Pith. sign in

REVIEW 3 major objections 4 minor 50 references

The paper argues that decoupling ad insertion from generation—using a lightweight sidecar rewriter on the finished answer—is a strictly better insertion paradigm than prompting, sampling, or tuning a single model, with consistent gains acro

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

2026-08-01 02:00 UTC pith:FRPWYGL2

load-bearing objection New formulation, weak validation: the decoupled sidecar idea is worth engaging, but the current evidence is a single-family judge with no human checks. the 3 major comments →

arxiv 2607.25590 v1 pith:FRPWYGL2 submitted 2026-07-28 cs.CL

PILA: Plug-and-Play Insertion for LLM-native Advertising

Zhaowei Zhang , Yuhan Fu , Yihang Zhang , Xiaohan Liu , Ceyao Zhang , Xiaoyuan Zhang , Yipeng Kang , Tonghan Wang
show 1 more author
Yaodong Yang
This is my paper
classification cs.CL
keywords LLM-native advertisingresponse rewritingsidecar moduleplug-and-playcontrastive decodingad intensity controlpersuasion knowledge modelmodel-agnostic
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.

PILA reframes LLM-native advertising as a conditional response-rewriting problem: instead of asking the model that answers the user to also generate sponsored content, a small sidecar model rewrites the already-finished answer to weave in an ad. The paper claims this decoupling improves both user-side response quality and ad-side effectiveness, reporting that 4B and 8B PILA variants raise scores of seven frontier commercial models by 17.2% and 18.4% on average, and beat single-model prompt, sampling, and tuning baselines by 34.2%, 47.3%, and 7.7% respectively. It also introduces an inference-time ad-intensity knob that trades naturalness against ad salience without retraining. A sympathetic reader would care because, if correct, this offers a practical, model-agnostic route to monetizing LLM APIs and agentic workflows without touching the base model or its pipeline.

Core claim

The central discovery is that post-hoc rewriting outperforms entangled generation: a rewriter trained to insert sponsored spans into a completed, ad-free answer yields higher user-side relevance/coherence and higher ad effectiveness than having one model generate both at once. In the paper's experiments, PILA-4B and PILA-8B consistently improve every score across every evaluated upstream model and ad domain, with the 8B variant improving the seven commercial upstream models by 18.4% on average; against single-LLM baselines it improves over SFT, Base, and MOSAIC by 7.7%, 34.2%, and 47.3%. The authors attribute this to decoupling: the upstream system remains responsible for serving the user, w

What carries the argument

The load-bearing object is the sidecar rewriter πθ, a lightweight fine-tuned model (Qwen3-4B/8B) that takes the user query, the upstream ad-free response, and the ad (name + copy) and outputs a rewritten response with the sponsored span marked by <ad>...</ad> tags. Carrying the argument is the training corpus of roughly 25k (query, response, ad, rewritten-response) instances created by synthesis followed by diversity augmentation, plus an inference-time intensity controller defined by contrastive decoding: log πρ = log π_PILA + ρ log(π_PILA/π_ref), where ρ is a deployment-time knob grounded in the persuasion-knowledge model that shifts between subtle and salient insertion.

Load-bearing premise

The evaluator scores come from an LLM judge, not from real users, and the paper assumes those scores reflect how humans actually experience the responses.

What would settle it

Take a sample of the benchmark pairs where PILA beats Base, collect human ratings of relevance, coherence, and ad effectiveness on the same 1–5 scale, and compare human preference against the LLM judge's scores. If humans do not prefer PILA's rewrites at near the same rate—or prefer Base in a nontrivial share of cases—the central claim that decoupled rewriting improves user-side quality and ad effectiveness would fail its load-bearing validation.

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

If this is right

  • Ad insertion becomes a drop-in module for any LLM service, including API-only and workflow/agentic systems, since it only needs the final text response.
  • PILA-8B improves seven commercial upstream models by 18.4% on average, with the largest gains on Gemini 3.1 Flash (33.2%) and Deepseek V3.2 (27.0%) — indicating that stronger rewriting helps where native ad generation is weakest.
  • The intensity factor ρ gives operators a deployment-time trade-off knob: increasing ρ monotonically lowers user-side scores while first raising then lowering ad-side scores, so a deployment can be tuned to its pricing or UX constraints without retraining.
  • The paradigm suggests a two-stage architecture is preferable to single-model end-to-end generation for ad-insertion quality, as the decoupled method Pareto-dominates the base prompt-based approach on user- and ad-side scores.
  • Because the rewriter is trained on the same data used to build the SFT baseline, the gains are attributed to the rewriting formulation itself rather than to more or better training data.

Where Pith is reading between the lines

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

  • The reported gains may partly reflect the taste of the LLM judge: the training corpus was synthesized and filtered by the same model family used for evaluation, so a human-rating study could shrink the margins. This is an open question the paper does not settle.
  • Because insertion is separated from generation, the sidecar layer is a natural place for future auction mechanisms — advertisers could bid on bridging strategies or insertion positions per response, without touching the upstream model.
  • The same rewrite-layer architecture could generalize beyond ads to other conditional insertions (disclaimers, citations, content warnings, personalized callouts), making the paper's contribution a template for post-hoc text augmentation.
  • A concrete testable extension: vary ρ per user or per query context to see whether the monotone user-side decline can be partially avoided, which the current aggregate curves do not rule out.

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

3 major / 4 minor

Summary. The paper introduces PILA, a sidecar rewriter that inserts advertisements into an already completed upstream LLM response rather than generating the response and ad jointly. The method is model-agnostic, uses a 25k synthetic training corpus built on NaiAD with Claude Opus 4.5/Haiku 4.5 filtering, fine-tunes Qwen3-4B/8B as rewriters, and exposes an ad-intensity knob via contrastive decoding (Eq. 3). Against prompt-, sampling-, and tuning-based single-LLM baselines and across seven commercial upstream models, the paper reports large average gains (e.g., PILA-8B improving on Base by 34.2% in Table 1 and by 18.4% on upstream models in Table 2) while preserving or improving response quality.

Significance. The architectural proposal is timely and practically motivated: decoupling ad insertion as post-hoc rewriting is compatible with API-only and workflow-based systems, requires no access to upstream parameters, and the ρ-controlled contrastive decoding in Eq. (3) is a clean, parameter-free mechanism for navigating the user/ad trade-off. The paper is also transparent about its data construction pipeline and evaluates over diverse commercial models and categories. If the empirical gains survive proper evaluation, this would be a solid contribution. The currently missing piece is validation that the single automated judge (Claude Opus 4.6) yields conclusions that transfer to human-perceived quality.

major comments (3)
  1. [Experiments, Benchmark and Evaluation; Case Study] All headline numbers are point estimates from a single LLM judge, Claude Opus 4.6. The only validation offered is the qualitative case study in the Case Study section, which is not human evaluation. The training corpus was synthesized and quality-filtered with Claude Opus 4.5/Haiku 4.5, so PILA is optimized to imitate a Claude-style insertion; the same-family judge may systematically prefer that style. Because Q1/Q2/Q3 are the sole dependent measure, the central claim rests on an unvalidated metric. I would want either a human rating study (even a few hundred examples) or at least a second judge from a different model family plus an agreement analysis.
  2. [Table 2; Experiments, Results on Improving Upstream Models] No error bars, significance tests, or seed variance are reported for any table. The claim that PILA 'consistently improve[s] all scores across all frontier upstream models and ad topics' is contradicted by the table itself: PILA-4B lowers Q1 on GPT-5.4 Food & Beverage (3.26→3.11) and Home (3.38→3.15), and lowers Q2 on Claude Haiku 4.5 Home (2.96→2.87); PILA-8B lowers Q3 on Gemini 3.1 Pro Home (3.93→3.77) and Gemini 3.1 Flash Travel (3.67→3.60). At minimum, the authors should report bootstrap or other uncertainty intervals and soften the consistency claim to what the data support.
  3. [Experiments, Baseline Methods; Table 1] The SFT baseline is trained on 'the high-quality responses from the same data used for PILA' but deployed as a single-model generator rather than as a rewriter. It is unclear whether the SFT model receives the same 25k instances in the same format, whether it uses the same LoRA recipe and training epochs, and whether the comparison controls for training compute. The reported 7.7% average advantage over SFT is load-bearing for the decoupling thesis; if the SFT baseline is undertrained or format-mismatched, the comparison is not apples-to-apples.
minor comments (4)
  1. [Experiments, Results on Improving Upstream Models] The percentages 17.2%/18.4% (Table 2) and 24.9%/20.0%/8.1%/17.2% are stated without saying whether they cover the four main categories only or all six including the appendix Table 5. Please clarify.
  2. [Experiments, Results on Ad Intensity Control] Figure 4 uses a descending x-axis, which is unintuitive; please annotate the direction of increasing ρ and label the ρ values on the curve.
  3. [Case Study] The sentence 'This comparison also serves to validate that our automated evaluation metrics... accurately reflect human-perceived quality' overstates what two qualitative examples can show. Rephrase as illustrative rather than validating.
  4. [PILA Training Details] No statement of code/data release, random seeds, or evaluation prompt details is provided. Please include these for reproducibility.

Circularity Check

2 steps flagged

Evaluation loop is closed: training labels and final scores both come from Claude-family judges, and the benchmark/evaluator are inherited from the authors' own NaiAD; no human validation anchors the metric.

specific steps
  1. other [Methodology: Data Construction (Figure 2); Experiments: Benchmark and Evaluation]
    "Following the NaiAD protocol, the model scores the responses and we retain only high-quality candidates. ... Following NaiAD, we use Claude Opus 4.6 (Anthropic 2026b) as the evaluator backbone and retain the prompts and scoring settings."

    The supervised training targets are synthesized and filtered by Claude Opus 4.5/Haiku 4.5 (the seed corpus accepts examples that 'lose no more than 2 points' under Claude's self-judging). The final quality scores (Q1-Q3) are then assigned by Claude Opus 4.6 using NaiAD prompts. PILA is therefore fine-tuned to imitate text that Claude-family judges rate highly, and the reported gains are measured with the same family of judges. The 'improvement' reduces to agreement with the judge family that defined the training labels, rather than to an independently anchored user-side or ad-side utility; no human ratings are reported.

  2. self citation load bearing [Experiments: Benchmark and Evaluation; Related Works; Data Construction]
    "NaiAD (Zhang et al. 2026) is currently the most comprehensive LLM-native advertising dataset... We build our evaluation benchmark from its six largest advertising categories... Following NaiAD, we use Claude Opus 4.6 as the evaluator backbone and retain the prompts and scoring settings."

    The central empirical claim (PILA improves seven upstream models) is measured exclusively on a benchmark whose categories, rendering data, prompts, and evaluator choice are taken from NaiAD, a prior paper with overlapping authors (Zhang, Kang, Wang). The paper supplies no external validation that NaiAD's Claude-based scores correspond to human-perceived quality; the only check offered is a qualitative case study that asserts the metrics 'accurately reflect human-perceived quality' without collecting ratings. Thus the load-bearing evidence for the headline claim is a self-citation chain, not an independent benchmark.

full rationale

There is no equation-level circularity: Eq. (3) is a standard contrastive-decoding combination of a tuned and untuned policy with no fitted constants, and the ρ sweep is a reported trade-off measurement rather than a prediction. The circularity burden is in the evaluation loop. The training corpus is built by Claude Opus 4.5 and Haiku 4.5 and filtered by self-judging; the final evaluation is performed by Claude Opus 4.6 under NaiAD prompts; and NaiAD itself is the authors' own prior work. Consequently the central claim that PILA 'consistently improves ad effectiveness while preserving response quality' is supported only by scores from the same model family that produced and filtered the training data, with no human ratings or external metric to anchor the definition of 'quality.' The case study's assertion that the automated metrics 'accurately reflect human-perceived quality' is not human validation; it is a qualitative example selected to illustrate the metrics. The self-citation is load-bearing because every quantitative headline (e.g., 17.2%/18.4% average improvements in Table 2) is computed inside this NaiAD/Claude loop. I do not claim the method is a pure fit: the baseline comparisons are real and the rewriting outputs are not constructed to equal the judge's scores. But the measured advantage is partly an artifact of training and evaluating within one judge family, which is a partial circularity. The claim in the paper that PILA 'consistently improve[s] all scores across all frontier upstream models and ad topics' is also not supported by Table 2 itself (e.g., PILA-4B lowers Q1 on GPT-5.4 Food & Beverage from 3.26 to 3.11; PILA-8B lowers Q3 on Gemini 3.1 Pro Home from 3.93 to 3.77), which further weakens the reliability of the aggregate gains. Overall score: 5 — partial circularity via self-referential evaluation and a load-bearing self-citation, but not a derivation that reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The empirical claims rest on hand-set data-construction choices (λ=0.15, pass threshold ≤2 points, 3 paraphrases per seed) and an inference-time intensity knob ρ; on domain assumptions about LLM-as-a-judge fidelity, PKM's predicted quality trade-off, NaiAD's representativeness, and rewrite faithfulness; and on standard contrastive-decoding math in Eq. (3). No new physical or formal entities are introduced; the sidecar module and <ad> tagging protocol are software artifacts whose only evidence is the paper's own benchmark scores.

free parameters (4)
  • ad-intensity factor ρ (inference-time knob) = evaluated at {0,1,2,3,4}
    Exposed as a deployment-time control in Eq. (3), not fitted to data; the paper reports user-side score decreasing monotonically with ρ. It is a hand-chosen interface, not a learned constant, but it is a free control variable in the reported intensity experiments.
  • diversity penalty λ for ad retrieval = 0.15
    Data Construction: sentence-embedding retrieval pairs each query with the top-4 ads under diversity penalty λ=0.15. Hand-set; directly shapes the training corpus composition.
  • quality pass threshold for seed acceptance = loss of ≤ 2 points from total score
    Data Construction: 'each example is required to lose no more than 2 points from the total score to pass,' yielding 10,014 accepted seeds from 24,883 attempts. A hand-set quality gate determining what the rewriter learns.
  • number of paraphrases per seed (augmentation) = 3
    Data Construction: each seed generates three paraphrase variants, yielding ~15k augmented examples. Hand-set multiplicity that defines the final 25k training corpus.
axioms (5)
  • domain assumption LLM-as-a-judge scores (Claude Opus 4.6) faithfully represent user-side response quality and ad effectiveness
    Invoked in Experiments, 'Benchmark and Evaluation'; the entire empirical comparison rests on this, with no systematic human calibration and no error-bar or agreement analysis.
  • domain assumption Persuasion Knowledge Model (PKM) correctly predicts that increasing ad salience monotonically lowers user satisfaction
    Used in 'Controlling Ad Intensity' to justify the ρ knob; the expected monotone trade-off is assumed to follow PKM and is only checked on a 10% sample without variance.
  • standard math Contrastive decoding combination log π_ρ = log π_PILA + ρ log(π_PILA/π_ref) yields a valid decoding distribution over the same support
    Eq. (3): standard log-linear interpolation of two distributions (Li et al. 2023; Mitchell et al. 2023). No new mathematics; normalization and support assumptions are implicit.
  • domain assumption The NaiAD corpus and evaluation protocol are representative of real LLM-native advertising scenarios
    The benchmark is built from 'NaiAD's six largest advertising categories' and 'retain[s] the prompts and scoring settings' of NaiAD (Zhang et al. 2026), co-authored by two of PILA's corresponding authors; its public availability is not disclosed.
  • domain assumption Rewriting a completed answer with an <ad>-tagged span preserves the facts and utility of the original response
    The PILA prompt instructs 'preserve all facts, formatting, and overall quality,' but no faithfulness metric checks whether rewriting altered the upstream answer's content; the no-ad response y is never scored.

pith-pipeline@v1.3.0-alltime-deepseek · 20631 in / 21965 out tokens · 193750 ms · 2026-08-01T02:00:00.337447+00:00 · methodology

0 comments
read the original abstract

How to monetize large language models (LLMs) by naturally integrating sponsored content into their responses, known as LLM-native advertising, has recently emerged as a critical problem. However, existing solutions entangle advertising with content generation inside a single model, which is incompatible with modern API-only or workflow-based LLM applications and inevitably compromises the original response quality. To address this, we propose PILA, which reformulates ad insertion as a conditional response rewriting problem and decouples it from the upstream service as a lightweight sidecar module. PILA is model-agnostic and can be seamlessly integrated with existing LLM services without modifying the base model or its workflow. It also exposes a controllable trade-off between user-side naturalness and ad-side exposure, offering a practical interface for downstream pricing and deployment. Experiments across diverse upstream models show that \pila consistently improves ad effectiveness while preserving response quality, highlighting its promise as a practical solution for LLM-native advertising.

Figures

Figures reproduced from arXiv: 2607.25590 by Ceyao Zhang, Tonghan Wang, Xiaohan Liu, Xiaoyuan Zhang, Yaodong Yang, Yihang Zhang, Yipeng Kang, Yuhan Fu, Zhaowei Zhang.

Figure 1
Figure 1. Figure 1: An overview of our PILA framework. Given a user query along with an ad name and ad content, existing baselines either rely on a single LLM with prompting, sampling, or tuning strategies, or follow predefined workflows that often produce poor cases while losing flexibility. In contrast, PILA acts as a lightweight plug-and-play module that adapts diverse upstream models to generate high-quality, naturally in… view at source ↗
Figure 2
Figure 2. Figure 2: The data construction process of the PILA framework. Starting from NaiAD data covering various ad classes, Claude Opus 4.5 generates initial paired responses (with and without ads) for each query–ad tuple. A quality judge then performs self-judging and score-based filtering to retain a 10k qualified seed set. Each seed is further expanded through diverse data augmentation, where Claude Haiku 4.5 produces t… view at source ↗
Figure 3
Figure 3. Figure 3: Performance of PILA against base commercial [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results of ad intensity control. The x- and y-axes [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

50 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:2512.10551 , year=

    LLM-Auction: Generative Auction towards LLM-Native Advertising , author=. arXiv preprint arXiv:2512.10551 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2510.11438 , year=

    What generative search engines like and how to optimize web content cooperatively , author=. arXiv preprint arXiv:2510.11438 , year=

    arXiv

  3. [3]

    arXiv preprint arXiv:2512.02556 , year=

    Deepseek-v3.2: Pushing the frontier of open large language models , author=. arXiv preprint arXiv:2512.02556 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2306.03314 , year=

    Multi-agent collaboration: Harnessing the power of intelligent llm agents , author=. arXiv preprint arXiv:2306.03314 , year=

    Pith/arXiv arXiv

  5. [5]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Mechanism design for llm fine-tuning with multiple reward models , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  8. [8]

    arXiv preprint arXiv:2605.09918 , year=

    NaiAD: Initiate Data-Driven Research for LLM Advertising , author=. arXiv preprint arXiv:2605.09918 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2311.07601 , year=

    Online advertisements with llms: Opportunities and challenges , author=. arXiv preprint arXiv:2311.07601 , year=

    Pith/arXiv arXiv

  10. [10]

    Proceedings of the ACM Web Conference 2024 , pages=

    Mechanism design for large language models , author=. Proceedings of the ACM Web Conference 2024 , pages=

    2024

  11. [11]

    Chen, Zhaohua and Yang, Mingwei and Wang, Chang and Li, Jicheng and Cai, Zheng and Ren, Yukun and Zhu, Zhihua and Deng, Xiaotie , month = may, year =. Budget-. Proceedings of the. doi:10.1145/3589334.3645344 , language =

    arXiv

  12. [12]

    Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pages=

    Geo: Generative engine optimization , author=. Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining , pages=

  13. [13]

    Auctions with

    Dubey, Kumar Avinava and Feng, Zhe and Kidambi, Rahul and Mehta, Aranyak and Wang, Di , month = apr, year =. Auctions with. doi:10.48550/arXiv.2404.08126 , language =

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.08126

  14. [14]

    Hajiaghayi, MohammadTaghi and Lahaie, Sébastien and Rezaei, Keivan and Shin, Suho , year =. Ad. Advances in. doi:10.52202/079017-0585 , language =

    work page doi:10.52202/079017-0585

  15. [15]

    Proceedings of the ACM Web Conference 2026 , pages=

    Position auctions in ai-generated content , author=. Proceedings of the ACM Web Conference 2026 , pages=

    2026

  16. [16]

    Mathematics of operations research , volume=

    Optimal auction design , author=. Mathematics of operations research , volume=. 1981 , publisher=

    1981

  17. [17]

    International Conference on Artificial General Intelligence , pages=

    Roadmap on incentive compatibility for ai alignment and governance in sociotechnical systems , author=. International Conference on Artificial General Intelligence , pages=. 2025 , organization=

    2025

  18. [18]

    arXiv preprint arXiv:2604.06263 , year=

    Incentive-Aware Multi-Fidelity Optimization for Generative Advertising in Large Language Models , author=. arXiv preprint arXiv:2604.06263 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2605.08326 , year=

    LLM Advertisement based on Neuron Auctions , author=. arXiv preprint arXiv:2605.08326 , year=

    Pith/arXiv arXiv

  20. [20]

    Sponsored

    Mordo, Tommy and Tennenholtz, Moshe and Kurland, Oren , month = aug, year =. Sponsored. Proceedings of the 2024. doi:10.1145/3664190.3672517 , language =

    arXiv 2024

  21. [21]

    and Seuken, Sven , month = feb, year =

    Soumalias, Ermis and Curry, Michael J. and Seuken, Sven , month = feb, year =. Truthful. doi:10.48550/arXiv.2405.05905 , language =

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.05905

  22. [22]

    arXiv preprint arXiv:2601.19435 , year=

    Ad Insertion in LLM-Generated Responses , author=. arXiv preprint arXiv:2601.19435 , year=

    arXiv

  23. [23]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  24. [24]

    2026 , month =

    Spivack, Jameson , title =. 2026 , month =

    2026

  25. [25]

    2026 , note =

    OpenAI , title =. 2026 , note =

    2026

  26. [26]

    2026 , note =

    Anthropic , title =. 2026 , note =

    2026

  27. [27]

    2025 , note =

    Anthropic , title =. 2025 , note =

    2025

  28. [28]

    2025 , url =

    OpenAI Developers , title =. 2025 , url =

    2025

  29. [29]

    2025 , url =

    Anthropic , title =. 2025 , url =

    2025

  30. [30]

    2026 , note =

    Google DeepMind , title =. 2026 , note =

    2026

  31. [31]

    2026 , url =

    Google DeepMind , title =. 2026 , url =

    2026

  32. [32]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=

  33. [33]

    Forty-first International Conference on Machine Learning , year=

    Gptswarm: Language agents as optimizable graphs , author=. Forty-first International Conference on Machine Learning , year=

  34. [34]

    International Conference on Learning Representations , volume=

    Aflow: Automating agentic workflow generation , author=. International Conference on Learning Representations , volume=

  35. [35]

    International Conference on Learning Representations , volume=

    Amulet: Realignment during test time for personalized preference adaptation of LLMs , author=. International Conference on Learning Representations , volume=

  36. [36]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Contrastive decoding: Open-ended text generation as optimization , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  37. [37]

    arXiv preprint arXiv:2310.12962 , year=

    An emulator for fine-tuning large language models using small language models , author=. arXiv preprint arXiv:2310.12962 , year=

    Pith/arXiv arXiv

  38. [38]

    8th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud 16) , year=

    Design patterns for container-based distributed systems , author=. 8th USENIX Workshop on Hot Topics in Cloud Computing (HotCloud 16) , year=

  39. [39]

    Journal of consumer research , volume=

    The persuasion knowledge model: How people cope with persuasion attempts , author=. Journal of consumer research , volume=. 1994 , publisher=

    1994

  40. [40]

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , author=

  41. [41]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  42. [42]

    2025 , url =

    Claude Code documentation , author =. 2025 , url =

    2025

  43. [43]

    2026 , url =

    Harness design for long-running application development , author =. 2026 , url =

    2026

  44. [44]

    2026 , url =

    Harness engineering: leveraging Codex in an agent-first world , author =. 2026 , url =

    2026

  45. [45]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  46. [46]

    Procedia computer science , volume=

    A Survey on RAG with LLMs , author=. Procedia computer science , volume=. 2024 , publisher=

    2024

  47. [47]

    2025 , url =

    Codex documentation , author =. 2025 , url =

    2025

  48. [48]

    2026 , url =

    OpenClaw documentation , author =. 2026 , url =

    2026

  49. [49]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 3: system demonstrations) , pages=

    Llamafactory: Unified efficient fine-tuning of 100+ language models , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 3: system demonstrations) , pages=

  50. [50]

    Hashimoto , title =

    Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto , title =. GitHub repository , howpublished =. 2023 , publisher =

    2023