Pith. sign in

REVIEW 4 major objections 4 minor 29 references

A 0.6B-parameter language model, trained with supervised fine-tuning then reinforcement learning on retrieval-quality rewards, selects retrieval agents more accurately than larger models that route on intent alone.

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 01:38 UTC pith:AON34PKG

load-bearing objection Interesting recipe, but the evaluation metric is self-referential and the headline NDCG deltas don't support the central claim. the 4 major comments →

arxiv 2608.00030 v1 pith:AON34PKG submitted 2026-07-15 cs.CL cs.AI

SLMs as Multi-Agent Routers: A Progressive SFT and Reinforcement Learning Approach

Gayathri V Kondapalli , Alexander Ng , Hirsh Pithadia , Rahul Monish , Harvey Yorke , Amir Kayhani This is my paper
classification cs.CL cs.AI
keywords multi-agent routingquery routingsmall language modelsreinforcement learningsupervised fine-tuningretrieval relevanceNDCG@10agent-query mismatch
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 paper tries to establish that agent selection quality is determined by the training signal, not model size. It trains a 0.6B-parameter language model in two stages: supervised fine-tuning teaches it to output a structured routing tuple (agent pool selection, keywords, time range), then reinforcement learning provides a reward based on judged retrieval relevance, teaching it when a topically appropriate agent is actually a poor fit. The resulting router achieves mean NDCG@10 of 0.771 versus 0.594 and 0.552 for two larger LLM routers that use only query intent, and 0.918 versus 0.539 and 0.490 on a targeted subset of agent-query mismatches, with mean selection latency of 120.1ms. If correct, this would mean accurate, cheap, fast agent selection can sit on the critical path of production retrieval systems.

Core claim

The central claim is that a 0.6B language model can outperform much larger LLM routers if its training signal includes retrieval relevance. Supervised fine-tuning alone reaches 89.4% source exact match and learns to parse queries into structured routing parameters, but it cannot detect when a topically aligned agent retrieves poorly because it never observes results. Reinforcement learning, using a hierarchical reward that combines routing fidelity with judged relevance of retrieved content, adds that signal. On the full test set the trained router reaches mean NDCG@10 of 0.771, beating two larger intent-only LLMs (0.594 and 0.552); on a curated subset of queries where the topical specialist

What carries the argument

The load-bearing piece is the hierarchical reward function used during reinforcement learning: R = λ_src R_src + λ_con R_con. R_src scores how well the routing decision matches query intent (source bitmask precision/recall, date accuracy, keyword accuracy), while R_con scores the quality of retrieved content using an LLM judge across direct answerability, completeness, semantic alignment, and trust. This reward, applied with a KL penalty to keep the policy near the SFT checkpoint, is what lets the small model correlate selection decisions with actual retrieval outcomes and learn to redirect mismatched queries.

Load-bearing premise

The load-bearing premise is that the LLM judge used to create SFT labels, to compute the content-quality reward, and to score final NDCG is a valid and impartial proxy for human relevance judgments.

What would settle it

Have human raters score the same pooled result lists and recompute NDCG@10 per router. If the trained router's advantage over the two LLM routers shrinks or disappears under human labels, the central claim fails; if the advantage persists, it is confirmed. A complementary check is to replace the LLM judge with a different judge and see if the rank ordering of the routers changes.

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

If this is right

  • Selection latency of 120.1ms mean (P99 179.4ms) puts the router under production-friendly budgets, whereas the two LLM baselines take 683.6ms and 2,457ms on average.
  • Routing can be grounded in observed retrieval quality, so the system adapts to which agents actually produce relevant results for a given query distribution.
  • Small language models can substitute for LLM-based selectors, lowering per-query inference cost without sacrificing selection quality.
  • The reinforcement-learning stage is necessary for mismatch detection; SFT alone cannot learn that a topically aligned agent may be the wrong choice.
  • Because median NDCG scores are close across routers, the gains concentrate on hard, ambiguous queries, meaning the approach improves worst-case routing rather than easy cases.

Where Pith is reading between the lines

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

  • The same progressive training recipe could be applied to other agentic decisions with observable downstream outcomes, such as choosing among tools, databases, or query reformulation strategies.
  • If the reward hierarchy transfers, a router trained on a fixed agent pool could be extended to a dynamic pool by conditioning on explicit agent capability descriptions, a direction the paper leaves open.
  • The current evaluation compares against intent-only LLM routers; a testable next step is to train a similarly RL-grounded larger model and compare, to isolate the contribution of the training signal from model size.
  • The reported advantage is measured against one LLM judge's relevance labels; using human relevance judgments or a different judge would reveal whether the gains represent genuine retrieval quality for users.

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 / 4 minor

Summary. The paper proposes a two-stage post-training method for a 0.6B Qwen SLM to act as a router over 11 domain-specific retrieval agents. It uses SFT on 56k labelled query-output tuples, then REINFORCE++ with a hierarchical reward combining query-intrinsic routing quality (source bitmask, keywords, dates) and LLM-judged content quality of retrieved results. The main claimed result is that this router achieves NDCG@10 of 0.771 vs 0.594/0.552 for Nova Lite/Haiku, and 0.918 vs 0.539/0.490 on a targeted 'agent-query mismatch' subset, with 120.1ms mean latency. The paper attributes this to retrieval-quality-grounded RL and concludes that agent selection quality is determined by the training signal, not model size.

Significance. If the central comparison were valid, this would be a useful demonstration that a small router can outperform much larger prompted LLMs on agent selection when trained with retrieval-grounded rewards, with substantial latency savings. The two-stage recipe is straightforwardly specified (LoRA hyperparameters, RL coefficients, prompts), and the latency measurements are concrete. However, the evaluation as reported does not support the claimed NDCG advantages: the per-router IDCG normalization makes the headline metric uninterpretable, and the use of one LLM as labeler, reward, robustness-subset constructor, and final judge raises serious circularity concerns. The result is therefore currently an interesting hypothesis rather than an established empirical finding.

major comments (4)
  1. [§4.5.1, Eq. (10)] IDCG@10 is computed by sorting the same 10 relevance scores returned by each router. This makes every NDCG self-normalized: a router returning 3 relevant results at ranks 1–3 and 7 irrelevant results scores 1.0, while a router returning 8 relevant results with one non-relevant at rank 2 scores lower. Even with perfect human labels, this denominator prevents cross-router comparison, because it measures ranking quality within each router's own retrieved set, not which router retrieved more relevant content. Since the abstract and §6 rely on the Table 2/4 deltas, the central quantitative claim is unsupported. Please recompute NDCG with a common IDCG (e.g., over the pooled relevant set or the full relevant set for the query), or use recall-oriented metrics with a fixed denominator.
  2. [§3.4, §3.5, §4.5.1] Claude Sonnet 4.6 is used for SFT labels (§3.4), for the content-quality reward R_con in Eq. (3), and for final NDCG relevance judgments (§4.5.1); the robustness subset is also constructed by LLM-as-judge (§4.5.2). The router is thus trained, rewarded, and scored by one subjective model. The reported gains may measure alignment with that judge's preferences rather than retrieval quality for users. The §6 limitation asks for human relevance assessments, but that acknowledgement does not address the fact that the current pipeline contains no independent signal whatever. Please add human relevance judgments on a held-out sample and use a different judge for final evaluation than for training/reward, with agreement statistics.
  3. [§4.5.2, Table 4] The 'targeted subset' of agent-query mismatches has no reported size, selection rule, or confidence intervals. Because the subset is chosen by the same judge used in training and reward, the 0.918 vs 0.539/0.490 deltas may be an artifact of overfitting to that judge's notion of mismatch. Report the subset size, how queries were filtered, per-query NDCG distributions, and a significance test against the baselines.
  4. [§5.2] No retrieval-quality metric is reported for the SFT-only model. Table 3 gives source match, keyword similarity, and date accuracy, but no NDCG@10. The paper claims RL provides the improvement over SFT and intent-only baselines, but without an SFT-only NDCG@10 value (and ideally an RL-only ablation) the marginal contribution of RL cannot be isolated. Add paired evaluations of SFT-only vs SFT+RL on the same test queries.
minor comments (4)
  1. [Table 1] Column headers (Sources%, Recency%, Full%) are not defined in the text; please define the benchmark tasks or rename.
  2. [§3.5] Equation (4) introduces weights w_i and a long list of reward coefficients without instantiated values or sensitivity analysis; Figure 1 uses 0.65/0.35, which §3.5 does not state. Please align or explicitly mark as illustrative.
  3. [§5.1] 'lower variance (±0.372)' is ambiguous; report the standard deviation/confidence interval over queries rather than a bare ± symbol.
  4. [§4.5.3/Table 3] Text says 'date exact accuracy', while Table 3 reports start/end date accuracy separately; unify the nomenclature.

Circularity Check

2 steps flagged

The headline NDCG deltas are not independent of the training objective: the reward's content-quality term and the NDCG judgments are both defined by Claude Sonnet 4.6, and the per-router IDCG normalizes away the amount of relevant content retrieved.

specific steps
  1. fitted input called prediction [§3.5 Eq. (3); §4.5.1 Eqs. (9)-(10); §3.4]
    "R_con = α·Rel + (1−α)·Trust (3) ... Both are scored by LLM-as-judge on a normalised scale. ... Relevance judgments are obtained from Claude Sonnet 4.6 acting as a blind judge. ... Claude Sonnet 4.6 was found to be the strongest judge and was used to resolve the remaining 20 percent."

    The RL objective includes a content-quality reward R_con defined as LLM-as-judge relevance (Eq. 3), and the evaluation NDCG is computed from relevance judgments obtained from Claude Sonnet 4.6 (Section 4.5.1). SFT labels are also partly resolved by Claude Sonnet 4.6 (Section 3.4). The model is therefore trained, initialized, and scored with the same oracle. The reported NDCG gains (Abstract, Table 2, Table 4, Section 6) measure alignment with that oracle rather than retrieval quality independent of the training signal, so the central claim's quantitative support is built from its own target.

  2. other [§4.5.1 Eq. (10); Tables 2 and 4]
    "IDCG@10 is computed by sorting the same 10 relevance scores in descending order and applying the same formula, representing the maximum achievable DCG given the retrieved set."

    Because IDCG normalizes by the best ranking of each router's own retrieved set, a router returning only 3 relevant results can score NDCG@10=1.0, while a router returning 8 relevant results with one misplaced item scores lower. The metric therefore compares each router's ranking against its own retrieved content, not how much relevant content each router's agent selection retrieved. All cross-system comparisons in the Abstract, Table 2, and Table 4 rest on this self-normalized denominator, so the reported NDCG deltas do not support the claim that the SLM's selections surface more relevant results.

full rationale

The central derivation chain is: define retrieval relevance via LLM-as-judge (Eq. 3 and Section 4.5.1), train the router to maximize that relevance through RL, and then announce that the same judge scores the router's NDCG higher. That is a partial circularity: the prediction is not an independent test, because the evaluation oracle is the same function being optimized. The flaw is compounded by Eq. (10), where IDCG@10 is computed over each router's own retrieved set; this makes cross-router NDCG comparisons self-normalized and unable to reward retrieving more relevant content. The paper's own conclusion admits that complementary human relevance assessments are needed, which is consistent with this assessment. Latency results and SFT intrinsic metrics (source match, keyword similarity, date accuracy) are independent measurements and are not circular; however, the abstract's headline retrieval-quality claims depend on the contaminated NDCG comparisons. There is no load-bearing self-citation in the reference list, so the circularity is not a self-citation chain but a train-and-evaluate-with-the-same-oracle problem.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on a substantial set of domain assumptions: the judge oracle, the representativeness of the data, the mismatch subset construction, and the baseline choice. The free parameters in the reward function are not reported, so the policy is under-specified.

free parameters (2)
  • Reward coefficients λ_src, λ_con, α, β, γ, α+, α−, ρ_m, δ_miss, w_i = not reported (Fig. 1 shows λ_src=0.65, λ_con=0.35; text calls all tunable)
    Defined in Eqs (2)-(8) and described as tunable hyperparameters with no reported values; they determine the routing policy, so the central results are not reproducible.
  • Keyword semantic-match cosine threshold = 0.6
    Evaluation metric (Sec 4.5.3) hand-sets a 60% cosine similarity threshold to count keyword matches; changing it changes query-intrinsic accuracy.
axioms (4)
  • domain assumption Claude Sonnet 4.6's relevance judgments are valid ground truth for retrieval quality and are stable across training and evaluation.
    Used for SFT labels (§3.4), reward (§3.5), and NDCG judgments (§4.5.1); no human validation or held-out judge is provided.
  • domain assumption The 'targeted subset' of agent-query mismatches can be identified by LLM-as-judge independently of routing outcome and is representative.
    Robustness evaluation (§4.5.2) selects queries where the ground-truth agent, as determined by LLM-as-judge, is mismatched; no sample size or selection rule is reported, so the subset may be cherry-picked.
  • domain assumption Production search logs plus synthetic queries over 11 domains are representative of deployment traffic.
    SFT (56k queries, §4.3.1) and RL (11k queries, §4.4.1) rely on this; no distribution analysis is given.
  • domain assumption Prompted Nova Lite and Claude Haiku with elaborate system prompts are a strong proxy for the class of intent-based routers.
    Baselines (§4.1) are zero-shot prompted LLMs; no learned routing baselines are compared, so 'outperforms intent-based routing' is only established against this specific prompting setup.

pith-pipeline@v1.3.0-alltime-deepseek · 12055 in / 15948 out tokens · 138038 ms · 2026-08-04T01:38:36.124742+00:00 · methodology

0 comments
read the original abstract

Specialised retrieval agents typically surface higher quality results than general-purpose search, but selecting the optimal agent for a given query remains an open problem. Current approaches route queries based on inferred topic or intent, however intent-based selection is fundamentally limited: it does not incorporate signal from retrieved content, and cannot detect when a topically aligned agent produces low-relevance results. We address this by training a small language model via supervised fine-tuning followed by reinforcement learning to jointly perform agent selection and structured parameter generation for downstream tool calls, using a hierarchical reward function grounded in retrieval relevance along with query-agent topic alignment. This enables the model to learn task-dependent agent suitability from retrieval performance: which agents reliably yield high-relevance results for which query distributions, and when to redirect queries away from specialised agents despite surface-level topical overlap. On a targeted subset of such agent-query mismatches, the trained model achieves an NDCG@10 of 0.918 compared to 0.539 and 0.490 for two LLM baselines (Amazon Nova Lite and Claude Haiku 4.5) that route on intent alone. Overall, it achieves a mean NDCG@10 of 0.771 (+0.177 over Nova Lite, +0.219 over Haiku) with a mean selection latency of 120.1ms, an 82.4% reduction over Nova Lite.

Figures

Figures reproduced from arXiv: 2608.00030 by Alexander Ng, Amir Kayhani, Gayathri V Kondapalli, Harvey Yorke, Hirsh Pithadia, Rahul Monish.

Figure 1
Figure 1. Figure 1: Multi-agent query routing pipeline with reinforcement learning feedback over ranked retrieval results [PITH_FULL_IMAGE:figures/full_fig_p004_1.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

29 extracted references · 16 linked inside Pith

  1. [1]

    Avishek Anand, Venktesh V, Abhijit Anand, and Vinay Setty. 2023. Query Un- derstanding in the Age of Large Language Models. arXiv:2306.16004 [cs.IR] https://arxiv.org/abs/2306.16004

    Pith/arXiv arXiv 2023

  2. [2]

    Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin, and Pavlo Molchanov. 2025. Small Lan- guage Models are the Future of Agentic AI. arXiv:2506.02153 [cs.AI] https: //arxiv.org/abs/2506.02153

    Pith/arXiv arXiv 2025

  3. [3]

    Broder, Marcus Fontoura, Evgeniy Gabrilovich, Amruta Joshi, Vanja Josifovski, and Tong Zhang

    Andrei Z. Broder, Marcus Fontoura, Evgeniy Gabrilovich, Amruta Joshi, Vanja Josifovski, and Tong Zhang. 2007. Robust classification of rare queries using web knowledge. InProceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval(Amsterdam, The Nether- lands)(SIGIR ’07). Association for Computing...

    arXiv 2007

  4. [4]

    Huanhuan Cao, Derek Hao Hu, Dou Shen, Daxin Jiang, Jian-Tao Sun, Enhong Chen, and Qiang Yang. 2009. Context-aware query classification. InProceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval(Boston, MA, USA)(SIGIR ’09). Association for Computing Machinery, New York, NY, USA, 3–10. doi:10.1145/1571...

    arXiv 2009

  5. [5]

    Le, Sergey Levine, and Yi Ma

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V. Le, Sergey Levine, and Yi Ma. 2025. SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training.CoRR abs/2501.17161 (2025), 1–14. https://arxiv.org/abs/2501.17161

    Pith/arXiv arXiv 2025

  6. [6]

    Yung-Sung Chuang, Wei Fang, Shang-Wen Li, Wen tau Yih, and James Glass

  7. [7]

    Jeffrey Dalton, Laura Dietz, and James Allan. 2014. Entity query feature expansion using knowledge base links. InProceedings of the 37th International ACM SIGIR Conference on Research & Development in Information Retrieval(Gold Coast, Queensland, Australia)(SIGIR ’14). Association for Computing Machinery, New York, NY, USA, 365–374. doi:10.1145/2600428.2609628

    arXiv 2014

  8. [8]

    Akash Dhasade, Rachid Guerraoui, Anne-Marie Kermarrec, Diana Petrescu, Rafael Pires, Mathis Randl, and Martijn de Vos. 2026. Efficient Feder- ated Search for Retrieval-Augmented Generation using Lightweight Routing. arXiv:2502.19280 [cs.LG] https://arxiv.org/abs/2502.19280

    Pith/arXiv arXiv 2026

  9. [9]

    Dujian Ding, Ankur Mallick, Shaokun Zhang, Chi Wang, Daniel Madrigal, Mirian Del Carmen Hipolito Garcia, Menglin Xia, Laks V. S. Lakshmanan, Qingyun Wu, and Victor Rühle. 2025. BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute. arXiv:2506.22716 [cs.LG] https://arxiv.org/abs/2506.22716

    Pith/arXiv arXiv 2025

  10. [10]

    Pengcheng Jiang, Jiacheng Lin, Lang Cao, Runchu Tian, SeongKu Kang, Zifeng Wang, Jimeng Sun, and Jiawei Han. 2025. DeepRetrieval: Hacking Real Search Engines and Retrievers with Large Language Models via Reinforcement Learning. arXiv:2503.00223 [cs.IR] https://arxiv.org/abs/2503.00223

    Pith/arXiv arXiv 2025

  11. [11]

    Imran Khan. 2025. You Don’t Need Prompt Engineering Anymore: The Prompting Inversion. arXiv:2510.22251 [cs.CL] https://arxiv.org/abs/2510.22251

    arXiv 2025

  12. [12]

    To Eun Kim and Fernando Diaz. 2025. LTRR: Learning To Rank Retrievers for LLMs. arXiv:2506.13743 [cs.CL]

    Pith/arXiv arXiv 2025

  13. [13]

    Pierre Lepagnol, Thomas Gerald, Sahar Ghannay, Christophe Servan, and Sophie Rosset. 2024. Small Language Models Are Good Too: An Empirical Study of Zero-Shot Classification.CoRRabs/2404.11122 (2024), 1–3. https://arxiv.org/abs/ 2404.11122

    Pith/arXiv arXiv 2024

  14. [14]

    Chen Luo, Rahul Goutam, Haiyang Zhang, Chao Zhang, Yangqiu Song, and Bing Yin. 2023. Implicit query parsing at Amazon Product Search. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, New York, NY, USA, 3380–3384. doi:10.1145/3539618. 3591858

    doi:10.1145/3539618 2023

  15. [15]

    Ali Montazeralghaem, Hamed Zamani, and James Allan. 2020. A Reinforcement Learning Framework for Relevance Feedback. InProceedings of the 43rd Interna- tional ACM SIGIR Conference on Research and Development in Information Retrieval (Virtual Event, China)(SIGIR ’20). Association for Computing Machinery, New York, NY, USA, 59–68. doi:10.1145/3397271.3401099

    arXiv 2020

  16. [16]

    Feiteng Mu, Liwen Zhang, Yong Jiang, Wenjie Li, Zhen Zhang, Pengjun Xie, and Fei Huang. 2025. Unsupervised Query Routing for Retrieval Augmented Generation. arXiv:2501.07793 [cs.IR] https://arxiv.org/abs/2501.07793

    Pith/arXiv arXiv 2025

  17. [17]

    Enrico Palumbo, Marcus Isaksson, Alexandre Tamborrino, Maria Movin, Catalin Dincu, Ali Vardasbi, Lev Nikeshkin, Oksana Gorobets, Anders Nyman, Poppy Newdick, Hugues Bouchard, Paul Bennett, Mounia Lalmas, Dani Doro, Christine Doig Cardet, and Ziad Sultan. 2025. You Say Search, I Say Recs: A Scalable Agentic Approach to Query Understanding and Exploratory S...

    arXiv 2025

  18. [18]

    Milad Shokouhi and Luo Si. 2011. Federated Search.Foundations and Trends in Information Retrieval5, 1 (2011), 1–102

    2011

  19. [19]

    Mingyang Song and Mao Zheng. 2026. A Survey of Query Optimization in Large Language Models. arXiv:2412.17558 [cs.CL] https://arxiv.org/abs/2412.17558

    arXiv 2026

  20. [20]

    Shreyas Subramanian, Vikram Elango, and Mecit Gungor. 2025. Small Language Models (SLMs) Can Still Pack a Punch: A survey. arXiv:2501.05465 [cs.CL] https://arxiv.org/abs/2501.05465

    Pith/arXiv arXiv 2025

  21. [21]

    Atsushi Sugiura and Oren Etzioni. 2000. Query routing for Web search engines: architecture and experiments.Comput. Netw.33, 1 (June 2000), 417–429. doi:10. 1016/S1389-1286(00)00059-1

    2000

  22. [22]

    Yongjian Tang, Rakebul Hasan, and Thomas Runkler. 2025. FsPONER: Few-shot Prompt Optimization for Named Entity Recognition in Domain-specific Scenarios. arXiv:2407.08035 [cs.CL] https://arxiv.org/abs/2407.08035

    Pith/arXiv arXiv 2025

  23. [23]

    Jianyou Wang, Kaicheng Wang, Xiaoyue Wang, Weili Cao, Ramamohan Paturi, and Leon Bergen. 2025. IR2: Information Regularization for Information Retrieval. arXiv:2402.16200 [cs.IR] https://arxiv.org/abs/2402.16200

    Pith/arXiv arXiv 2025

  24. [24]

    Shuai Wang, Shengyao Zhuang, Bevan Koopman, and Guido Zuccon. 2025. ReSLLM: Large Language Models are Strong Resource Selectors for Federated Search. InCompanion Proceedings of the ACM on Web Conference 2025(Sydney NSW, Australia)(WWW ’25). Association for Computing Machinery, New York, NY, USA, 1360–1364. doi:10.1145/3701716.3715595

    arXiv 2025

  25. [25]

    Yining Wang, Liwei Wang, Yuanzhi Li, Di He, Tie-Yan Liu, and Wei Chen. 2013. A Theoretical Analysis of NDCG Type Ranking Measures. arXiv:1304.6480 [cs.LG] https://arxiv.org/abs/1304.6480

    Pith/arXiv arXiv 2013

  26. [26]

    Lai Wei, Yuting Li, Kaipeng Zheng, Chen Wang, Yue Wang, Linghe Kong, Lichao Sun, and Weiran Huang. 2025. Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start. arXiv:2505.22334 [cs.CL] https: //arxiv.org/abs/2505.22334

    Pith/arXiv arXiv 2025

  27. [27]

    Yongliang Wu, Yizhou Zhou, Zhou Ziheng, Yingzhe Peng, Xinyu Ye, Xinting Hu, Wenbo Zhu, Lu Qi, Ming-Hsuan Yang, and Xu Yang. 2026. On the Generaliza- tion of SFT: A Reinforcement Learning Perspective with Reward Rectification. arXiv:2508.05629 [cs.LG] https://arxiv.org/abs/2508.05629

    arXiv 2026

  28. [28]

    Xing, Hao Zhang, Joseph E

    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. arXiv:2306.05685 [cs.CL] https://arxiv.org/abs/2306.05685 Received 15 April 2026

    Pith/arXiv arXiv 2023

  29. [2023]

    arXiv:2305.17080 [cs.CL] https://arxiv.org/abs/2305.17080

    Expand, Rerank, and Retrieve: Query Reranking for Open-Domain Question Answering. arXiv:2305.17080 [cs.CL] https://arxiv.org/abs/2305.17080

    Pith/arXiv arXiv