Pith. sign in

REVIEW 2 major objections 5 minor 50 references

When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Task-vector interference is measurable and input-dependent, not a universal semantic.

desk verdict A careful empirical mapping of task-vector interference: input- and format-conditioned, honestly bounded, but the adapters' semantic identity is left unverified. read the letter →

arxiv 2608.09490 v1 pith:CTBHGZRD submitted 2026-08-10 cs.LG

Chencheng Zhu , Xiaoyang Li , Taotao Cai This is my paper · ORCID
classification cs.LG
keywords taskarithmeticvectorsmodelmergingfunctionalinterferenceweight-spacecompositionLoRAinput-conditionedgeometryvalidityboundaries
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks when adding fine-tuning task vectors in weight space produces predictable changes in model behavior. It separates parameter geometry from functional geometry and measures pairwise non-additivity over a two-task composition surface, using a first-token logit-space interaction ratio normalized by base displacement. On Qwen2.5-1.5B, code+safety is more non-additive than code+math on code and instruction prompts but not on math prompts. A prospectively specified six-task expansion yields all eight predicted high-versus-low signs for unseen task pairs, and the contrast persists under full fine-tuning, larger scales, and a second model family. The conclusion is that weight-space composition supports coarse, input- and format-conditioned functional statements, not a universal merging-performance predictor.

What carries the argument

The central object is the input-conditioned interaction ratio $R^{(s)}_{a,b}(X)$ from Eq. (5): the ratio of the expected first-token Jensen-Shannon divergence between the observed composed distribution and an additive counterfactual, to the expected divergence from the base. The counterfactual is built in logit space as $\ell_{\alpha,0}+\ell_{0,\beta}-\ell_{0,0}$, which removes the marginal nonlinearity of each axis before measuring their joint effect. Norm-matched controls equalize vector radii across tasks and seeds, prospective bins and frozen success rules provide transfer tests on unseen pairs, and a hierarchical bootstrap over seeds and prompts supplies uncertainty intervals.

What would settle it

Recompute the central code+safety versus code+math contrast with a different operationalization of functional interaction-for example, full-sequence likelihood or fine-grained task accuracy on code and instruction benchmarks-using the same seeds and norm-matched endpoints; if the direction reverses or vanishes on code and instruction prompts, the claimed boundary is an artifact of the first-token JSD ratio.

Watch

Extended reading notes

Core claim

The central claim is that functional interference between task vectors is jointly determined by the task pair and the input distribution, and that this interference can be measured by a first-token interaction ratio $R^{(s)}_{a,b}(X)$: the ratio of Jensen-Shannon divergence between the actual merged distribution and an additive logit-space counterfactual, to the divergence from the base. The paper establishes that the code+safety pair exceeds the code+math pair by 5.96 and 7.58 percentage points on code and instruction prompts but only 0.10 on math prompts; all eight prospective high-versus-low comparisons on unseen task pairs have the predicted sign; and the ordering persists under full-parameter fine-tuning at 0.5B, Qwen2.5 scales up to 7B, and a Llama-3.1-8B cross-architecture audit. The same signal collapses when identical public code prompts are wrapped in an instruction template and is not robustly reproduced by EvalPlus pass@1. Therefore task-vector interference is conditionally generalizable, not a global semantic coordinate system in weight space.

Load-bearing premise

The measurement rests on the assumption that the first-token interaction ratio R, computed in logit space and normalized by base displacement, captures the meaningful functional interference between task vectors; the paper does not report task accuracy for adapters, so task labels describe training data rather than verified capabilities.

Editorial extensions

If this is right

  • If a practitioner needs to predict whether two adapters interfere, the input prompt distribution and its serialization must be specified; there is no input-free answer.
  • Parameter-space summaries such as cosine similarity or clustering are not reliable proxies for functional interference: at 0.5B the functional contrast persists without detectable task clustering.
  • The code+safety versus code+math hierarchy transfers across rank-16 LoRA, full fine-tuning at 0.5B, Qwen2.5 scaling to 7B, and Llama-3.1-8B, so the coarse ordering is robust to adaptation method, scale, and one additional model family.
  • Continuous first-token interaction does not predict discrete benchmark merge performance: EvalPlus pass@1 interactions are wide and inconsistent, so evaluation on intended prompts remains necessary.
  • Because prompt format flips the result on identical prompts, wrapper templates are part of the conditioning distribution rather than a neutral serialization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same ratio machinery could be applied layer-wise or at multiple decoding steps to localize where interference enters the network, which the paper names as the natural next test.
  • The results suggest that a safety adapter is not a semantically clean axis: the code+safety contrast may partly reflect training-data characteristics or refusal behavior rather than a stable safety direction.
  • If the boundary is correct, task-arithmetic methods that optimize a single global coefficient set should be re-evaluated per prompt stratum, since a fixed coefficient set cannot be optimal across formats.
  • A testable extension is to use the interaction ratio R as a screening score on many task pairs and compare its ranking with measured merge accuracy on a fixed prompt distribution; the paper's pass@1 result suggests the ranking may fail on discrete metrics.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper studies functional non-additivity of task-vector composition by defining an input-conditioned interaction ratio R (Eq. 5) that compares the observed merged logit distribution against a logit-space additive counterfactual built from the two axis paths, normalized by displacement from the base. On Qwen2.5-1.5B with norm-matched LoRA vectors, the paper reports that code+safety has higher R than code+math on code and instruction prompts but not math prompts; a prospectively specified six-task expansion yields 8/8 high-vs-low sign agreements; the contrast persists under 0.5B full fine-tuning, Qwen2.5-7B LoRA, and Llama-3.1-8B; and public raw prompts preserve but an instruction wrapper collapses the code-prompt contrast, while EvalPlus pass@1 interactions do not robustly reproduce it. The paper concludes that weight-space composition supports coarse, input- and format-conditioned functional statements but is not a universal merging-performance predictor.

Significance. The result, if it holds, is a useful boundary result for task arithmetic: it separates parameter geometry from functional geometry and provides a concrete measurement surface rather than a fitted merge-quality predictor. The methodology has notable strengths: matched checkpoints within seed, core median norm matching, frozen prospective specifications with dated summaries, a cached estimator audit that reproduces the frozen bootstrap values, explicit denominator-boundary checks, a design-preserving exact randomization test, and a claim-boundary checklist that distinguishes supported from unsupported conclusions. The paper is honest about its limits, including the absence of per-adapter task accuracy and the failure of the metric to predict pass@1. Its significance is therefore conditional and primarily negative/descriptive: it maps where composition behaves non-additively rather than offering a new merging algorithm.

major comments (2)
  1. [Discussion and Limitations; Eq. (5)] The absence of any per-adapter capability validation is load-bearing for the interpretation of the central code+safety vs. code+math contrast as task-vector interference. Because the code checkpoint is shared within each seed, the uncontrolled comparison is between the safety adapter and the math adapter; if the safety adapter is degenerate, miscalibrated, or dominated by a generic refusal/format shift, the elevated R on code and instruction prompts could reflect properties of that adapter's output distribution rather than an interaction between task directions. This concern propagates to the OOS ordering and to every stress-test row, which inherit the same task labels. The Discussion explicitly acknowledges "We do not report task accuracy for the individual adapters," but the acknowledgement does not resolve the ambiguity. I request per-adapter validation (for example, held-out loss or task accuracy for each adapter on its own evaluation split, plus perhaps a norm-matched random-direction control) or, alternatively, a systematic reframing of all task-level conclusions as statements about data-conditional fine-tuning displacements rather than about "tasks."
  2. [Expanded Methods; Claim Boundary Checklist] The paper's own limitation statement notes the absence of same-task and random-direction anchors. Without a norm-matched random-direction baseline, the paper cannot distinguish task-vector-specific interference from generic properties of any large norm-matched displacement in the model's weight space. The central claim that "task-vector interference is measurable" requires at least one such control; the existing comparisons between task pairs are informative but do not establish that the effect is specific to task vectors. I recommend adding a random-direction condition (e.g., Gaussian or permuted directions matched in norm) or softening the conclusion accordingly.
minor comments (5)
  1. [Results, Tables 2-5 and Eq. (5)] R is a dimensionless ratio, but the text reports it in "percentage points" and "percentages" (e.g., Table 3 caption). Please state explicitly that all reported values are 100 times R.
  2. [Functional Structure Persists Without Detectable Parameter Clustering] The section heading overstates the 1.5B finding, where the clustering diagnostics are significant (cosine gap p=0.0002); the heading applies to the 0.5B block. Consider rewording to "Persistence without Detectable Clustering at 0.5B."
  3. [Prospective Directional Test on Unseen Pairs] The sentence "Accuracy for comparisons between high and low bins is 37.5% on math prompts, 50% on code prompts..." uses "accuracy" for a sign-agreement rate; rename to avoid confusion with model task accuracy, especially given the paper's own caution about missing task accuracy.
  4. [Estimator and Denominator Audit] The sentence "Table 7 shows that the absolute numerator and denominator vary substantially..., so the main result is not inferred from a contrast alone" is unclear; the displayed variation does not by itself support that conclusion, and the sentence could be removed or rewritten.
  5. [Abstract and Prospective Directional Test] The abstract's "all eight ... predicted sign" should be accompanied by the exact design-preserving p-value (0.083) or a qualifier such as "directionally, though not statistically definitive," to prevent overinterpretation by casual readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical measurements with prospectively frozen specifications, and the interaction ratio is explicitly presented as a chosen operationalization rather than a derived prediction.

full rationale

I find no circular step that reduces a claimed prediction to its own inputs. The central quantity, R in Eq. (5), is a measurement operationalization defined as the ratio of a JSD-based interaction numerator to a base-displacement denominator; the paper explicitly states 'It is a chosen operationalization of additivity between directions, not a definition of functional interaction that is independent of the model.' The headline contrast between code+safety and code+math is an empirical comparison across trained adapters and prompt strata, not a parameter fitted to produce that ordering. The prospective six-task expansion was specified before the summarization and translation adapters were trained, as the paper states: 'The comparison bins and success rule were recorded before the summarization and translation adapters were trained.' No fitted parameter is renamed as a prediction, and no load-bearing argument depends on a self-citation: the closest-work citations (Zhou et al. 2026; Sivaramakrishnan et al. 2026) are not by the present authors. The acknowledged limitations, such as not reporting per-adapter task accuracy and lacking same-task or random-direction anchors, are validity concerns about what the task labels and absolute ratio values mean; they do not make the empirical contrast equivalent to its definition by construction. The derivation chain, insofar as one exists, is self-contained measurement plus preregistered directional tests, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a chosen interaction metric, linear weight-space interpolation, and unverified task labels. No new physical or conceptual entity is posited; the contribution is a measurement and its boundary conditions.

free parameters (2)
  • median norm-matching radius = median Frobenius norm of rank-16 task vectors in the corresponding audit block
    Central contrasts are reported after rescaling every task vector to this radius. The choice of norm is an intervention and affects absolute magnitudes, though the authors say natural radii give similar results.
  • interaction grid and epsilon = alpha, beta in {0.2, 0.4, ..., 1.2}; epsilon = 1e-12
    The ratio of expectations is computed over these 36 interior grid points. The grid and epsilon are chosen by hand and define the estimator.
assumptions (4)
  • domain assumption Task vectors are additive displacements from a shared base model (theta0 + alpha*Delta_a + beta*Delta_b).
    The entire surface evaluation assumes fine-tuning differences can be interpolated linearly in weight space; this comes from prior task arithmetic work.
  • ad hoc to paper Logit-space additive counterfactual l_add = l_alpha0 + l_0beta - l_00 defines no-interaction.
    The paper states this is a chosen operationalization, not an independent definition of functional interaction.
  • ad hoc to paper Normalizing interaction by base displacement makes surfaces comparable.
    If the denominator M is small, R is unstable; the paper audits denominators but relies on this normalization for all main comparisons.
  • domain assumption Task labels describe learned capabilities.
    The authors explicitly do not report task accuracy, so safety/code/math labels are training-data labels, not verified skills.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition." pith.science (2026). https://pith.science/paper/CTBHGZRD

@misc{pith2026260809490,
  author       = {Pith},
  title        = {Pith review of: When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTBHGZRD}},
  note         = {Machine review of arXiv:2608.09490}
}
read the original abstract

Task arithmetic treats fine-tuning displacements as composable directions in weight space, yet it remains unclear when parameter addition reflects predictable changes in model function. We separate parameter geometry from functional geometry and measure pairwise functional non-additivity over a two-dimensional task-vector surface, using a first-token predictive-distribution interaction ratio conditioned on an input distribution and evaluated with norm-matched controls, three training seeds, and response-only fine-tuning. On Qwen2.5-1.5B, code+safety is more non-additive than the matched code+math control on code and instruction prompts, but not on math prompts. In a prospectively specified six-task expansion, all eight high-versus-low comparisons of unseen task pairs have the predicted sign. The primary ordering further persists under full-parameter fine-tuning at 0.5B, Qwen2.5 LoRA scale tests up to 7B, and a Llama-3.1-8B cross-architecture audit. External validation exposes a sharper boundary: raw public code, instruction, and safety prompts preserve the continuous contrast, whereas an instruction-style wrapper collapses it on the identical public-code prompts, and EvalPlus pass@1 interactions do not robustly reproduce it. Weight-space composition therefore supports coarse, input- and format-conditioned functional statements across adaptation methods, scales, and one additional model family, not a universal merging-performance predictor.

Figures

Figures reproduced from arXiv: 2608.09490 by the authors.

Figure 1
Figure 1. Measurement pipeline for input-conditioned functional interaction. Two task vectors from a shared base define a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Functional interference depends on input stratum, prompt format, and evaluation metric. (a) Contrasts between [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. A prospectively specified directional score transfers to unseen pairs, but depends on the input. (a) Mean interaction [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    International Conference on Learning Representations , year =

    Gabriel Ilharco and Marco Tulio Ribeiro and Mitchell Wortsman and Suchin Gururangan and Ludwig Schmidt and Hannaneh Hajishirzi and Ali Farhadi , title =. International Conference on Learning Representations , year =

    work page

  2. [2]

    Advances in Neural Information Processing Systems , volume =

    Guillermo Ortiz-Jimenez and Alessandro Favero and Pascal Frossard , title =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    work page 2023

  3. [3]

    Advances in Neural Information Processing Systems , year =

    Prateek Yadav and Derek Tam and Leshem Choshen and Colin Raffel and Mohit Bansal , title =. Advances in Neural Information Processing Systems , year =

    work page

  4. [4]

    International Conference on Learning Representations , year =

    Enneng Yang and Zhenyi Wang and Li Shen and Shiwei Liu and Guibing Guo and Xingwei Wang and Dacheng Tao , title =. International Conference on Learning Representations , year =

    work page

  5. [5]

    Transactions on Machine Learning Research , year =

    Yifei He and Yuzheng Hu and Yong Lin and Tong Zhang and Han Zhao , title =. Transactions on Machine Learning Research , year =

    work page

  6. [6]

    A Systematic Study of In-the-Wild Model Merging for Large Language Models , journal =

    O. A Systematic Study of In-the-Wild Model Merging for Large Language Models , journal =. 2026 , url =

    work page 2026

  7. [7]

    Demystifying Mergeability: Interpretable Properties to Predict Model Merging Success , journal =

    Luca Zhou and Bo Zhao and Rose Yu and Emanuele Rodol. Demystifying Mergeability: Interpretable Properties to Predict Model Merging Success , journal =. 2026 , url =

    work page 2026

  8. [8]

    PermDoRA -- Understanding Adapter Interference in Language Models: Limits of Parameter-Space Geometry

    Gowtham Sivaramakrishnan and Sarvesha Kumar Kombaiah Seetha and Kishan Gupta Balaji and Santhosh Baradwaj Vaduvur Ranganathan , title =. arXiv preprint arXiv:2606.11262 , year =

    work page Pith review arXiv

Show all 50 references
  1. [9]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , title =

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , title =. International Conference on Learning Representations , year =

  2. [10]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , year =

    Armen Aghajanyan and Luke Zettlemoyer and Sonal Gupta , title =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , year =

  3. [11]

    arXiv preprint arXiv:2405.16833 , year =

    Chia-Yi Hsu and Yu-Lin Tsai and Chih-Hsun Lin and Pin-Yu Chen and Chia-Mu Yu and Chun-Ying Huang , title =. arXiv preprint arXiv:2405.16833 , year =

    arXiv

  4. [12]

    arXiv preprint arXiv:2412.15115 , year =

    arXiv

  5. [13]

    arXiv preprint arXiv:2407.21783 , year =

    Abhimanyu Dubey and Abhinav Jauhri and Abhinav Pandey and Abhishek Kadian and Ahmad Al-Dahle and Aiesha Letman and Akhil Mathur and Alan Schelten and Amy Yang and Angela Fan and others , title =. arXiv preprint arXiv:2407.21783 , year =

    arXiv

  6. [14]

    arXiv preprint arXiv:2305.01210 , year =

    Jiawei Liu and Chunqiu Steven Xia and Yuyao Wang and Lingming Zhang , title =. arXiv preprint arXiv:2305.01210 , year =

    arXiv

  7. [15]

    arXiv preprint arXiv:2107.03374 , year =

    Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan and Henrique Ponde de Oliveira Pinto and Jared Kaplan and Harri Edwards and Yuri Burda and Nicholas Joseph and Greg Brockman and others , title =. arXiv preprint arXiv:2107.03374 , year =

    arXiv

  8. [16]

    Le and Charles Sutton , title =

    Jacob Austin and Augustus Odena and Maxwell Nye and Maarten Bosma and Henryk Michalewski and David Dohan and Ellen Jiang and Carrie Cai and Michael Terry and Quoc V. Le and Charles Sutton , title =. arXiv preprint arXiv:2108.07732 , year =

    arXiv

  9. [17]

    arXiv preprint arXiv:2110.14168 , year =

    Karl Cobbe and Vineet Kosaraju and Mohammad Bavarian and Mark Chen and Heewoo Jun and Lukasz Kaiser and Matthias Plappert and Jerry Tworek and Jacob Hilton and Reiichiro Nakano and Christopher Hesse and John Schulman , title =. arXiv preprint arXiv:2110.14168 , year =

    arXiv

  10. [18]

    Cohen and Mirella Lapata , title =

    Shashi Narayan and Shay B. Cohen and Mirella Lapata , title =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year =

    2018

  11. [19]

    Parallel Data, Tools and Interfaces in

    J. Parallel Data, Tools and Interfaces in. Proceedings of the Eighth International Conference on Language Resources and Evaluation , pages =. 2012 , url =

    2012

  12. [20]

    arXiv preprint arXiv:2406.15513 , year =

    Jiaming Ji and Donghai Hong and Borong Zhang and Boyuan Chen and Josef Dai and Boren Zheng and Tianyi Qiu and Boxun Li and Yaodong Yang , title =. arXiv preprint arXiv:2406.15513 , year =

    arXiv

  13. [21]

    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 =. 2023 , howpublished =

    2023

  14. [22]

    2023 , howpublished =

    Sahil Chaudhary , title =. 2023 , howpublished =

    2023

  15. [23]

    Morcos and Hongseok Namkoong and Ali Farhadi and Yair Carmon and Simon Kornblith and Ludwig Schmidt , title =

    Mitchell Wortsman and Gabriel Ilharco and Samir Ya Gadre and Rebecca Roelofs and Raphael Gontijo-Lopes and Ari S. Morcos and Hongseok Namkoong and Ali Farhadi and Yair Carmon and Simon Kornblith and Ludwig Schmidt , title =. Proceedings of the 39th International Conference on ...

    2022

  16. [24]

    Matena and Colin A

    Michael S. Matena and Colin A. Raffel , title =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  17. [25]

    International Conference on Learning Representations , year =

    Rahim Entezari and Hanie Sedghi and Olga Saukh and Behnam Neyshabur , title =. International Conference on Learning Representations , year =

  18. [26]

    Ainsworth and Jonathan Hayase and Siddhartha Srinivasa , title =

    Samuel K. Ainsworth and Jonathan Hayase and Siddhartha Srinivasa , title =. International Conference on Learning Representations , year =

  19. [27]

    Proceedings of the 40th International Conference on Machine Learning , volume =

    Sadhika Malladi and Alexander Wettig and Dingli Yu and Danqi Chen and Sanjeev Arora , title =. Proceedings of the 40th International Conference on Machine Learning , volume =. 2023 , url =

    2023

  20. [28]

    Lee and Ernest K

    Uijeong Jang and Jason D. Lee and Ernest K. Ryu , title =. Proceedings of the 41st International Conference on Machine Learning , volume =. 2024 , url =

    2024

  21. [29]

    Localizing Task Information for Improved Model Merging and Compression , booktitle =

    Ke Wang and Nikolaos Dimitriadis and Guillermo Ortiz-Jimenez and Fran. Localizing Task Information for Improved Model Merging and Compression , booktitle =. 2024 , url =

    2024

  22. [30]

    International Conference on Learning Representations , year =

    George Stoica and Daniel Bolya and Jakob Brandt Bjorner and Pratik Ramesh and Taylor Hearn and Judy Hoffman , title =. International Conference on Learning Representations , year =

  23. [31]

    Smith and Hannaneh Hajishirzi and Pang Wei Koh and Jesse Dodge and Pradeep Dasigi , title =

    Jacob Morrison and Noah A. Smith and Hannaneh Hajishirzi and Pang Wei Koh and Jesse Dodge and Pradeep Dasigi , title =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , doi =

    2024

  24. [32]

    International Conference on Learning Representations , year =

    George Stoica and Pratik Ramesh and Boglarka Ecsedi and Leshem Choshen and Judy Hoffman , title =. International Conference on Learning Representations , year =

  25. [33]

    International Conference on Learning Representations , year =

    Ziyu Zhao and Tao Shen and Didi Zhu and Zexi Li and Jing Su and Xuwu Wang and Fei Wu , title =. International Conference on Learning Representations , year =

  26. [34]

    Proceedings of the 42nd International Conference on Machine Learning , year =

    Runxi Cheng and Feng Xiong and Yongxian Wei and Wanyun Zhu and Chun Yuan , title =. Proceedings of the 42nd International Conference on Machine Learning , year =

  27. [35]

    Vetrov and Andrew Gordon Wilson , title =

    Pavel Izmailov and Dmitrii Podoprikhin and Timur Garipov and Dmitry P. Vetrov and Andrew Gordon Wilson , title =. Conference on Uncertainty in Artificial Intelligence , pages =. 2018 , url =

    2018

  28. [36]

    Vetrov and Andrew G

    Timur Garipov and Pavel Izmailov and Dmitrii Podoprikhin and Dmitry P. Vetrov and Andrew G. Wilson , title =. Advances in Neural Information Processing Systems , volume =. 2018 , url =

    2018

  29. [37]

    Hamprecht , title =

    Felix Draxler and Kambis Veschgini and Manfred Salmhofer and Fred A. Hamprecht , title =. Proceedings of the 35th International Conference on Machine Learning , volume =. 2018 , url =

    2018

  30. [38]

    Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalization , booktitle =

    Alexandre Ram. Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalization , booktitle =. 2023 , url =

    2023

  31. [39]

    International Conference on Learning Representations , year =

    Xisen Jin and Xiang Ren and Daniel Preotiuc-Pietro and Pengxiang Cheng , title =. International Conference on Learning Representations , year =

  32. [40]

    Proceedings of the 41st International Conference on Machine Learning , volume =

    Le Yu and Bowen Yu and Haiyang Yu and Fei Huang and Yongbin Li , title =. Proceedings of the 41st International Conference on Machine Learning , volume =. 2024 , url =

    2024

  33. [41]

    Transactions on Machine Learning Research , year =

    Derek Tam and Mohit Bansal and Colin Raffel , title =. Transactions on Machine Learning Research , year =

  34. [42]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics , pages =

    Jonas Pfeiffer and Aishwarya Kamath and Andreas R. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics , pages =. 2021 , doi =

    2021

  35. [43]

    Findings of the Association for Computational Linguistics: EACL 2023 , pages =

    Alexandra Chronopoulou and Matthew Peters and Alexander Fraser and Jesse Dodge , title =. Findings of the Association for Computational Linguistics: EACL 2023 , pages =. 2023 , doi =

    2023

  36. [44]

    Fowlkes and Stefano Soatto and Pietro Perona , title =

    Alessandro Achille and Michael Lam and Rahul Tewari and Avinash Ravichandran and Subhransu Maji and Charless C. Fowlkes and Stefano Soatto and Pietro Perona , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2019 , url =

    2019

  37. [45]

    Proceedings of the 37th International Conference on Machine Learning , volume =

    Trevor Standley and Amir Zamir and Dawn Chen and Leonidas Guibas and Jitendra Malik and Silvio Savarese , title =. Proceedings of the 37th International Conference on Machine Learning , volume =. 2020 , url =

    2020

  38. [46]

    Advances in Neural Information Processing Systems , volume =

    Chris Fifty and Ehsan Amid and Zhe Zhao and Tianhe Yu and Rohan Anil and Chelsea Finn , title =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

    2021

  39. [47]

    Length-Controlled

    Yann Dubois and Bal. Length-Controlled. First Conference on Language Modeling , year =

  40. [48]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =

    Paul R. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages =. 2024 , doi =

    2024

  41. [49]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

    Almog Gueta and Elad Venezian and Colin Raffel and Noam Slonim and Yoav Katz and Leshem Choshen , title =. Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

    2023

  42. [50]

    Conference on Language Modeling , year =

    Chengsong Huang and Qian Liu and Bill Yuchen Lin and Tianyu Pang and Chao Du and Min Lin , title =. Conference on Language Modeling , year =

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.