Pith.

Computer Vision and Pattern Recognition · arXiv:2611.22164 v1 · submitted 17 Nov 2026 · 99 references

HandRSI: Recursive Self-Improvement for Hand Pose Estimation

Zhekai Wang, Luqing Luo, Wenhao Yao, Junjie Lu, Huanan Zhu, Feifei Gao, Yuhua Jiang

The paper says

“Egocentric 3D hand pose estimators still miss hands at the image border, misplace occluded hands and jitter. Such failures are usually fixed by hand-written inference refinements and by fine-tuning, two interdependent steps that are tuned separately.”

more of the abstract

We present HandRSI, a recursive self-improvement framework for a hand pose system that consists of a video estimator and its executable inference harness. In each round, coding agents refine the harness, and a change is accepted only if it improves a composite of all ten benchmark metrics on held-out data without significantly regressing any of them. The accepted harness then guides the estimator through selective distillation, which supervises hand presence and temporal consistency on frames without ground truth, while ground truth supervises pose. Since the updated estimator may change which refinements remain useful, the harness is re-validated on it. Starting from the released ACE-Ego-Hand model, five rounds reduce MPJPE-p by 35% on HOT3D and 11% on ARCTIC and roughly halve jitter. On HOT3D, HandRSI's MPJPE-p is 16% lower than that of ACE-Ego-Hand fine-tuned on the same ground truth. Under a unified protocol, it outperforms nine re-run published methods on all ten metrics of both benchmarks. Hypervolume selection on the accuracy-speed front yields Lite, Medium and Heavy systems running at 20.7 to 41.1 FPS. This report gives the full method, every stage of every round and the complete protocol. Code, evaluator, harnesses and weights are public.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 18 Nov 2026 · read the full text (132,503 characters) · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. HandRSI delivers state-of-the-art egocentric hand pose by having coding agents iteratively improve the inference harness and distill the gains into the network, with a careful multi-metric gate that prevents any single metric from being traded off. This line is wrong

no objections raised

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 132,503 characters. Not checked: figures, code, what the 99 references say.

What would settle it

Verify that the released code and weights reproduce Table 1 row 'HandRSI (full)' on HOT3D test: run the provided evaluator with the final harness and weights on the same 1,162 segments; if MPJPE-p differs by more than 0.1 mm from the reported 8.37 mm, investigate the discrepancy.

from the skeptic's reading

How the review went

5 passes by deepseek-v4-flash · 18 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 7 s · 48,268 tokens read, 749 written
  3. Referee no objections raised 6 s · 48,452 tokens read, 280 written
  4. Check against the paper
  5. Editor wrote the verdict above 9 s · 97,913 tokens read, 445 written

Not run on this paper: the simulated authors' defence.

Pith checked this review's objections against the text of the paper it read.

The review in full

The review in brief, machine-written

This paper presents a well-engineered system that improves a hand pose estimator by recursively refining the code around it. The core idea is sound: coding agents write harness modules that post-process the network's outputs, a multi-metric gate (covering detection, pose, placement, and jitter) accepts only changes that improve the composite without regressing any single metric, and the accepted harness then serves as a teacher for selective distillation. The evaluation is thorough: nine baselines re-run under a unified protocol, ten metrics, three test sets, per-round controls, and ablation studies. The results are strong: on HOT3D and ARCTIC, HandRSI outperforms all re-run methods on every metric, and the gains compound over five rounds. The paper is clearly written, the method is described in detail, and the code and weights are released. The main soft spots are acknowledged by the authors: the operators consulted test scores for three decisions in rounds 1-2, which weakens the claim of a fully sealed test; the round-1 FAcc regression past the 5% cap was accepted because it was not statistically significant, but the cap is a hard rule in the gate definition; the evaluator is a re-implementation with a known discrepancy in EPE_2D-p; distillation saturates after round 2, and the paper does not fully explain why the student cannot improve further; and the HOI4D zero-shot results show that 2D error and detection lag behind ACE-Ego-Hand, which the paper attributes to camera differences but does not fully resolve. These are real concerns, but they are not fatal. The paper is a serious piece of work that deserves a rigorous peer review. The method is novel, the evaluation is careful, and the results are convincing. I recommend that the paper be sent to peer review.

The strongest objection

The paper presents a well-engineered system with thorough evaluation, public code and weights, and detailed per-round controls. The central claim that HandRSI improves hand pose estimation through recursive self-improvement is supported by extensive evidence. The authors acknowledge the main limitations: (1) operators consulted test scores for three decisions in rounds 1-2 (Sec. 4.4), which weakens the claim of a fully sealed test; (2) the round-1 FAcc regression past the 5% cap was accepted because it was not statistically significant; (3) the evaluator is a re-implementation with a known discrepancy in EPE_2D-p; (4) distillation saturates after round 2; (5) HOI4D zero-shot results show 2D error and detection lag behind ACE-Ego-Hand. These are all stated in the paper. The weakest assumption identified by the reader—that coding agents can write useful harness modules and the gate correctly identifies improvements—is supported by the evidence: 98 of 156 submissions passed the dev pre-check, 73 were accepted by the gate, and the per-round controls show that the harness alone improves the untouched base network. The gate mechanism (Eq. 3-5) is well-defined with bootstrap confidence intervals and per-metric regression checks. The paper's own data shows that the gate rejected submissions with significant composite gain when a single metric regressed (7 submissions), demonstrating it works as intended. I cannot identify a load-bearing concern that the paper does not already acknowledge and address.

The reader's verdict is CONDITIONAL with HIGH confidence, and the weakest assumption identified is that the coding agents can write useful harness modules and the gate correctly identifies improvements. I agree with this assessment. The paper provides substantial evidence supporting this assumption: 98 of 156 submissions passed the dev pre-check, 73 were accepted by the gate, and the per-round controls show that the harness alone improves the untouched base network. The gate mechanism is well-defined with bootstrap confidence intervals and per-metric regression checks. The paper's own data shows that the gate rejected submissions with significant composite gain when a single metric regressed (7 submissions), demonstrating it works as intended. The five concerns raised by the reader are all acknowledged by the authors in the paper: (1) operators consulted test scores for three decisions in rounds 1-2 (Sec. 4.4); (2) the round-1 FAcc regression past the 5% cap was accepted because it was not statistically significant; (3) the evaluator is a re-implementation with a known discrepancy in EPE_2D-p; (4) distillation saturates after round 2; (5) HOI4D zero-shot results show 2D error and detection lag behind ACE-Ego-Hand. These are not fatal and are addressed in the paper. The paper is thorough, the evaluation is comprehensive, and the code and weights are publicly released. I see no reason to change the verdict.

The referee's report

If the results hold, this is a significant contribution to egocentric 3D hand pose estimation. The recursive self-improvement framework is novel in the perception domain, and the thorough evaluation against nine re-run baselines under a unified protocol is a strength. The paper provides extensive controls and ablations. The release of code, evaluator, harnesses, and weights supports reproducibility.

Pith's reading of the paper (a machine's summary, not the authors')

The paper presents HandRSI, a framework in which coding agents recursively improve an egocentric 3D hand pose system consisting of a video estimator and an executable inference harness. In each round, agents write harness modules that are accepted only if they improve a composite of all ten benchmark metrics on held-out data without significantly regressing any single metric. The accepted harness then serves as a teacher for selective distillation, where the student learns pose from ground truth but takes hand presence and temporal consistency from the teacher. The harness is re-validated after each network update. Starting from the ACE-Ego-Hand model, five rounds reduce MPJPE-p by 35% on HOT3D and 11% on ARCTIC, roughly halve jitter, and outperform nine re-run published methods on all ten metrics of both benchmarks. The paper also selects accuracy-speed tiers (Lite, Medium, Heavy) via hypervolume optimization.

What carries the argument. The key machinery is the recursive self-improvement loop with three nested loops: Loop A evolves the inference harness via coding agents (Claude Opus 5.5 and GPT-6-sol) under a gate that requires improvement on a composite of ten metrics without significant regression on any single metric; Loop B distills the harness's gains into the network via selective distillation (pose from ground truth, presence and temporal behavior from the teacher); Loop C reflects on the archive to guide the next round. The gate mechanism (Eq. 3-5) uses a composite gain with per-metric clipping and a no-regression rule based on paired bootstrap confidence intervals.

Core claim. The paper claims that a recursive self-improvement loop, where coding agents refine an inference harness under a multi-metric no-regression gate and the harness's gains are selectively distilled into the network by information source, can substantially improve egocentric 3D hand pose estimation beyond what fine-tuning or classical post-processing achieves. The central claim is that this joint harness-network optimization, with the harness re-validated after each network update, compounds improvements across rounds: the harness alone improves the untouched base network, distillation transfers temporal stability, and the combination reaches state-of-the-art results on HOT3D and ARCTIC benchmarks.

The premise it rests on. The coding agents (Claude Opus 5.5 and GPT-6-sol) can write harness modules that improve the system, and the gate mechanism correctly identifies improvements without overfitting to the validation set, with the agents capped at 10 charged val queries per round.

If the paper is right (Pith's inferences, not the authors')

If HandRSI's approach generalizes, other 3D perception tasks (body pose, object pose, scene reconstruction) could be improved by similar recursive self-improvement loops with multi-metric gates.

The separation of supervision by information source (pose from GT, temporal behavior from teacher) could become a standard design pattern for distillation in perception systems.

The finding that distillation saturates after round 2 while the harness continues to improve suggests that for some refinements (placement, articulation fitted to image evidence), the network cannot absorb them without losing pose accuracy at a given LoRA budget.

The three deployment tiers (Lite, Medium, Heavy) show that most of the gain comes from cheap modules, with the WiLoR-based image evidence module providing the remaining pose gain at a speed cost.

The framework's reliance on specific coding agents (Claude Opus 5.5, GPT-6-sol) raises the question of whether the results transfer to other LLMs or future models; the paper's release of code and harnesses enables such tests.

The human-in-the-loop elements (operator approval for pipeline changes, test score consultation for three decisions in rounds 1-2) mean the system is not fully automated; a fully sealed version would require automating these decisions.

The saturation of distillation after round 2 suggests a fundamental limit on what can be transferred from a refined harness to a network at a given capacity; higher-capacity student architectures or different distillation objectives might extend the gains.

The zero-shot HOI4D results show that the harness can improve 3D pose and jitter on unseen cameras without retraining, but 2D alignment may degrade; this suggests a trade-off between temporal consistency and 2D accuracy that depends on camera parameters.

Check this review

What was read

The passes were given all 132,503 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): 0e7e0f8b9d2fbe7c25750b711a6fe7de87a3960e46eddb3bb2997d677b13dfb7

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.22164

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.0099 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0045 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0045 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Editor deepseek-v4-flash · medium effort · $0.0054 · instructions of the Editor, fingerprint e7211f6a3d27…
  5. Circularity check deepseek-v4-flash · medium effort · $0.0053 · instructions of the Circularity check, fingerprint e9946e39d67b…
  6. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0048 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.034 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.505d180dcaef

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 18 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): a0c8aaa4b39b9a8d54d38974570ac311e94970565dbab667974c54def3c3de36

curl -s 'https://pith.science/api/v1/verdicts/2611.22164?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Egocentric 3D hand pose estimators still miss hands at the image border, misplace occluded hands and jitter. Such failures are usually fixed by hand-written inference refinements and by fine-tuning, two interdependent steps that are tuned separately. We present HandRSI, a recursive self-improvement framework for a hand pose system that consists of a video estimator and its executable inference harness. In each round, coding agents refine the harness, and a change is accepted only if it improves a composite of all ten benchmark metrics on held-out data without significantly regressing any of them. The accepted harness then guides the estimator through selective distillation, which supervises hand presence and temporal consistency on frames without ground truth, while ground truth supervises pose. Since the updated estimator may change which refinements remain useful, the harness is re-validated on it. Starting from the released ACE-Ego-Hand model, five rounds reduce MPJPE-p by 35% on HOT3D and 11% on ARCTIC and roughly halve jitter. On HOT3D, HandRSI's MPJPE-p is 16% lower than that of ACE-Ego-Hand fine-tuned on the same ground truth. Under a unified protocol, it outperforms nine re-run published methods on all ten metrics of both benchmarks. Hypervolume selection on the accuracy-speed front yields Lite, Medium and Heavy systems running at 20.7 to 41.1 FPS. This report gives the full method, every stage of every round and the complete protocol. Code, evaluator, harnesses and weights are public.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-18

The authors' main claim, as the classifier picked it out 0.97

“Under a unified protocol, it outperforms nine re-run published methods on all ten metrics of both benchmarks.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.16 · human participants 0.07 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.99

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Egocentric 3D hand pose estimators still miss hands at the image border, misplace occluded hands and jitter. background 0.89 0.15 0.02 asserted 1.00
Such failures are usually fixed by hand-written inference refinements and by fine-tuning, two interdependent steps that are tuned separately. background 0.87 0.07 0.02 asserted 0.83
We present HandRSI, a recursive self-improvement framework for a hand pose system that consists of a video estimator and its executable inference harness. contribution 0.98 0.85 0.07 asserted 1.00
In each round, coding agents refine the harness, and a change is accepted only if it improves a composite of all ten benchmark metrics on held-out data without significantly regressing any of them. method 0.99 0.90 0.05 asserted 0.99
The accepted harness then guides the estimator through selective distillation, which supervises hand presence and temporal consistency on frames without ground truth, while ground truth supervises pose. method 1.00 0.87 0.05 asserted 1.00
Since the updated estimator may change which refinements remain useful, the harness is re-validated on it. method 1.00 0.66 0.03 hedged possibility 0.77
Starting from the released ACE-Ego-Hand model, five rounds reduce MPJPE-p by 35% on HOT3D and 11% on ARCTIC and roughly halve jitter. result 1.00 0.96 0.27 asserted 1.00
On HOT3D, HandRSI's MPJPE-p is 16% lower than that of ACE-Ego-Hand fine-tuned on the same ground truth. result 1.00 0.96 0.18 asserted 1.00
Under a unified protocol, it outperforms nine re-run published methods on all ten metrics of both benchmarks. result 1.00 0.94 0.35 asserted 1.00
Hypervolume selection on the accuracy-speed front yields Lite, Medium and Heavy systems running at 20.7 to 41.1 FPS. result 0.89 0.94 0.06 asserted 1.00
This report gives the full method, every stage of every round and the complete protocol. contribution 0.95 0.66 0.03 asserted 0.99
Code, evaluator, harnesses and weights are public. contribution 0.98 0.67 0.03 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-18: 1 request, 8,884 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (17)
Figure 1 of the paper
Figure 1. HandRSI at a glance. Left: HandRSI’s predictions on frames of the sealed HOT3D and ARCTIC tests under occlusion, interaction and fast motion; below, the MANO hands of each top frame from a novel view. Center: coding agents refine the inference harness and distill it into the estimator in nested feedback loops; Lite, Medium and Heavy systems are selected after the loop. Right: HandRSI against ACE-Ego-Hand [45] and the strongest re-run baselines on the sealed HOT3D and ARCTIC tests; each axis scores a method relative to the best one (100%). Abstract Egocentric 3D hand pose estimators…
Figure 2 of the paper
Figure 2. Overview of one HandRSI round. Loop A evolves the inference harness under the gate GA; Loop B distills the refined harness into the estimator and re-validates it on the new network; the round gate G1% accepts the new system on val; Loop C turns the experimental archive into guidance for the next round. that are scored on test for reporting only. Round k of the tables and figures is the step from Zk−1 to Zk. 3.1. Problem Formulation System and round state. Given a clip X = {It} T t=1 and intrinsics K, the estimator fθ predicts per frame and hand a presence score, MANO [68]…
Figure 3 of the paper
Figure 3. Re-validation after a model update (schematic). With θk+1 fixed, paired dev ablations on the same clips reassess every module of H I,⋆ k as useful, redundant or harmful; the weak mod￾ules are removed jointly, and the pruned chain is kept only if it passes the benchmark’s rule, otherwise H I,⋆ k is kept. Modules A– C illustrate one possible outcome. Re-validation. H I,⋆ k was written for θk, and on the new network some of its modules may correct errors that are gone (
Figure 4 of the paper
Figure 4. Top: 3D PCK on the sealed tests (wrist-aligned; missed hands count as errors); legend: area under the curve up to 50 mm, HOT3D / ARCTIC; shaded: HandRSI’s margin over ACE-Ego-Hand. Bottom: temporal stability on one HOT3D test clip, image position of the index fingertip and per-frame joint acceleration: ACE-Ego-Hand’s acceleration peaks every four frames, HandRSI follows the ground truth without these steps (clip jitter 4.83→1.70 mm/frame2 ). missed on-screen hand is charged a canonical MANO [68] hand at the camera origin; PA-p aligns each matched hand to the GT by the least-squares
Figure 5 of the paper
Figure 5. Comparison with the strongest baselines on the sealed tests (same crop per row; white outline: ground truth; tag: 2D error in that frame). The baselines miss occluded or close hands and misplace interacting ones; HandRSI recovers and aligns them. that converts a crop camera to the full image receives the clip’s GT intrinsics; HaWoR runs its camera-space front end with SLAM and infilling off, EgoForce runs its causal Kalman filter [30], and InterWild uses its egocentric checkpoint (trained with the egocentric cameras of ReInter￾Hand [53]). ACE-Ego-Hand’s K-free row takes existence…
Figure 6 of the paper
Figure 6. Two HOT3D test clips with the network alone: two frames (cropped to the hand that ACE-Ego-Hand gets most wrong) and the clip’s hands in camera space with wrist trails (ground truth in gray). Below: the clip’s MPJPE-p and EPE2D-p. ACE-Ego-Hand misses the hand at the image border (left) and misplaces the occluded hand (right); the HandRSI network recovers both. gates, Loop B training, re-validation and the tier search), the final harness and the three tiers, our re-implementation of the ACE-Ego-Hand evaluator, the runners of the pub￾lished baselines and of the classical and ensemble…
Figure 7 of the paper
Figure 7. Recursive improvement on the sealed test sets: the system Sk, its network θk, and the round’s harness on the original network (C 2 k); round 0 is ACE-Ego-Hand; val gains in Tab. 6. Ground truth ACE-Ego-Hand S1 S2 S3 S4 S5 (HandRSI) frame 24 frame 51 error per-frame error (0–40 mm) missed MPJPE-p 47.0 mm 28.7 mm 16.2 mm 15.6 mm 15.0 mm 14.9 mm
Figure 8 of the paper
Figure 8. One HOT3D test clip after every round. Top: two frames of the right hand (same crop per column); bottom: its wrist-aligned joint error in each of the 81 frames (red: missed; ticks: the frames above) and the clip’s MPJPE-p. ACE-Ego-Hand misses the hand in 22 frames and S1 in 8; from round 2 on it is always found, and later rounds lower its error on the detected frames (14.4→12.5 mm)
Figure 10 of the paper
Figure 10. The harness over the rounds: modules of each ac￾cepted HI k by function (from each module’s own description); above the bars, modules added / removed relative to the previous round. adds, placement and articulation fitted to image evidence, is what Tab. 9 shows a student cannot take from its teacher without losing pose. Every stage of every round. Appendix A gives every stage of every round on the three test sets, with the GT￾only control, the teacher and the harness on the original network (Tabs. 14–16). In every round the system lowers −10 −8 −6 −4 −2 0 0 1 2 3 τ = 0.3% −5% −2%…
Figure 11 of the paper
Figure 11. Every Loop A gate decision on val (the 98 of 156 submissions that passed the dev pre-check, five rounds): compos￾ite gain against the agent’s incumbent and the worst single-metric change. Dashed: τ ; dotted: the regression thresholds (rm < −2% when significant, rm < −5% always). Two rejections beyond −10% are drawn at the border
Figure 12 of the paper
Figure 12. All 124 searched configurations on val against the summed per-module cost (an estimate that overstates WiLoR about 2.5×;
Figure 13 of the paper
Figure 13. Decoded sequence length: network alone on HOT3D test against the length decoded in one pass (dashed: 81-frame seg￾ments). [97] Eckart Zitzler and Lothar Thiele. Multiobjective Evolution￾ary Algorithms: A Comparative Case Study and the Strength Pareto Approach. IEEE Trans. Evol. Comput., 3(4):257–271, 1999. 3, 6 A. Complete Results Every stage of every round. Tables 14–16 give every stage of every round on the three test sets: the GT-only con￾trol C 1 k , the teacher, the network θk, the system Sk, and the round’s harness on the original network (C 2 k ). In the block of round k,…
Figure 14 of the paper
Figure 14. Extended comparison on the sealed HOT3D test. Per example: the frame with each method’s meshes (white outline: ground truth; tag: mean 2D joint error of the visible hands, “missed” if one of them is not output) and the same hands in 3D from one third-person view above a floor, identical in every column. Frames: the two with the largest gap between the baselines’ and HandRSI’s 2D error, then the two largest among frames where every method outputs every visible hand; one per recording. 22
Figure 15 of the paper
Figure 15. Extended comparison on the sealed ARCTIC test (held-out subject s05), as in
Figure 16 of the paper
Figure 16. Extended comparison on HOI4D, zero-shot for every method, as in
Figure 17 of the paper
Figure 17. Three more HOT3D test clips after every round (as

Showing 16 of the paper's 17 figures. Figures are the authors', reproduced from the paper.

References (99)
  1. [1] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, et al. HOT3D: Hand and Object Track- ing in 3D from Egocentric Multi-View Videos. InCVPR, on Pith
  2. [2] Avrim Blum and Moritz Hardt. The Ladder: A Reliable Leaderboard for Machine Learning Competitions. InICML, on Pith
  3. [3] Adnane Boukhayma, Rodrigo de Bem, and Philip H. S. Torr. 3D Hand Shape and Pose from Images in the Wild. InCVPR, on Pith
  4. [4] Karl Bringmann, Tobias Friedrich, and Patrick Klitzke. Two- Dimensional Subset Selection for Hypervolume and Epsilon- Indicator. InGenetic and Evolutionary Computation Confer- ence, pages 589–596, 2014. 7 on Pith
  5. [5] Jinkun Cao. Fast and Accurate Hand Estimation in Pixels, Interaction and World Space.arXiv preprint arXiv:2611.07711, 2026. 2 on PithPith's reviewarXiv
  6. [6] G ´ery Casiez, Nicolas Roussel, and Daniel V ogel. 1C Filter: A Simple Speed-Based Low-Pass Filter for Noisy Input in Interactive Systems. InCHI, pages 2527–2530, 2012. 14 on Pith
  7. [7] Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, et al. MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering. InICLR, 2025. 3 on Pith
  8. [8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, et al. DexYCB: A Benchmark for Capturing Hand Grasping of Objects. InCVPR, 2021. 2 on Pith
  9. [9] Xingyu Chen, Zhuheng Song, Xiaoke Jiang, Yaoqing Hu, Junzhi Yu, and Lei Zhang. HandOS: 3D Hand Reconstruc- tion in One Stage. InCVPR, 2025. 2 on Pith
  10. [10] Zhengyu Chen, Teng Xiao, Huaisheng Zhu, Yige Yuan, Luan Zhang, and Jingang Wang. Co-Harness: Co-Evolving Har- nesses and Model Weights for LLM Agents.arXiv preprint arXiv:2607.22688, 2026. 2, 3 on PithPith's reviewarXiv
  11. [11] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Ky- oung Mu Lee. Beyond Static Features for Temporally Con- sistent 3D Human Pose and Shape from a Video. InCVPR, on Pith
  12. [12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, et al. Rescaling Egocentric Vision: Col- lection, Pipeline and Challenges for EPIC-KITCHENS-100. IJCV, 130(1):33–55, 2022. 2 on Pith
  13. [13] Haoye Dong, Aviral Chharia, Wenbo Gou, Francisco Vi- cente Carrasco, and Fernando De la Torre. Hamba: Single-view 3D Hand Reconstruction with Graph-guided Bi- Scanning Mamba. InNeurIPS, 2024. 2, 10, 21 on Pith
  14. [14] Enes Duran, Muhammed Kocabas, Vasileios Choutas, Zi- cong Fan, and Michael J. Black. HMP: Hand Motion Priors for Pose and Shape Estimation From Video. InWACV, pages 6353–6363, 2024. 2 on Pith
  15. [15] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Aaron Roth. The Reusable Holdout: Preserving Validity in Adaptive Data Analysis.Sci- ence, 349(6248):636–638, 2015. 3, 4, 5 on Pith
  16. [16] Bradley Efron and Robert J. Tibshirani.An Introduction to the Bootstrap. Chapman and Hall/CRC, 1994. 4, 5 on Pith
  17. [17] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, et al. ARCTIC: A Dataset for Dexterous Bimanual Hand-Object Manipulation. InCVPR, 2023. 2, 7 on Pith
  18. [18] Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, et al. Benchmarks and Challenges in Pose Estimation for Egocen- tric Hand Interactions with Objects. InECCV, 2024. 2 on Pith
  19. [19] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M. Kitani. Deformer: Dynamic Fusion Transformer for Robust Hand Pose Estimation. InICCV, 2023. 2 on Pith
  20. [20] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, et al. 3D Hand Shape and Pose Estimation from a Single RGB Image. InCVPR, 2019. 2 on Pith
show 79 more
  1. [21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, et al. Ego4D: Around the World in 3,000 Hours of Egocentric Video. InCVPR, 2022. 2 on Pith
  2. [22] Tanmay Gupta and Aniruddha Kembhavi. Visual Program- ming: Compositional Visual Reasoning without Training. In CVPR, 2023. 3 on Pith
  3. [23] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vin- cent Lepetit. HOnnotate: A Method for 3D Annotation of Hand and Object Poses. InCVPR, 2020. 2 on Pith
  4. [24] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. LoRA+: Ef- ficient Low Rank Adaptation of Large Models. InICML, on Pith
  5. [25] Bingxiang He, Pengqi Sun, Shihao Peng, Shaoxuan Li, et al. AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning.arXiv preprint arXiv:2610.22194, 2026. 3 on PithPith's reviewarXiv
  6. [26] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531, 2015. 3 on PithPith's reviewarXiv
  7. [27] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, et al. LoRA: Low-Rank Adaptation of Large Language Models. InICLR, 2022. 3, 5, 6 on Pith
  8. [28] Peter J. Huber. Robust Estimation of a Location Parameter. The Annals of Mathematical Statistics, 35(1):73–101, 1964. 5 on Pith
  9. [29] Hisao Ishibuchi, Ryo Imada, Yu Setoguchi, and Yusuke No- jima. How to Specify a Reference Point in Hypervolume Calculation for Fair Performance Comparison.Evolutionary Computation, 26(3):411–440, 2018. 7 on Pith
  10. [30] Rudolph E. Kalman. A New Approach to Linear Filtering and Prediction Problems.Journal of Basic Engineering, 82 (1):35–45, 1960. 9 16 on Pith
  11. [31] Sayash Kapoor, Benedikt Stroebl, Zachary S. Siegel, Nitya Nadgir, and Arvind Narayanan. AI Agents That Matter. arXiv preprint arXiv:2407.01502, 2024. 3 on PithPith's reviewarXiv
  12. [32] Haechan Kim, Yoonho Lee, Gisang Lee, Chelsea Finn, and Kangwook Lee. WHALE: A Simple Recipe for Joint Harness-Weight Optimization.arXiv preprint arXiv:2609.00196, 2026. 2, 3 on PithPith's reviewarXiv
  13. [33] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to Reconstruct 3D Human Pose and Shape via Model-fitting in the Loop. InICCV, 2019. 3 on Pith
  14. [34] Tobias Kuhn, Carlos M. Fonseca, Lu ´ıs Paquete, Stefan Ruzika, Miguel M. Duarte, and Jos ´e Rui Figueira. Hyper- volume Subset Selection in Two Dimensions: Formulations and Algorithms.Evolutionary Computation, 24(3):411–425, on Pith
  15. [35] Taein Kwon, Bugra Tekin, Jan St ¨uhmer, Federica Bogo, and Marc Pollefeys. H2O: Two Hands Manipulating Objects for First Person Interaction Recognition. InICCV, 2021. 2 on Pith
  16. [36] Marco Laumanns, Lothar Thiele, Kalyanmoy Deb, and Eckart Zitzler. Combining Convergence and Diversity in Evolutionary Multiobjective Optimization.Evolutionary Computation, 10(3):263–282, 2002. 3, 7 on Pith
  17. [37] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta-Harness: End- to-End Optimization of Model Harnesses.arXiv preprint arXiv:2603.28052, 2026. 2, 3 on PithPith's reviewarXiv
  18. [38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring Plain Vision Transformer Backbones for Object Detection. InECCV, 2022. 8 on Pith
  19. [39] Yu Li, Chenyang Shao, Xinyang Liu, et al. AutoSOTA: An End-to-End Automated Research System for State-of-the- Art AI Model Discovery.arXiv preprint arXiv:2604.05550, on PithPith's reviewarXiv
  20. [40] Dixuan Lin, Yuxiang Zhang, Mengcheng Li, Yebin Liu, et al. OmniHands: Towards Robust 4D Hand Mesh Recovery via A Versatile Transformer.arXiv preprint arXiv:2405.20330, on PithPith's reviewarXiv
  21. [41] Zihao Lin, Haoyu Dong, Haobo Xu, Chi-Chih Chang, et al. VideoHarness-SRSI: Structured Recursive Self- Improvement for Video Understanding Harness.arXiv preprint arXiv:2610.19424, 2026. 2, 3 on PithPith's reviewarXiv
  22. [42] Changzhi Liu, Yilun Liu, Sikuan Yan, V olker Tresp, and Yunpu Ma. Mendel G ¨odel Machine: Recursive Self- Improving Coding Agents via Comparative Evolution.arXiv preprint arXiv:2608.07645, 2026. 3 on PithPith's reviewarXiv
  23. [43] Ruicong Liu, Takehiko Ohkawa, Mingfang Zhang, and Yoichi Sato. Single-to-Dual-View Adaptation for Egocen- tric 3D Hand Pose Estimation. InCVPR, pages 677–686, on Pith
  24. [44] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, et al. HOI4D: A 4D Egocentric Dataset for Category-Level Human-Object Interaction. InCVPR, 2022. 2, 7 on Pith
  25. [45] Yufei Liu, Xixi Wang, Hao Li, Ganlong Zhao, et al. ACE-Ego-Hand: Repurposing Video Diffusion Models for Occlusion-Robust Egocentric 3D Hand Motion Recovery. arXiv preprint arXiv:2608.20308, 2026. 1, 2, 4, 7, 8, 9, 10, 21, 22, 23, 24 on PithPith's reviewarXiv
  26. [46] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic Gradi- ent Descent with Warm Restarts. InICLR, 2017. 6 on Pith
  27. [47] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. InICLR, 2019. 6 on Pith
  28. [48] Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.arXiv preprint arXiv:2408.06292, 2024. 3 on PithPith's reviewarXiv
  29. [49] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hal- linan, et al. Self-Refine: Iterative Refinement with Self- Feedback. InNeurIPS, 2023. 3 on Pith
  30. [50] Christen Millerdurai, Shaoxiang Wang, Yaxu Xie, Vladislav Golyanik, Didier Stricker, and Alain Pagani. EgoForce: Forearm-Guided Camera-Space 3D Hand Pose from a Monocular Egocentric Camera. InACM SIGGRAPH, 2026. 2, 8, 10, 21, 22, 23, 24 on Pith
  31. [51] Gyeongsik Moon. Bringing Inputs to Shared Domains for 3D Interacting Hands Recovery in the Wild. InCVPR, 2023. 2, 10, 21 on Pith
  32. [52] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A Dataset and Base- line for 3D Interacting Hand Pose Estimation from a Single RGB Image. InECCV, 2020. 2 on Pith
  33. [53] Gyeongsik Moon, Shunsuke Saito, Weipeng Xu, Rohan Joshi, et al. A Dataset of Relighted 3D Interacting Hands. InNeurIPS Datasets and Benchmarks Track, 2023. 2, 9 on Pith
  34. [54] Seungjun Moon, Subin Jeon, Sangwoo Kim, Hanbyul Joo, and Jinwoo Shin. RLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning. arXiv preprint arXiv:2610.09455, 2026. 2 on PithPith's reviewarXiv
  35. [55] Wiktor Mucha, Michael Wray, and Martin Kampel. Towards Egocentric 3D Hand Pose Estimation in Unseen Domains. In WACV, 2026. 2, 3 on Pith
  36. [56] Alexander Novikov, Ng ˆan V ˜u, Marvin Eisenberger, Em- ilien Dupont, et al. AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery.arXiv preprint arXiv:2506.13131, 2025. 3 on PithPith's reviewarXiv
  37. [57] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. AssemblyHands: Towards Egocentric Activity Understanding via 3D Hand Pose Esti- mation. InCVPR, 2023. 2 on Pith
  38. [58] Zhuoshi Pan, ZhiYuan Feng, Yukuo Cen, Honglin Lin, et al. HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs.arXiv preprint arXiv:2610.26770, 2026. 2, 3 on PithPith's reviewarXiv
  39. [59] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstruct- ing Hands in 3D with Transformers. InCVPR, 2024. 2, 10, 21 on Pith
  40. [60] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. InICCV, 2023. 6 on Pith
  41. [61] Boris T. Polyak and Anatoli B. Juditsky. Acceleration of Stochastic Approximation by Averaging.SIAM Journal on Control and Optimization, 30(4):838–855, 1992. 6 on Pith
  42. [62] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. WiLoR: End-to-end 3D Hand Localization and Reconstruction in-the-wild. InCVPR, on Pith
  43. [63] Aditya Prakash, Ruisen Tu, Matthew Chang, and Saurabh Gupta. 3D Hand Pose Estimation in Everyday Egocentric Images. InECCV, 2024. 2, 10, 21 on Pith
  44. [64] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, et al. DexMV: Imitation Learning for Dexterous Manipu- lation from Human Videos. InECCV, 2022. 2 on Pith
  45. [65] Kerui Ren, Kaiwen Song, Weiguang Zhao, Yuxi Wang, et al. InfiniHand: Streaming World-Space Hand Mo- tion Estimation from Egocentric Video.arXiv preprint arXiv:2609.35743, 2026. 2 on PithPith's reviewarXiv
  46. [66] Patrick Rim, Kevin Harris, Braden Copple, Shangchen Han, et al. SHOW3D: Capturing Scenes of 3D Hands and Objects in the Wild. InCVPR, 2026. 2 on Pith
  47. [67] Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, et al. Mathematical Dis- coveries from Program Search with Large Language Models. Nature, 625(7995):468–475, 2024. 3 on Pith
  48. [68] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied Hands: Modeling and Capturing Hands and Bod- ies Together.ACM TOG, 36(6), 2017. 2, 4, 7 on Pith
  49. [69] Abraham Savitzky and Marcel J. E. Golay. Smoothing and Differentiation of Data by Simplified Least Squares Proce- dures.Analytical Chemistry, 36(8):1627–1639, 1964. 14 on Pith
  50. [70] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F. Fouhey. Understanding Human Hands in Contact at Inter- net Scale. InCVPR, 2020. 2 on Pith
  51. [71] Kenneth Shaw, Shikhar Bahl, and Deepak Pathak. VideoDex: Learning Dexterity from Internet Videos. In CoRL, 2022. 2 on Pith
  52. [72] Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflex- ion: Language Agents with Verbal Reinforcement Learning. InNeurIPS, 2023. 3, 6 on Pith
  53. [73] Chen Si, Yulin Liu, Bo Ai, Jianwen Xie, et al. AnyHand: A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Esti- mation.arXiv preprint arXiv:2603.25726, 2026. 2 on PithPith's reviewarXiv
  54. [74] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand Keypoint Detection in Single Images using Multiview Bootstrapping. InCVPR, 2017. 3 on Pith
  55. [75] Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, and Zhengyao Jiang. Recursive Self-Improvement of AI Re- search Agents.arXiv preprint arXiv:2609.26457, 2026. 3 on PithPith's reviewarXiv
  56. [76] Zihan Tan, Leixin Sun, Zitong Shi, et al. MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recur- sive Self-Improving Systems Themselves.arXiv preprint arXiv:2609.06396, 2026. 3 on PithPith's reviewarXiv
  57. [77] Shinji Umeyama. Least-Squares Estimation of Transforma- tion Parameters Between Two Point Patterns.IEEE TPAMI, 13(4):376–380, 1991. 7 on Pith
  58. [78] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, et al. MimicPlay: Long-Horizon Imitation Learning by Watching Human Play. InCoRL, 2023. 2 on Pith
  59. [79] Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, et al. OpenHands: An Open Platform for AI Software Developers as Generalist Agents. InICLR, 2025. 3 on Pith
  60. [80] Yuxi Wang, Chengkai Jin, Yufei Liu, Wenqi Ouyang, et al. The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction.arXiv preprint arXiv:2606.30308, 2026. 2, 9 on PithPith's reviewarXiv
  61. [81] Siwei Wu, Jincheng Ren, Yizhi Li, Haau-Sing Li, et al. Mod- ularRSI: Modular and Generalizable Recursive Harness Self- Improvement.arXiv preprint arXiv:2609.14857, 2026. 3 on PithPith's reviewarXiv
  62. [82] Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, et al. RRSI: Regularized Recursive Self-Improvement of Agent Harnesses.arXiv preprint arXiv:2609.24972, 2026. 2, 3 on PithPith's reviewarXiv
  63. [83] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V . Le. Self-Training with Noisy Student Improves ImageNet Classification. InCVPR, 2020. 3 on Pith
  64. [84] Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu, Bhavana Dalvi Mishra, et al. Harness 2: Recursive Agent Harnessing for an Open World.arXiv preprint arXiv:2610.14224, 2026. 3 on PithPith's reviewarXiv
  65. [85] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation. InNeurIPS, 2022. 8 on Pith
  66. [86] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose++: Vision Transformer for Generic Body Pose Es- timation.IEEE TPAMI, 46(2):1212–1230, 2024. 8 on Pith
  67. [87] John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, et al. SWE-agent: Agent-Computer Interfaces En- able Automated Software Engineering. InNeurIPS, 2024. 3 on Pith
  68. [88] Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, and Guojie Song. Harness-Zero: Harness Distillation via Agent- as-Harness.arXiv preprint arXiv:2609.24974, 2026. 2, 3 on PithPith's reviewarXiv
  69. [89] Yufei Ye, Yao Feng, Omid Taheri, Haiwen Feng, Shubham Tulsiani, and Michael J. Black. Predicting 4D Hand Trajec- tory from Monocular Videos. In3DV, 2026. 2 on Pith
  70. [90] Zhengdi Yu, Stefanos Zafeiriou, and Tolga Birdal. Dyn- HaMR: Recovering 4D Interacting Hand Motion from a Dy- namic Camera. InCVPR, 2025. 9 on Pith
  71. [91] Huajian Zeng, Chaohua Yao, Yuantai Zhang, Jiaqi Yang, Rolandos Alexandros Potamias, and Xingxing Zuo. Sta- bleHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric Video. In NeurIPS, 2026. To appear on Pith
  72. [92] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolan- dos Alexandros Potamias. HaWoR: World-Space Hand Mo- tion Reconstruction from Egocentric Videos. InCVPR, 2025. 2, 7, 8, 9, 10, 21, 22, 23, 24 on Pith
  73. [93] Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin G ¨odel Machine: Open-Ended Evolution of Self-Improving Agents. InICLR, 2026. 2, 3 on Pith
  74. [94] Zijie Zhu, Weiren Cai, Yizhou Wang, Zhenjie Yang, et al. MINT: A Unified Model for World-Space Camera and Hand Motion Estimation from Scalable Egocentric Pipeline Super- vision.arXiv preprint arXiv:2609.04958, 2026. 2 on PithPith's reviewarXiv
  75. [95] Christian Zimmermann and Thomas Brox. Learning to Es- timate 3D Hand Pose from Single RGB Images. InICCV, on Pith
  76. [96] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max J. Argus, and Thomas Brox. FreiHAND: A Dataset for Markerless Capture of Hand Pose and Shape from Single RGB Images. InICCV, 2019. 2 18 20 50 100 200 500 1k 2k 5k 0 2 4 Lite Medium Heavy S5 summed module cost p... on Pith
  77. [97] Eckart Zitzler and Lothar Thiele. Multiobjective Evolution- ary Algorithms: A Comparative Case Study and the Strength Pareto Approach.IEEE Trans. Evol. Comput., 3(4):257–271, on Pith
  78. [1999] 3, 6 A. Complete Results Every stage of every round.Tables 14–16 give every stage of every round on the three test sets: the GT-only con- trolC 1k, the teacher, the networkθ k, the systemS k, and the round’s harness on the original network (C2k). In the block of roundk, the te... on Pith
  79. [2025] 2, 5, 8, 9, 10, 14, 15, 21, 22, 23, 24 17 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment