metrics do not directly measure hallucination or business risk, so the paper's headline safety improvements require independent validation. The transferability results are suggestive but likewise rely on the same circular offline metrics.

major comments (4)
  1. [Section 3.3.4 vs. Section 4, Eqs. (5)–(7) and Table 2] The offline headline metrics in Table 2 (Human-Likeness Score, Dialogue GSB Score, Critical Business Risk Rate, Hallucination Rate) are computed by the same judge functions used as RL rewards: Jhuman, JGSB, and Jrisk appear in Eq. (6), and OlaMind-Halluc-Judge provides Rhl in Eq. (7). No held-out judge, human–judge agreement (e.g., Cohen's kappa), or confidence intervals are reported. The policy is optimized to maximize these exact scores, so the reported 8.6% hallucination and 6.0% risk rates may reflect reward hacking rather than genuine safety or quality. The online A/B test (Table 3) is user-facing and directional, but it measures issue resolution and human takeover, not hallucination or business risk. Please add independent validation: human annotation on a random subset, a held-out judge from a different model family, and bootstrap confidence intervals.
  2. [Section 3.3.4, hallucination judge; Table 2] OlaMind-Halluc-Judge serves both as the reward model (Rhl in Eq. 7) and as the hallucination-rate estimator in offline evaluation. The training pipeline includes human-in-the-loop annotation, but its precision/recall/F1 on a held-out human-labeled set is never reported. Without this, the 8.6% hallucination rate is not interpretable — the judge's false positive/negative rates could dominate the differences between models. Please report the judge's accuracy against a held-out human-labeled sample and compare it with at least one independent strong LLM judge.
  3. [Appendix B and Section 4, human evaluation] The human evaluation in Appendix B covers only dialogue GSB, for two comparisons (OlaMind-Basic-SFT vs. baseline, and OlaMind-Hard-RL vs. OlaMind-Basic-SFT). It does not validate human-likeness scores, risk rates, or hallucination rates. These safety-related metrics are the central claims of the paper, and in the offline section they rely entirely on LLM judges that are also training rewards. Please add human evaluation, or at minimum judge–human agreement, for risk and hallucination on a representative subsample.
  4. [Abstract vs. full text] The provided abstract (and the arXiv title) describe an OlaBench benchmark and report OlaBench results (83.64 vs. GPT-5.2/Gemini 3 Pro). The full text contains no OlaBench section, no GPT-5.2/Gemini 3 Pro comparisons, and the benchmark is not described anywhere in the body. The manuscript's actual abstract and title are about OlaMind. Please align the abstract with the submitted content, or include the OlaBench material if it is intended to be part of the contribution.
minor comments (6)
  1. [Section 3.3.1, Eq. (1)] The notation J_GSB=1 is used to mean 'better than the reference response,' but the GSB prompt outputs 'Good/Same/Bad.' Please clarify the mapping to numeric values.
  2. [Section 3.3.4, Eq. (4)] The cache ratio ρ in the length reward is listed as a free parameter but its value is not reported. Please provide the value and, ideally, a sensitivity analysis.
  3. [Section 3.3.4, reward weights] The reward weights α1=0.2, α2=0.5, α3=1, β1=β2=β3=1, γ=5 are given without ablation or sensitivity analysis. A brief study of the most important weights (e.g., γ for hallucination) would strengthen the paper.
  4. [Section 4, evaluation setup] The risk and hallucination test sets are described as 'constructed' and 'manually annotated,' but no annotation guidelines, inter-annotator agreement, or sample-selection details are reported. Please add these details, or state that the data will be released.
  5. [Table 3, online A/B] Online A/B metrics are reported as relative improvements without absolute values, sample sizes, or statistical significance. Please include confidence intervals or significance tests.
  6. [Section 4.2, Figure 8] The reward-dynamics plot would benefit from a description of how the judge-based rewards are normalized and what the validation set consists of.

Circularity Check

4 steps flagged

Same LLM judges used as RL rewards also score the headline offline metrics; the hallucination rate is measured by the same reward model, so the 8.6% figure is not an independent result.

specific steps
  1. fitted input called prediction [Section 3.3.4 (Eqs. 6-7); Section 4, Implementation Details / Evaluation Setup; Table 2]
    "LLM-as-a-judge reward (R judge) (1) Human-likeness reward ( Rhuman) follows human-likeness evaluation J human. ... we construct an automatic evaluation system for human-likeness, as well as dialogue GSB assessment, critical business risk evaluation, and hallucination detection. ... We utilize Doubao-1.5-pro and DeepSeek-R1 for J human, J GSB and J risk evaluation."

    Rhuman is an explicit reward term in Eq. (6) and Eq. (7), so the policy is trained to maximize Jhuman. The headline 'Human-Likeness Score' in Table 2 is then produced by the same Jhuman evaluation described in Section 4. No separate held-out human-likeness judge is described. The reported 4.19 score therefore partly measures success at optimizing the exact scoring function used as the reward, not an independent human-likeness assessment.

  2. fitted input called prediction [Section 3.3.4 (Eqs. 6-7); Section 4 Evaluation Setup (Eq. 8); Table 2]
    "Dialogue GSB reward( RGSB) follows dialogue GSB evaluation JGSB. ... we construct an automatic evaluation system for human-likeness, as well as dialogue GSB assessment ... ScoreGSB = Ngood - Nbad / (Ngood + Nsame + Nbad)."

    RGSB is a reward component in Eq. (6) and Eq. (7), and the offline 'Dialogue GSB Score' is computed by the automatic GSB evaluation system described in Section 4. The paper does not report a distinct judge for offline GSB evaluation. Thus the 41.3% GSB score is the same criterion the RL policy was rewarded to maximize, so the improvement is at least partly an artifact of optimizing the evaluator rather than an independent measure of dialogue quality.

  3. fitted input called prediction [Section 3.3.4 (Eqs. 6-7); Section 4, Evaluation Setup; Table 2]
    "Risk reward( Rrisk) follows risk evaluation Jrisk. ... we construct an automatic evaluation system for human-likeness, as well as dialogue GSB assessment, critical business risk evaluation, and hallucination detection."

    Rrisk is rewarded during RL through Eq. (6) and Eq. (7), and the 'Critical Business Risk Rate' in Table 2 is measured by the same automatic risk evaluation system. No independent risk judge, human agreement rate, or held-out risk set with external labels is reported. The 6.0% risk rate is therefore a measure of the policy's success under the same Jrisk criterion used to train it.

  4. fitted input called prediction [Section 3.3.4 (Hallucination reward, Eq. 7); Section 4 Evaluation Setup; Table 2]
    "Hallucination reward (R hl) Rhl is incorporated to mitigate hallucinations ... and provided by a dedicated hallucination-judge system with human&LLM interaction ... Hallucination detection: Another 1,000 manually annotated hallucination-prone cases are used for hallucination detection with OlaMind-Halluc-Judge."

    OlaMind-Halluc-Judge is the model that supplies Rhl in Eq. (7), and it is also the model used to compute the offline hallucination rate in Table 2. The paper reports no held-out human-labeled accuracy for this judge and no independent hallucination metric. The 8.6% hallucination rate is thus the output of the same fitted judge that was used as the reward signal, so the claim that RL 'drives hallucination down' is not independently verified; it shows that the policy learned to satisfy this particular judge.

full rationale

The paper's derivation chain contains a genuine circular segment: the policy is optimized with Jhuman, JGSB, Jrisk, and OlaMind-Halluc-Judge as rewards (Eqs. 6-7), and then the headline offline results in Table 2 — human-likeness 4.19, GSB 41.3%, risk 6.0%, hallucination 8.6% — are scored with the same automatic judges. For hallucination specifically, the reward model and the estimator are literally the same system. No independent held-out judge or human-agreement statistics are reported for these absolute rates, so the offline safety and human-likeness numbers are not external confirmations. However, the paper is not wholly circular: the online A/B metrics (IRR and HTR) are user-facing and independent of the LLM judge functions, and Appendix B provides human GSB annotations on the same 1,000 dialogues, though these are relative comparisons rather than absolute hallucination or risk rates. The core training recipe may still be effective, but the central offline 'hallucination-safe' and 'human-like' claims are partially reduced to fitting the same judge functions. Score 6 reflects this partial, multi-metric circularity while recognizing the independent online signal.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central empirical claims rest on several domain assumptions that are not independently benchmarked: LLM judges are treated as ground truth for quality and safety, LLM-extracted expert reasoning is treated as faithful, the authors' hallucination classifier is treated as a valid measurement instrument, and online metrics are treated as unbiased. The free parameters (reward weights, cache ratio) are hand-set and directly shape the reported tradeoffs.

free parameters (3)
  • Reward weights α1, α2, α3, β1, β2, β3, γ = 0.2, 0.5, 1, 1, 1, 1, 5
    Set empirically in Eqs. (5)-(7); directly control the objective optimized by RL, so the reported tradeoffs among risk, hallucination, human-likeness, and length depend on hand-chosen weights.
  • Cache ratio ρ in length reward = not reported
    Defines Lcache=ρLref in Eq. (4) and controls the soft penalty margin for overlong responses; the value is not stated in the paper.
  • Rejection sampling count NRJ = 8
    Sampling budget for candidate CoT/response selection in the basic stage; larger values would improve coverage at added compute cost and affect selected data quality.
axioms (4)
  • domain assumption LLM-as-a-judge evaluations (Jhuman, JGSB, Jrisk) are valid proxies for human quality and safety judgments.
    Used both as rejection-sampling filters and RL rewards (Section 3.3) and as offline metrics (Section 4). If judge preferences diverge from human/user preferences, the claimed quality gains are not established; no agreement statistics with human annotations are reported for these judge scores.
  • domain assumption DeepSeek-R1 extraction of human CSRs' reasoning processes and response strategies captures real expert cognition and generalizes across scenarios.
    Learn-to-Think (Section 3.2) feeds these extracted CoTs into training; the prompt asks for generalizable strategies, but there is no validation that the extracted reasoning matches actual CSR cognition.
  • domain assumption OlaMind-Halluc-Judge's labels are a valid measure of true hallucination and a safe reward signal.
    The judge is trained via the human-and-LLM pipeline (Section 3.3.4) and in turn measures the hallucination rate in Table 2; no external benchmark or agreement metric is given.
  • domain assumption Intelligent resolution rate (IRR) and human takeover rate (HTR) are unbiased measures of service quality in online A/B testing.
    IRR is user-marked resolution and HTR reflects user demand for human takeover. No absolute rates, sample sizes, or significance tests are reported for Table 3.
invented entities (2)
  • OlaMind-Think intermediate model no independent evidence
    purpose: Trained to produce distilled reasoning and response strategies from expert dialogues; used as the CoT generator for cold-start response sampling.
    New model component with no external validation beyond downstream OlaMind scores; its quality is only assessed indirectly through final system metrics.
  • OlaMind-Halluc-Judge no independent evidence
    purpose: Classifies hallucinations into three types, provides the hallucination reward, and also reports hallucination rate in evaluation.
    The authors' own classifier; no external benchmark or human-agreement metric is reported, and it is used both as a training reward and as the evaluation instrument.

pith-pipeline@v1.3.0-alltime-deepseek · 14861 in / 12567 out tokens · 112069 ms · 2026-08-04T08:08:48.653333+00:00 · methodology

0 comments
read the original abstract

Existing benchmarks and training pipelines for industrial intelligent customer service (ICS) remain misaligned with real-world dialogue requirements, overemphasizing verifiable task success while under-measuring subjective service quality and realistic failure modes, leaving a gap between offline gains and deployable dialogue behavior. We close this gap with a benchmark-to-optimization loop: we first introduce OlaBench, an ICS benchmark spanning retrieval-augmented generation, workflow-based systems, and agentic settings, which evaluates service capability, safety, and latency sensitivity; moreover, motivated by OlaBench results showing state-of-the-art LLMs still fall short, we propose OlaMind, which distills reusable reasoning patterns and service strategies from expert dialogues and applies staged exploration--exploitation reinforcement learning with instance-level rubric-aware guidance to improve model capability. OlaMind surpasses GPT-5.2 and Gemini 3 Pro on OlaBench (83.64 vs. 70.58/70.84) and, in online A/B tests, delivers an average +23.67% issue resolution and -6.6% human transfer rate versus the baseline, bridging offline gains to deployment. Together, OlaBench and OlaMind advance ICS systems toward more anthropomorphic, professional, and reliable deployment. The project page and evaluation are available at https://olamind-olabench.github.io.

Figures

Figures reproduced from arXiv: 2510.22143 by Bei Shi, Huiyu Yu, Jiapeng Wang, Jundong Shen, Junfeng Yao, Tianhong Gao, Ying Ju.

Figure 1
Figure 1. Figure 1: Comparisons between human customer ser￾vice, intelligent customer service with general LLMs, and our proposed OlaMind. Notably, OlaMind can think and respond with human-likeness and hallucination￾safety, thus significantly bridging the gap. 2022; Olujimi and Ade-Ibijola, 2023), offering piv￾otal benefits of automation, personalized support, and optimized efficiency. Recently, large language models (LLMs) (… view at source ↗
Figure 2
Figure 2. Figure 2: The overall framework of OlaMind. It introduces [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the reasoning process and response strategy between OlaMind-Think and general LLMs. LLMs to focus on human-likeness while ensuring the practical applicability of the generated outputs. Our primary objective is to transcend case-specific problems or context-dependent solutions, thereby demonstrating generalizability across different ser￾vice scenarios. The related prompt is given in Ap￾pendix … view at source ↗
Figure 4
Figure 4. Figure 4: Training pipeline for hallucination detection model. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performances across different stages of Ola [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Conversation snippets between user and intelligent customer service of different models. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Trends of key reward metrics in validation set. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 18 linked inside Pith

  1. [1]

    Guy Azov, Tatiana Pelc, Adi Fledel Alon, and Gila Kamhi. 2024. Self-improving customer review response generation based on llms. arXiv preprint arXiv:2405.03845

    Pith/arXiv arXiv 2024

  2. [2]

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. ^ 2 -bench: Evaluating conversational agents in a dual-control environment. arXiv preprint arXiv:2506.07982

    Pith/arXiv arXiv 2025

  3. [3]

    Antonin Brun, Ruying Liu, Aryan Shukla, Frances Watson, and Jonathan Gratch. 2025. Exploring emotion-sensitive llm-based conversational ai. arXiv preprint arXiv:2502.08920

    Pith/arXiv arXiv 2025

  4. [4]

    Ethem Ya g z C al k and Talha R \"u zgar Akku s . 2025. Enhancing human-like responses in large language models. arXiv preprint arXiv:2501.05032

    arXiv 2025

  5. [5]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261

    Pith/arXiv arXiv 2025

  6. [6]

    Huifang Du, Shuqin Li, Minghao Wu, Xuejing Feng, Yuan-Fang Li, and Haofen Wang. 2024. Rewarding what matters: Step-by-step reinforcement learning for task-oriented dialogue. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 8030--8046

    2024

  7. [7]

    Tianhong Gao, Genhang Shen, Yuxuan Wu, Zunlei Feng, Jinshan Zhang, and Sheng Zhou. 2025. Ecommir: Towards intelligent multimodal intent recognition in e-commerce dialogue systems. In Companion Proceedings of the ACM on Web Conference 2025, pages 3049--3052

    2025

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    Pith/arXiv arXiv 2025

  9. [9]

    Mengze Hong, Chen Jason Zhang, Di Jiang, Yuanfeng Song, Lu Wang, Yuanqin He, Zhiyang Su, and Qing Li. 2024. Expanding chatbot knowledge in customer service: Context-aware similar question generation using large language models. arXiv preprint arXiv:2410.12444

    arXiv 2024

  10. [10]

    Xinyi Jiang, Tianyi Hu, Yuheng Qin, Guoming Wang, Zhou Huan, Kehan Chen, Gang Huang, Rongxing Lu, and Siliang Tang. 2025. Chatmap: Mining human thought processes for customer service chatbots via multi-agent collaboration. In Findings of the Association for Computational Linguistics: ACL 2025, pages 11927--11947

    2025

  11. [11]

    Sunwoo Lee, Dhammiko Arya, Seung-Mo Cho, Gyoung-eun Han, Seokyoung Hong, Wonbeom Jang, Seojin Lee, Sohee Park, Sereimony Sek, Injee Song, and 1 others. 2024. Telbench: A benchmark for evaluating telco-specific large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 609--626

    2024

  12. [12]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459--9474

    2020

  13. [13]

    Junhua Liu, Yong Keat Tan, Bin Fu, and Kwan Hui Lim. 2024. Lara: Linguistic-adaptive retrieval-augmentation for multi-turn intent classification. arXiv preprint arXiv:2403.16504

    Pith/arXiv arXiv 2024

  14. [14]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534--46594

    2023

  15. [15]

    Malak Mashaabi, Areej Alotaibi, Hala Qudaih, Raghad Alnashwan, and Hend Al-Khalifa. 2022. Natural language processing in customer service: A systematic review. arXiv preprint arXiv:2212.09523

    Pith/arXiv arXiv 2022

  16. [16]

    Agada Joseph Oche, Ademola Glory Folashade, Tirthankar Ghosal, and Arpan Biswas. 2025. A systematic review of key retrieval-augmented generation (rag) systems: progress, gaps, and future directions. arXiv preprint arXiv:2507.18910

    Pith/arXiv arXiv 2025

  17. [17]

    Peter Adebowale Olujimi and Abejide Ade-Ibijola. 2023. Nlp techniques for automating responses to customer queries: A systematic review. Discover Artificial Intelligence, 3(1):20

    2023

  18. [18]

    OpenAI. 2025. Gpt-5 system card. https://cdn.openai.com/gpt-5-system-card.pdf

    2025

  19. [19]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730--27744

    2022

  20. [20]

    Pragaash Ponnusamy, Clint Solomon Mathialagan, Gustavo Aguilar, Chengyuan Ma, and Chenlei Guo. 2022. Self-aware feedback-based self-learning in large-scale conversational ai. NAACL-HLT 2022, page 324

    2022

  21. [21]

    Akshara Prabhakar, Zuxin Liu, Ming Zhu, Jianguo Zhang, Tulika Awalgaonkar, Shiyu Wang, Zhiwei Liu, Haolin Chen, Thai Hoang, Juan Carlos Niebles, and 1 others. 2025. Apigen-mt: Agentic pipeline for multi-turn data generation via simulated agent-human interplay. arXiv preprint arXiv:2504.03601

    Pith/arXiv arXiv 2025

  22. [22]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728--53741

    2023

  23. [23]

    Nirmal Roy, Leonardo Ribeiro, Rexhina Blloshmi, and Kevin Small. 2024. Learning when to retrieve, what to rewrite, and how to respond in conversational qa. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 10604--10625

    2024

  24. [24]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    Pith/arXiv arXiv 2017

  25. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    Pith/arXiv arXiv 2024

  26. [26]

    Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. 2022. Defining and characterizing reward hacking. Advances in Neural Information Processing Systems, 35:9460--9471

    2022

  27. [27]

    Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534

    Pith/arXiv arXiv 2025

  28. [28]

    Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z Pan, and Kam-Fai Wong. 2024 a . Unims-rag: A unified multi-source retrieval-augmented generation for personalized dialogue systems. arXiv preprint arXiv:2401.13256

    Pith/arXiv arXiv 2024

  29. [29]

    Zilong Wang, Zifeng Wang, Long Le, Huaixiu Steven Zheng, Swaroop Mishra, Vincent Perot, Yuwei Zhang, Anush Mattapalli, Ankur Taly, Jingbo Shang, and 1 others. 2024 b . Speculative rag: Enhancing retrieval augmented generation through drafting. arXiv preprint arXiv:2407.08223

    Pith/arXiv arXiv 2024

  30. [30]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  31. [31]

    Yufei Xiang, Yiqun Shen, Yeqin Zhang, and Nguyen Cam-Tu. 2024. Retrospex: Language agent meets offline reinforcement learning critic. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4650--4666

    2024

  32. [32]

    Shicheng Xu, Liang Pang, Jun Xu, Huawei Shen, and Xueqi Cheng. 2024 a . List-aware reranking-truncation joint model for search and retrieval-augmented generation. In Proceedings of the ACM Web Conference 2024, pages 1330--1340

    2024

  33. [33]

    Zhentao Xu, Mark Jerome Cruz, Matthew Guevara, Tie Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024 b . Retrieval-augmented generation with knowledge graphs for customer service question answering. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pages 2905--2909

    2024

  34. [34]

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024 c . Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817

    Pith/arXiv arXiv 2024

  35. [35]

    Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling. 2024. Corrective retrieval augmented generation. arXiv preprint arXiv:2401.15884

    Pith/arXiv arXiv 2024

  36. [36]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    Pith/arXiv arXiv 2025

  37. [37]

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, and 1 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476

    Pith/arXiv arXiv 2025

  38. [38]

    Yangyang Zhao, Ben Niu, Mehdi Dastani, and Shihan Wang. 2024. Bootstrapped policy learning for task-oriented dialogue through goal shaping. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4566--4580

    2024

  39. [39]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  40. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...