Pith. sign in

REVIEW 3 major objections 5 minor 18 references

A frozen, text-only language model can set per-object force limits and pick recovery moves from force signatures, while a hard clamp in the control loop—not the model—enforces the ceiling; one policy then passes 256/256 clean insertions on

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 08:04 UTC pith:GP2BAYEX

load-bearing objection A carefully scoped, honestly reported simulation study whose safety findings are solid, but the headline recovery advantage rests on 25-episode cells and the LLM budget-setter is effectively a fixed constant. the 3 major comments →

arxiv 2607.21227 v1 pith:GP2BAYEX submitted 2026-07-23 cs.RO

FORGE-plus: Force-Budgeted Recovery for Contact-Rich Assembly with a Frozen LLM Supervisor

Kyupaeck Jeff Rah , Midum Oh This is my paper
classification cs.RO
keywords force-budgeted assemblyforce ceilinginsertion failure recoverylarge language model supervisionforce signaturefragile partsbehavior cloningrigid-body simulation
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 the two open questions left by force-conditioned assembly skills—who sets the force ceiling and what to do when insertion fails—can be answered by a deliberately thin semantic layer. A frozen, text-only language model reads the object's identity and sets a per-object force budget before the episode; on failure it reads a compact textual force signature and picks a fixed recovery maneuver. The language model never controls force: a hard clamp in the fast control loop enforces the ceiling, the recovery menu has no 'raise the ceiling' option, and the hidden breaking threshold is visible only to the evaluator. The central evidence is one unified policy checkpoint passing 256/256 clean insertion gates on both a fragile gear and a steel gear with zero breaks and correct learned release timing, plus a fully physical table-pick flow at 5.4 N mean peak force. Under an injected in-grip slip, the force-signature recovery resolves 40% and 64% of jams on two grippers, while a press-harder baseline is either futile or breaks 96% of parts; the paper also reports negative results (on-policy RL fails at this clearance, and several natural learned-release heads fail) and explicitly makes no sim-to-real claim.

Core claim

The core discovery is that a force-budgeted assembly policy can be made reliable for fragile parts without granting the language model any force authority. The paper shows that a frozen, text-only LLM, called at most twice per episode, can set Fmax from object identity and choose a recovery from a fixed menu by reading a compact text force signature (peak axial force, net insertion, axial-rising flag, lateral bias, contact persistence, slip events). The fast layer—a force-conditioned skill on an impedance controller with a hard clamp—enforces the ceiling. With the hidden per-episode breaking force Fbreak known only to the evaluator, one unified checkpoint passes 256/256 clean gates on both a

What carries the argument

The load-bearing mechanism is a two-layer separation of semantic authority and force authority. A slow, frozen text-only LLM layer contains a budget-setter (identity→Fmax) and a recovery-selector (force signature→menu action); a fast control layer contains the force-conditioned policy, an impedance controller, and a hard force clamp that saturates commanded wrench every step. The central object is the compact text force signature—a short token listing peak axial force, net insertion, axial-rising state, lateral bias, contact persistence, and slip events—because it is the only signal that discriminates failures that look alike on camera (a rim wedge versus a recurring contactless hover from a

Load-bearing premise

Everything rests on the assumption that the benchmark distribution—the same two objects, same rigid-body simulator, same injected faults (rim wedge, 5 mm in-grip slip), and the same detector thresholds and budgets the authors tuned—represents the failures a deployed system would actually meet. If real faults include burrs, cross-threading, or deformation-based breakage, the force-signature discriminator and the fixed recovery menu may not recognize them.

What would settle it

Give the released checkpoint two held-out fragile object classes and one new fault type, e.g. cross-threading, with hidden Fbreak drawn per episode, and measure clean-gate breakage and recovery success over 256 episodes. If the identity-derived budget produces breaks or the force-signature chain cannot route to the right maneuver, the paper's central claim about generalizable force-budgeted recovery is wrong. A cheaper physics check: measure peak contact force at funnel entry on hardware with the same Fmax; if overshoot reaches ~1.5× the budget, the clamp-fidelity assumption behind the oracle-

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

If this is right

  • Force-conditioned assembly skills can be made safe for fragile parts without per-part manual tuning: the force ceiling comes from object identity text, and the fast-loop clamp—not the LLM—is what enforces it.
  • Failures that look identical on camera can be distinguished from short force signatures, at least for the injected fault types studied; recovery value comes from this discrimination rather than from pressing harder.
  • Setting Fmax just below the breaking force is unsafe unless the controller's overshoot distribution is accounted for; conservative identity-derived budgets outperform an oracle budget that cheats by reading the hidden break threshold.
  • On-policy reinforcement learning is not a viable training method at 0.4 mm clearance for fragile parts, because the exploration noise needed to search the funnel already breaks the part; imitation learning with iterative relabeling plus weight interpolation is the working pipeline.
  • A learned release decision needs an input-skip linear head on raw observations; trained trunks provably discard the seat-state signal, so refining trunk features is the wrong path.

Where Pith is reading between the lines

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

  • If the force-signature separation holds on richer faults, failure diagnostics for assembly could be built entirely from force text tokens, avoiding vision pipelines for contact-stage errors.
  • The input-skip release-head result suggests a cheap diagnostic for any learned manipulation policy: probe raw observations with a linear classifier before training a refinement head, because trained trunks can discard the exact state a high-level decision needs.
  • A natural next experiment is to test the same frozen-LLM budget/recovery loop on a real robot with a contact-force kill switch; if the ~1.5× overshoot persists, budgets should be set from the overshoot distribution, not from Fbreak−ε.
  • The two failure faces of press-harder imply a design rule worth adopting elsewhere: any recovery policy that escalates actuation force should be considered unsafe for fragile assemblies unless the load path provably cannot engage.

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 / 5 minor

Summary. The paper proposes FORGE-plus, a two-layer force-budgeted assembly framework in which a frozen, text-only LLM sets a per-object force ceiling before execution and, on failure, selects a recovery maneuver from a fixed menu based on a compact force signature. The low-level controller enforces the ceiling through a hard clamp, and the recovery policy is forbidden from raising it. The hidden per-episode breaking force Fbreak is known only to the evaluator, making the breakage metric non-circular. The authors evaluate the framework in Isaac Lab rigid-body simulation on bottle placement and tight-clearance gear insertion with two grippers, reporting 256/256 clean gates on both fragile and robust classes, a learned release head, a 64/64 table-pick smoke flow, and a recovery sweep under injected in-grip slip. They also report negative results, including PPO failure at 0.4 mm clearance, collapse of a working policy under tiny-std PPO polish, and failures of three learned-release-head designs. The paper claims that the force-signature recovery chain resolves 40% and 64% of jams on the two grippers, whereas press-harder is futile or destructive, and that the identity-derived budget beats an oracle budget because it covers contact overshoot.

Significance. If accepted, the paper would make a useful contribution to contact-rich manipulation by separating semantic deliberation (LLM) from force authority (fast clamp), showing a concrete way to make force-budgeted recovery safe, and documenting strong negative results about on-policy RL at tight clearances. The design invariants are carefully stated: hidden Fbreak, immutable Fmax, and a fast-loop clamp. The 256/256 clean-gate results and the 96% breakage rate under press-harder on the Franka are strong point estimates. The paper is also commendably honest: it explicitly labels smoke tests, the vision baseline as a random-proxy, the simulation-only scope, and several unresolved caveats. The main concerns are statistical power of the recovery sweep and the lack of a true ablation of the LLM budget-setter, both of which are load-bearing for the headline claims.

major comments (3)
  1. [Table 3, §6.4] The paper's headline recovery-efficacy claim — 'worth a factor of ∼2 over selection-by-luck' (§8) — is supported only by 25-episode cells. On the 2F-140, ours vs. the vision-LLM proxy is 10/25 vs 7/25; a Fisher exact test gives p > 0.3, so the reported 40% vs 28% difference is within sampling noise. On the Franka, 16/25 vs 8/25 is borderline (p ≈ 0.02–0.05 uncorrected) and does not survive a multiple-comparison correction across five baselines and two grippers. The 12% breakage rates are 3/25 point estimates. The negative press-harder result (0% or 96% breaks) is a large effect and is not at risk, but the positive recovery-efficacy claim is load-bearing and underpowered. Please report confidence intervals, run larger N (simulation is cheap), or soften the factor-of-two claim.
  2. [§3.1, §4.1, §6.3] The role of the frozen LLM in setting the force ceiling is not empirically established. The budget-setter outputs are cached per class, and the only fragile/robust budgets used are 10 N and 100 N, which appear to have been chosen by the authors with knowledge of the Fbreak distributions (38±5 N and 180 N). There is no ablation against a hand-coded class-to-budget mapping, no report of the raw LLM outputs for the tested identities, and only two object classes. As written, the abstract's claim that a frozen LLM 'assigns a per-object force ceiling' is compatible with the LLM being a lookup table for manually tuned constants. Please add a non-LLM budget baseline (e.g., fixed heuristic from class/material tags) and report the LLM's actual outputs.
  3. [Table 2, §6.3] The 'Over-budget eps' column is internally inconsistent. The text defines clamp fidelity as contact overshoot above Fmax and stresses that clamping the command does not bound contact force. Yet the Ours row reports Fmax=10 N, peak contact force max 20.2 N, and over-budget episodes 0/200. If 'over-budget' counts contact force, this is impossible; if it counts the commanded force after clamping, it is trivially zero and cannot support the claim that a conservative budget 'absorbs' overshoot. Define the metric and report contact overshoot for each row.
minor comments (5)
  1. [§4.1, §6.1] The bottle class Fbreak is stated as ≈22 N in §4.1 and as 23 N / 23.3 N in §6.1. Distinguish the class-level distribution from per-episode draws to avoid apparent contradictions.
  2. [§5.2] The α selection for the weight soup is described as minimizing 'joint force tails' on the evaluation classes. Clarify whether the selection used held-out episodes, and how much selection over α could inflate the reported 256/256 clean-gate numbers.
  3. [Table 3] The note says rows are comparable within a table, not across tables, yet the abstract compares the two grippers' recovery rates (40% vs 64%). Add an explicit caution that different step caps and checkpoints separate the two halves of Table 3.
  4. [§6.2] The Franka clean gate uses 200/200 while the 2F-140 gate uses 256/256. State why the protocols differ; otherwise the different numbers look inconsistent.
  5. [§6.4] The 'vision-LLM proxy' is a random menu draw, as the paper honestly states. Since the paper's own claim is about beating selection-by-luck, this is acceptable; but the abstract/contribution wording should not imply a comparison against an actual vision-based system without repeating the proxy caveat.

Circularity Check

0 steps flagged

No significant circularity: the breakage metric is protected by hidden Fbreak, budgets are not constructed from Fbreak, and recovery success is empirical.

full rationale

Reviewing the derivation chain: (1) The breakage metric is explicitly protected by three design invariants (§3.2): Fbreak is visible only to the evaluator, Fmax is immutable during recovery, and force authority lives in the fast loop, with a runtime assertion guarding the signature encoder. This breaks the self-definitional loop that would arise if the agent could observe the threshold it is scored against. (2) The budget-setter maps object identity to Fmax; the paper never defines this mapping in terms of Fbreak, and because Fbreak is sampled per episode from a distribution not given to the LLM, the budget is a prediction rather than a renamed version of the hidden target. The oracle baseline (Fbreak−ε) is explicitly labeled as cheating and is used as a contrast, which further separates the measured result from the hidden variable. (3) The recovery-selector's advantage is not forced by the signature alone: the hand-coded heuristic consumes the same signature and fails (0% in Table 3), while the frozen LLM routes to regrasp; the comparison is therefore empirical, not a construction. (4) There are no load-bearing self-citations: references [1]–[17] are external prior work, and no uniqueness theorem by the present authors is invoked. (5) The paper's own limitations (§7) and honest caveats (§6.4) disclose that disturbances are injected and simple, the vision baseline is a proxy, and some results are smokes; these are generalization, statistical-power, and reproducibility caveats, not evidence that the outputs are equivalent to inputs by construction. The 25-episode recovery cells and the 10/25 vs 7/25 gap bear on the strength of the 'factor of ~2' claim, but that is a statistical-significance issue, not circularity. No step in the paper reduces to its own input, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The load-bearing 'free parameters' are the per-object force budgets and a handful of thresholds; the axioms are the simulation fidelity, the representativeness of injected faults, and the untested LLM budgeting capability.

free parameters (4)
  • Per-class force budget F_max = 10 N (fragile ABS gear), 8.8 N (fragile bottle), ~100 N (steel gear)
    Chosen by the design/LLM to be conservative relative to the class break threshold; not derived from first principles. The paper does not vary identity to test whether the LLM produces these values; they are cached constants.
  • Weight-soup interpolation coefficient α = 0.15
    Tuned on validation/triage; every α in [0.10,0.75] passes, so the exact value is not highly sensitive, but it is a fitted knob.
  • Release-head operating threshold and feature fold = set after standardization folding; exact value not reported
    Chosen to trade false positives/negatives; the paper describes the procedure but not the numeric threshold.
  • Jam-detector and hover-branch parameters = not reported
    Window lengths, contact-persist thresholds, and descent-window parameters are not specified numerically, and the recovery results depend on them.
axioms (4)
  • domain assumption Rigid-body simulation with breakage as a hidden scalar threshold on peak contact force is a valid proxy for real fragility
    The paper explicitly states there is no fracture/deformation modeling and no sim-to-real claim; the conclusions about safety are conditional on this model.
  • domain assumption The injected disturbance set (rim wedge, 5 mm in-grip slip) is representative of the failure modes the framework must handle
    Recovery is only evaluated on these two faults; richer jams are listed as future work.
  • ad hoc to paper The frozen LLM can set a sensible force ceiling from object identity text alone
    This is a core assumption of the budget-setter, but the paper does not show the LLM's raw outputs, test adversarial identities, or ablate the LLM against a hand-set budget.
  • ad hoc to paper Weight-soup interpolation is valid because bc7 was warm-started from bc3 (same basin)
    The paper argues this with empirical triage, but it is not proven for all α.

pith-pipeline@v1.3.0-alltime-deepseek · 197 in / 11151 out tokens · 149268 ms · 2026-08-01T08:04:45.814531+00:00 · methodology

0 comments
read the original abstract

Force-conditioned reinforcement learning (RL) enables tight-clearance assembly under a commanded force ceiling, but practical deployment requires determining an appropriate force limit for each object and recovering from insertion failures without exceeding it. We present a two-layer framework in which a frozen, text-only large language model (LLM) assigns a per-object force ceiling before execution and selects recovery maneuvers from a fixed action menu using compact textual force signatures. The LLM never controls force directly: a low-level controller enforces the force ceiling, the recovery policy cannot increase it, and the hidden breaking-force threshold is known only to the evaluator. We evaluate the framework on fragile bottle placement and 0.4 mm diametral-clearance gear insertion using two grippers (Robotiq 2F-140 and Franka Panda hand). A single policy passes 256/256 evaluation episodes on both fragile and robust objects without breakage, correctly predicts release timing, and completes a full table-pick-and-insert pipeline with a mean peak force of 5.4 N. Under injected in-grip slip, the force-signature recovery strategy resolves 40% and 64% of failures on the two grippers, whereas a press-harder baseline is either ineffective or causes frequent breakage. We also report negative results, including the failure of PPO to solve the task under strict force constraints and unsuccessful learned release strategies. All experiments are conducted in rigid-body simulation with hidden force-threshold breakage; no sim-to-real claim is made.

Figures

Figures reproduced from arXiv: 2607.21227 by Kyupaeck Jeff Rah, Midum Oh.

Figure 1
Figure 1. Figure 1: The two-layer architecture. The LLM only emits a number (Fmax) and a menu choice; the guarantee that commanded force never exceeds Fmax is produced by the clamp in the fast loop, backed by a fixed 120 N global hard cap and JSON range validation. The hidden Fbreak feeds only the evaluator’s breakage check. pick a recovery — with JSON in and JSON out, no images. The fast layer is the force-conditioned RL ski… view at source ↗
Figure 2
Figure 2. Figure 2: The rendered fragile recovery episode (Franka, glass-class bottle, Fbreak ≈ 23 N). The JAM card shows the actual signature the LLM saw — peak 16.1 N, net insert 1.6 mm, rising, lateral bias +x steady — its rotate_align decision, and the standing constraints: from the force signature alone (no vision); same Fmax — never press harder. operational-space controller — an OSC wrist-twist disturbance walked the e… view at source ↗
Figure 3
Figure 3. Figure 3: Stills from the released 2F-140 videos (HUD labels every phase learned/scripted with a live force gauge vs. Fmax and Fbreak). Left to right: the scripted-staging table pick (a real friction grasp of the resting gear); the learned release firing at 1.9 N after a clean insertion; a recovery episode mid-regrasp — the tilted gear has been placed back on the table for a physical re-pick (LLM decision wedge → re… view at source ↗
Figure 4
Figure 4. Figure 4: The recovery sweep of [PITH_FULL_IMAGE:figures/full_fig_p011_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

18 extracted references · 14 linked inside Pith

  1. [1]

    Factory: Fast contact for robotic assembly

    YashrajNarang, KierStorey, IretiayoAkinola, MilesMacklin, PhilippReist, LukaszWawrzyniak, Yunrong Guo, Adam Moravanszky, Gavriel State, Michelle Lu, Ankur Handa, and Dieter Fox. Factory: Fast contact for robotic assembly. InRobotics: Science and Systems (RSS), 2022. arXiv:2205.03532

    Pith/arXiv arXiv 2022

  2. [2]

    Lin, Iretiayo Akinola, Ankur Handa, Gaurav S

    Bingjie Tang, Michael A. Lin, Iretiayo Akinola, Ankur Handa, Gaurav S. Sukhatme, Fabio Ramos, Dieter Fox, and Yashraj Narang. IndustReal: Transferring contact-rich assembly tasks from simulation to reality. InRobotics: Science and Systems (RSS), 2023. arXiv:2305.17110

    Pith/arXiv arXiv 2023

  3. [3]

    Sukhatme, Fabio Ramos, and Yashraj Narang

    Bingjie Tang, Iretiayo Akinola, Jie Xu, Bowen Wen, Ankur Handa, Karl Van Wyk, Dieter Fox, Gaurav S. Sukhatme, Fabio Ramos, and Yashraj Narang. AutoMate: Specialist and generalist assembly policies over diverse geometries. InRobotics: Science and Systems (RSS),

  4. [4]

    FORGE: Force-guided explo- ration for robust contact-rich manipulation under uncertainty.IEEE Robotics and Automation Letters, 2025

    Michael Noseworthy, Bingjie Tang, Bowen Wen, Ankur Handa, Chad Kessens, Nicholas Roy, Dieter Fox, Fabio Ramos, Yashraj Narang, and Iretiayo Akinola. FORGE: Force-guided explo- ration for robust contact-rich manipulation under uncertainty.IEEE Robotics and Automation Letters, 2025. arXiv:2408.04587

    Pith/arXiv arXiv 2025

  5. [5]

    REFLECT: Summarizing robot experiences for failure explanation and correction

    Zeyi Liu, Arpit Bahety, and Shuran Song. REFLECT: Summarizing robot experiences for failure explanation and correction. InConference on Robot Learning (CoRL), 2023. arXiv:2306.15724

    Pith/arXiv arXiv 2023

  6. [6]

    DoReMi: Ground- ing language model by detecting and recovering from plan-execution misalignment.arXiv preprint arXiv:2307.00329, 2023

    Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Zheyuan Jiang, and Jianyu Chen. DoReMi: Ground- ing language model by detecting and recovering from plan-execution misalignment.arXiv preprint arXiv:2307.00329, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    AHA: A vision-language-model for detecting and reasoning over failures in robotic manipulation.arXiv preprint arXiv:2410.00371, 2024

    Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter Fox, Ajay Mandlekar, and Yijie Guo. AHA: A vision-language-model for detecting and reasoning over failures in robotic manipulation.arXiv preprint arXiv:2410.00371, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Tactile-VLA: Unlocking vision-language-action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160, 2025

    Jialei Huang, Shuo Wang, Fanqi Lin, Yihang Hu, Chuan Wen, and Yang Gao. Tactile-VLA: Unlocking vision-language-action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    PaCo-VLA: Passivity-shielded compliance prior for contact-rich vision-language- action manipulation.arXiv preprint arXiv:2606.00515, 2026

    Yifan Cao et al. PaCo-VLA: Passivity-shielded compliance prior for contact-rich vision-language- action manipulation.arXiv preprint arXiv:2606.00515, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation.arXiv preprint arXiv:2505.22159, 2025

    Jiawen Yu et al. ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation.arXiv preprint arXiv:2505.22159, 2025

    arXiv 2025

  11. [11]

    CompliantVLA-adaptor: VLM-guided variable impedance action for safe contact-rich manipulation.arXiv preprint arXiv:2601.15541, 2026

    Wei Zhang, Ying Huang, et al. CompliantVLA-adaptor: VLM-guided variable impedance action for safe contact-rich manipulation.arXiv preprint arXiv:2601.15541, 2026

    arXiv 2026

  12. [12]

    Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831, 2025

    Mayank Mittal et al. Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Orbit: A unified simulation framework for interactive robot learning environments.IEEE Robotics and Automation Letters, 2023

    Mayank Mittal, Calvin Yu, Qinxi Yu, Jingzhou Liu, Nikita Rudin, David Hoeller, Jia Lin Yuan, Ritvik Singh, Yunrong Guo, Hammad Mazhar, Ajay Mandlekar, Buck Babich, Gavriel State, Marco Hutter, and Animesh Garg. Orbit: A unified simulation framework for interactive robot learning environments.IEEE Robotics and Automation Letters, 2023. arXiv:2301.04195. 13

    Pith/arXiv arXiv 2023

  14. [14]

    GraspGen: A diffusion-based framework for 6-dof grasping.arXiv preprint arXiv:2507.13097, 2025

    Adithyavairavan Murali et al. GraspGen: A diffusion-based framework for 6-dof grasping.arXiv preprint arXiv:2507.13097, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

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

    Pith/arXiv arXiv 2017

  16. [16]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. InInternational Conference on Artificial Intelligence and Statistics (AISTATS), 2011. arXiv:1011.0686

    Pith/arXiv arXiv 2011

  17. [17]

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

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: Averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. InInternational Conference on Machine Learning (ICML),

  18. [2022]

    arXiv:2203.05482. 14

    Pith/arXiv arXiv