Pith. sign in

REVIEW 4 major objections 5 minor 11 references

Role-playing agents fail hardest on immoral characters, and a training-free decoding tweak recovers much of the gap.

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

2026-08-03 11:54 UTC pith:SDKJFQNT

load-bearing objection The disposition bottleneck is real, but the PersonaGym half of the evidence is a measurement artifact — the CoSER results and the new controlled dataset carry the paper. the 4 major comments →

arxiv 2601.04716 v3 pith:SDKJFQNT submitted 2026-01-08 cs.CL

Identifying and Mitigating Bottlenecks in Role-Playing Agents: A Systematic Study of Disentangling Character Profile Axes

Yonghyun Jun , Junhyuk Choi , Jeonghyun Park , Jihyeong Park , Liu Nicole Geumheon , Hwanhee Lee This is my paper
classification cs.CL
keywords role-playing agentscharacter profilemoral dispositionLLM alignmentsafety alignmentcontrastive decodingpersona evaluationtoxicity control
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 isolate which parts of a character profile actually determine how well an LLM role-playing agent stays in character. It builds a controlled set of 211 personas that vary along three axes — whether the character is known or unknown to the model, whether the profile is structured or free-form, and whether the character is moral or immoral — and evaluates five frontier LLMs on single-turn and multi-turn interactions. The central result is an asymmetry: the first two axes move performance little, while moral disposition moves it a lot. Immoral characters consistently score lower across every model and both benchmarks, with the drop concentrated in goals and morality rather than surface personality traits. The paper argues this is an alignment artifact — safety training suppresses the tokens needed to portray villains — and offers a training-free contrastive decoding method (FACD) that narrows the gap by selectively amplifying those suppressed signals.

Core claim

The paper's central empirical claim is that a character's moral disposition, not familiarity or profile format, is the dominant bottleneck in LLM role-playing. Across all five evaluated models and both interaction settings, immoral characters show a large, significant, and directionally consistent drop — about 0.29 to 0.40 points on the single-turn metric and 5.89 to 9.22 on the multi-turn metric — while known-vs-unknown and structured-vs-unstructured comparisons are small or inconsistent. The degradation concentrates in motivation-related fields such as Goal and Morality; personality traits such as Extraversion stay stable. The authors read this as evidence that safety alignment suppresses

What carries the argument

The load-bearing pieces are (1) a standardized hierarchical profile schema — 5 top-level dimensions and 28 leaf fields — that lets the paper vary one axis while holding the rest fixed; (2) a paired dataset of 211 personas (422 profiles including unstructured variants) with LLM- and human-validated moral/immoral labels; and (3) Field-Aware Contrastive Decoding (FACD), a training-free decoding intervention. FACD computes next-token logits under the full profile and under a 'negative' profile where immoral fields are removed (leaving only moral fields plus a small set of polarity-insensitive traits), then mixes them as z_pos + α(z_pos − z_neg) with α=1, so the decoder amplifies exactly those to

Load-bearing premise

The diagnosis assumes the evaluation measures role-playing fidelity independently of moral valence; the paper's own per-metric tables show the largest single moral–immoral gap is on a toxicity-control rubric (1.00–1.88), so if that rubric scores 'safe' rather than 'in character', the headline belief that disposition is the bottleneck is partly a scoring artifact.

What would settle it

Recompute the moral–immoral gap after removing the toxicity-control rubric (and any content-policy metric) from the single-turn benchmark, and run a second evaluation with human raters instructed to rate profile adherence while ignoring whether the character is good or evil. If the gap collapses on those runs, the central claim of an alignment-induced capability bottleneck fails; if it stays large and direction-consistent, the claim survives.

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

If this is right

  • Profile builders can stop treating format and fame as quality levers; structured, unstructured, known, and unknown profiles perform comparably, so effort can go into content rather than template.
  • Villainous or morally ambiguous characters will systematically underperform, with the biggest losses in goals, morality, and worldview; systems that need them require an explicit mitigation step or a lower fidelity expectation.
  • If the alignment-suppression explanation is right, the measured gap should grow as safety alignment grows and shrink for less-aligned models — a testable prediction across model families.
  • Benchmarks with a toxicity or safety rubric will score immoral personas down on that rubric by construction; reporting safety metrics separately from fidelity metrics would keep a measurement artifact from being read as an incapability.

Where Pith is reading between the lines

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

  • The same field-aware contrast recipe could plausibly amplify other suppressed signals beyond immorality — for instance, negative emotion, politically charged speech, or niche identity markers — since the mechanism is token suppression by alignment, not morality per se.
  • A sharper causal test of the paper's explanation would measure per-token logit suppression before and after alignment on a fixed profile: if goal/morality vocabulary is suppressed while personality-trait vocabulary is not, the localization story is confirmed at the mechanistic level.
  • The binary moral/immoral split leaves anti-heroes and morally ambiguous characters untested; an obvious extension is a continuous disposition scale to check whether the performance gap tracks the degree of immorality linearly or is thresholded.
  • Deploying FACD costs two forward passes per token, so for long multi-character conversations the efficiency cost may be nontrivial; distillation or a single-pass approximation would be a natural next step.

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 / 5 minor

Summary. The paper introduces a diagnostic framework for role-playing agents (RPAs) that disentangles three character-profile axes—Familiarity, Structure, and Disposition—using a controlled dataset of 211 personas built on a 28-field hierarchical schema. Five LLMs are evaluated on PersonaGym (single-turn) and CoSER (multi-turn). The central empirical claim is that Disposition is the dominant, consistent bottleneck: immoral characters incur large performance gaps on both benchmarks (PersonaGym Δ=0.29–0.40; CoSER Δ=5.89–9.22, all p<0.001), while Familiarity and Structure have negligible or inconsistent effects. The paper further proposes Field-Aware Contrastive Decoding (FACD), a training-free contrastive decoding variant, and reports that it narrows the Moral–Immoral gap on Qwen3-8B and Mistral-Small without hurting moral-character performance.

Significance. If the headline result holds, the paper would be a valuable contribution: it provides a controlled, reusable profile schema, a field-level diagnosis of where role-playing fidelity breaks down, and a simple training-free mitigation that shows large Character Fidelity recovery in at least one model. The strengths include the systematic dataset construction with human validation of disposition labels, the broad five-model evaluation, and the instructive finding that Motivations degrade more than Personality Traits under immoral settings. However, the central claim is currently over-stated. On PersonaGym, the disposition gap is almost entirely re-expressed from the Toxicity Control rubric, which directly penalizes the behavior under study. The CoSER result is more credible but relies on a safety-aligned LLM judge, and the human validation is too small and thinly documented to fully rule out evaluator bias. The inconsistency between Tables 3 and 20 for DeepSeek-v3.2 on CoSER also must be resolved. These issues are load-bearing for the bottleneck claim as written, but they are addressable within the manuscript's scope.

major comments (4)
  1. [§4.2, Table 20] The PersonaGym half of the headline result is largely a rubric artifact. PersonaScore is the unweighted mean of five metrics, and Table 20 shows the Moral–Immoral gap in Toxicity Control is Δ=-1.00 to -1.88 for every model, while the mean of the other four metrics is at most -0.26 in magnitude (e.g., +0.03 for Qwen3-8B and -0.02 for DeepSeek-v3.2 after averaging Expected Action, Linguistic Habits, Persona Consistency, and Action Justification). Since Toxicity Control explicitly penalizes toxic output, a model that faithfully portrays an immoral character is scored low by construction. The claim in §4.2 that Disposition produces 'large, consistent performance degradation' is therefore not cleanly supported on PersonaGym; the single-turn evidence reduces to a measurement confound unless the authors report a version of PersonaScore excluding Toxicity or use a moral-valence-neutral rubric.
  2. [§4.2, Appendix B, Appendix C.2] The CoSER result is better evidence but still vulnerable to judge bias: the critic is GPT-4o, a safety-aligned model, and the paper provides no test showing that it rates character fidelity independently of moral valence. The human validation is too thin to resolve this: C.2 samples 20% of characters, each item is scored by one of three annotators in non-overlapping groups, no inter-annotator reliability is reported, and the appendix does not state whether annotators were instructed to disregard moral valence. I recommend re-judging a subset with a rubric that explicitly excludes safety/toxicity considerations, and/or collecting human annotations with agreed moral-valence-agnostic criteria. The FACD results partly mitigate the concern—steering toward immoral content would likely lower scores under a strongly moralizing judge—but they do not eliminate it.
  3. [Table 3 vs. Table 20] There is an unresolved numerical conflict in the central result. Table 3 reports DeepSeek-v3.2 CoSER Moral 44.82, Immoral 35.60, Δ=-9.22; Table 20's per-metric averages are 42.82, 37.60, Δ=-5.22. These cannot both be correct, and the abstract and §4.2 use the -9.22 value to characterize the gap range (Δ 5.89–9.22, all p<0.001). The authors must correct the inconsistency and re-run the significance tests; this is load-bearing for the magnitude claim.
  4. [§5.3, Table 5] The mitigation claim rests on only two backbone models, with no significance tests or confidence intervals for the gap reduction, and the human validation table contains an arithmetic error: the FACD row for Qwen3-8B lists Moral 3.56, Immoral 3.25, but Avg. 3.64, whereas the mean is 3.405. The paper should scope the mitigation claim to the two evaluated models, report variability, and verify that the large Character-Fidelity reversal on Qwen3-8B is stable across seeds, turns, or repeated scenario draws.
minor comments (5)
  1. [Abstract / Figure 1] The abstract alternates between 'Disposition' and 'Valence'; Figure 1(C) also labels 'Moral Valence.' Please use one consistent term.
  2. [§5.3] Typo: 'pplying' should be 'Applying.' Also in Figure 1, 'Godam' should be 'Gotham.'
  3. [Appendix A / Appendix B] The Scenario Generation Prompt appears twice (Table 7 and Table 15) with identical content; keep only one copy.
  4. [Appendix C.2] Human evaluation tables report only means. Please add item counts, standard deviations, and a statement of whether each item received one or multiple annotations; the current description ('three non-overlapping groups') suggests single annotation per item, which limits interpretability.
  5. [§5.2, Eq. (2)] The field classifier M is described as MoralBERT, but the paper does not specify the score threshold used to label a field 'moral' versus 'immoral,' nor the exact ordering of the Fpad fallback fields. Please make both explicit for reproducibility.

Circularity Check

2 steps flagged

PersonaGym's disposition gap is arithmetically dominated by the Toxicity Control rubric, and FACD's Character Fidelity gain re-implements what the metric measures; CoSER provides partial independent support.

specific steps
  1. self definitional [Section 4.1 (PersonaGym setup) + Table 20 (per-metric disposition deltas)]
    "PersonaGym evaluates response quality via dynamic, persona-tailored interviews using five metrics: Persona Consistency, Linguistic Habits, Expected Action, Action Justification, and Toxicity Control. The final Persona Score is derived by averaging these metrics. Table 20 (Qwen3-8B): Expected Action Δ=-0.01, Toxicity Δ=-1.88, Linguistic Habits +0.06, Persona Consistency +0.13, Action Justification -0.07 → PersonaScore Δ=-0.36."

    The headline 'Disposition produces large, consistent performance degradation' on PersonaGym is not an independent behavioral measurement. Because PersonaScore is the unweighted mean of five rubric scores, the reported 0.29–0.40 gaps are by arithmetic the mean of the per-rubric deltas, and the Toxicity Control rubric contributes an outsized negative delta (up to -1.88) while the other four deltas are near zero, mixed, or much smaller. Toxicity Control by definition penalizes toxic content, so a faithful immoral persona is scored low by the rubric itself. The aggregate disposition gap therefore partly restates the rubric's toxicity penalty rather than demonstrating a model capability bottleneck.

  2. self definitional [Section 5.2, Eq. (2) and Section 5.3 (Mitigation Results), with CoSER metric definition in Appendix B.1]
    "The final negative prompt is: sneg = s(FPA ∪ Fmoral ∪ Fpad) ... In our setup, Immoral fields are absent from sneg ... α selectively amplifies only the tokens consistent with the omitted immoral fields. ... The most striking improvement appears in Character Fidelity, which is the dimension most directly tied to profile adherence."

    FACD constructs its negative prompt by deleting exactly the immoral profile fields that the contrastive term then amplifies, and CoSER's Character Fidelity is defined as 'the degree to which the character profile is faithfully reflected in the generated narrative.' Thus the reported recovery in Character Fidelity (e.g., 6.39→14.38 for Qwen3-8B) is the method optimizing precisely the quantity that the metric measures. Presenting this as confirmation that 'FACD successfully amplifies the exact immoral profile fields that safety alignment suppresses' makes the gain partly an artifact of the method-metric alignment, not an independent test of the alignment-suppression explanation. The non-fidelity dimensions (Anthropomorphism, Storyline Quality) are less forced, which is why the circularity is

full rationale

The paper's core diagnostic claim is empirical and much of it survives scrutiny: CoSER shows Moral–Immoral gaps across Anthropomorphism, Character Fidelity, and Storyline Quality, and the human evaluation on 20% of characters tracks the automatic trends, so the disposition finding is not wholly fabricated. However, two load-bearing pieces of evidence are partially circular. First, on PersonaGym the headline 'Disposition produces large, consistent performance degradation' is computed as the average of five fixed rubrics; the Toxicity Control rubric contributes an outsized negative delta (up to -1.88) while the other four deltas are near zero or mixed, so the aggregate gap is largely a re-expression of the rubric's decision to penalize toxic content rather than an independent measurement of role-playing capability. Second, FACD constructs its negative prompt by deleting exactly the immoral fields it then amplifies, and its largest reported success is on CoSER's Character Fidelity, a metric defined as profile adherence; boosting profile-specific tokens therefore moves the measured quantity by construction. This does not make the whole paper circular: the CoSER results and the non-fidelity dimensions are not forced, and no load-bearing self-citation or imported uniqueness theorem is used. But because one half of the RQ1 evidence and the headline RQ2 metric reduce to the definitions of the evaluation and decoding objectives, partial circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or formal entities. Its free parameters are concentrated in the dataset design (disposition cut, coherence threshold) and in FACD (α, fallback field selection). The most consequential assumption is the neutrality of the evaluation metrics, especially PersonaGym's Toxicity Control, which directly targets the behavior under study.

free parameters (5)
  • FACD amplification strength α = 1
    Set to 1 in §5.2 without a reported ablation; the headline FACD results depend on this value.
  • Disposition classification cutoff = score ≤5 Immoral, ≥6 Moral
    Binary split chosen in §3.2; the exact cutoff drives group membership and all downstream comparisons.
  • Synthetic profile coherence threshold = ≥8/10 to retain
    LLM-as-Judge coherence filter in §3.2; only profiles scoring 8+ are retained, affecting the Unknown set.
  • Fpad fallback field ordering = Fields ordered by empirical stability in Figure 4
    The fallback set for the FACD negative prompt is selected based on the paper's own stability ranking, i.e., the method is tuned on the same data used for the headline evaluation.
  • Fpad trigger threshold = fewer than six non-PA fields classified as moral
    The decision to add polarity-insensitive fallback fields is a hand-set threshold in §5.2; no robustness analysis is given.
axioms (5)
  • domain assumption Evaluation metrics (PersonaGym PersonaScore, CoSER Score) measure role-playing fidelity without moral-valence bias.
    PersonaGym's Toxicity Control rubric penalizes toxic output by construction; if this bias is not controlled, the Disposition gap is partly a yardstick effect. This is load-bearing for RQ1 (§4.2, Table 20).
  • domain assumption GPT-4o judge scores correlate with human judgment across Moral and Immoral groups.
    Human validation covers only a 20% sample on two backbone models (§C.2), not the full five-model, two-benchmark grid.
  • domain assumption Structured-to-unstructured rewriting by Claude-4.5-haiku preserves semantic content.
    The Structure axis comparison assumes the only difference between Structured and Unstructured conditions is formatting, not content (§3.2, Appendix A.3).
  • ad hoc to paper The three axes map to pretraining, instruction tuning, and value alignment phases.
    The mapping Familiarity→pretraining, Structure→instruction tuning, Disposition→alignment is a framing hypothesis in §1 and is not independently established.
  • ad hoc to paper Post-SFT alignment suppresses tokens needed for immoral portrayal.
    The causal mechanism is inferred from performance gaps (§4.3, §5.1); no base-model vs instruction-tuned-model control or direct token-level suppression evidence is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 23939 in / 11304 out tokens · 121679 ms · 2026-08-03T11:54:22.212138+00:00 · methodology

0 comments
read the original abstract

While Large Language Model (LLM) role-playing agents have advanced rapidly, it remains unclear which profile elements genuinely drive role-playing quality. To bridge this gap, we introduce a systematic diagnostic framework that disentangles the impact of character profiles along three axes: Familiarity (Known vs. Unknown), Structure (Structured vs. Unstructured), and Disposition (Moral vs. Immoral). Utilizing a unified hierarchical schema (5 dimensions, 28 fields), we construct a controlled dataset of 211 personas and evaluate five LLMs on both single- and multi-turn interactions. Our results reveal a striking asymmetry: Familiarity and Structure show negligible impact, while Disposition produces large, consistent performance degradation for immoral characters across all conditions. Further analyses suggest that the Moral--Immoral gap is amplified by post-SFT alignment, and that this degradation varies substantially across profile attributes. To mitigate this bottleneck, we propose Field-Aware Contrastive Decoding (FACD), a training-free strategy that amplifies suppressed disposition-sensitive signals, significantly closing the performance gap without sacrificing moral-character performance.

Figures

Figures reproduced from arXiv: 2601.04716 by Hwanhee Lee, Jeonghyun Park, Jihyeong Park, Junhyuk Choi, Liu Nicole Geumheon, Yonghyun Jun.

Figure 1
Figure 1. Figure 1: The three axes of character profile: (A) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hierarchical character profile schema: 28 leaf [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of profile sub-field embed [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of negative character on CoSER scores. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of steering performance and miti [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Turn-wise Trends in Role-Playing Perfor [PITH_FULL_IMAGE:figures/full_fig_p015_6.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

11 extracted references · 8 linked inside Pith

  1. [3]

    Yijiang River Dong, Tiancheng Hu, Zheng Hui, and Nigel Collier

    Gemini 2.5: Pushing the frontier with ad- vanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261. Yijiang River Dong, Tiancheng Hu, Zheng Hui, and Nigel Collier. 2026. Steer model beyond assistant: Controlling system prompt strength via contrastive decoding.arXiv preprint arXiv:2601.06403. I...

    Pith/arXiv arXiv 2026

  2. [4]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al

    Safety tax: Safety alignment makes your large reasoning models less reasonable.arXiv preprint arXiv:2503.00555. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276. Sarthak Jain and Byron C. Wallace. 2019....

    Pith/arXiv arXiv 2024

  3. [6]

    InProceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 7828–7840, Bangkok, Thailand

    Large language models are superpositions of all characters: Attaining arbitrary role-play via self-alignment. InProceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 7828–7840, Bangkok, Thailand. Association for Computational Linguistics. Laurens van der Maaten and Geoffrey Hinton. 2008. V...

    2008

  4. [9]

    InProceedings of the 2023 Conference on Empirical Methods in Natural Language Process- ing, pages 13153–13187, Singapore

    Character-LLM: A trainable agent for role- playing. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Process- ing, pages 13153–13187, Singapore. Association for Computational Linguistics. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Esin DURMUS, Zac Hatfield-Dodds, Scott R John- sto...

    2023

  5. [10]

    InThe Twelfth International Confer- ence on Learning Representations

    Towards understanding sycophancy in lan- guage models. InThe Twelfth International Confer- ence on Learning Representations. Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuo- han Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Ré, Ion Stoica, and Ce Zhang. 2023. Flexgen: High-throughput generative inference of large language models with a single gpu. ...

    Pith/arXiv arXiv 2023

  6. [11]

    reasoner

    Evaluating character understanding of large language models via character profiling from fictional works. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8015–8036, Miami, Florida, USA. Association for Computational Linguistics. Wenyuan Zhang, Shuaiyi Nie, Jiawei Sheng, Zefeng Zhang, Xinghua Zhang, Yongquan ...

    Pith/arXiv arXiv 2024

  7. [1991]

    Yonghyun Jun and Hwanhee Lee

    Big five inventory.Journal of personality and social psychology. Yonghyun Jun and Hwanhee Lee. 2025. Exploring per- sona sentiment sensitivity in personalized dialogue generation. InProceedings of the 63rd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18384–18402, Vi- enna, Austria. Association for Comput...

    Pith/arXiv arXiv 2025

  8. [2022]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein

    Training language models to follow instruc- tions with human feedback.Advances in neural in- formation processing systems, 35:27730–27744. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein. 2023. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th an- nual acm...

    2023

  9. [2023]

    merge conflicts!

    "merge conflicts!" exploring the impacts of external distractors to parametric knowledge graphs. arXiv preprint arXiv:2309.08594. Yiting Ran, Xintao Wang, Tian Qiu, Jiaqing Liang, Yanghua Xiao, and Deqing Yang. 2025. Bookworld: From novels to interactive agent societies for story creation. InProceedings of the 63rd Annual Meet- ing of the Association for ...

    Pith/arXiv arXiv 2025

  10. [2024]

    InFindings of the Association for Computational Linguistics: ACL 2024, pages 3291–3325, Bangkok, Thailand

    TimeChara: Evaluating point-in-time character hallucination of role-playing large language models. InFindings of the Association for Computational Linguistics: ACL 2024, pages 3291–3325, Bangkok, Thailand. Association for Computational Linguistics. Anthropic. 2025. Introducing claude sonnet 4.5. Hila Chefer, Shir Gur, and Lior Wolf. 2021. Trans- former in...

    Pith/arXiv arXiv 2024

  11. [2025]

    Amey Agrawal, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S Gulavani, and Ramachan- dran Ramjee

    gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925. Amey Agrawal, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S Gulavani, and Ramachan- dran Ramjee. 2023. Sarathi: Efficient llm infer- ence by piggybacking decodes with chunked prefills. arXiv preprint arXiv:2308.16369. Jaewoo Ahn, Taehyun Lee, Junyoung Lim, Jin-Hwa Kim, Sangdo...

    Pith/arXiv arXiv 2023