Reproduction of 'A General Framework for Fair and Robust Regression' (OpenReview R4W0oOjCKG): 3/3 claims verified.
Jansen Tang
ai-sherpa
AI & ML interests
None yet
Recent Activity
updated a Space 32 minutes ago
ai-sherpa/recursive-binding-subspace-carving-repro updated a Space about 4 hours ago
ai-sherpa/amortized-vi-partial-label-vipll-repro published a Space about 4 hours ago
ai-sherpa/amortized-vi-partial-label-vipll-reproOrganizations
Optimal Unconstrained Self-Distillation - Repro
Independent CPU reproduction of ridge self-distillation (arXiv 2602.17565): strict improvement, negative optimal weight, one-shot GCV tuning.
Attention & Frank-Wolfe — repro
Independent CPU reproduction (arXiv 2508.09628, orid zrn7rRuvhW): hardmax=Frank-Wolfe step, Voronoi exponential rate, finite-beta metastability.
SAM Stability Analysis — repro
Independent CPU reproduction of Stability Analysis of Sharpness-Aware Minimization (arXiv 2301.06308, orid tNvmMTsRqh).
Flat Minima and Generalization SCO repro
Independent reproduction of arXiv:2511.03548 (flat minima, SA-GD, SAM in smooth SCO).
Adam Convergence on Degenerate Polynomials repro
Independent numpy/mpmath reproduction of arXiv:2603.09581 (Adam local linear convergence + decoupling).
LatentMAS Repro (2511.20639)
Reproduction of LatentMAS (orid syG9I9ofd8): honest negative on accuracy + partial efficiency.
tau2-Bench Repro (2506.07982)
Reproduction of tau2-Bench dual-control claims (orids 71171 + OC2z7iSQKa).
Balancing Learning Rates Across Layers — repro
Independent CPU reproduction (arXiv 2606.00340, orid 4vztmTrGhd): exact 2-step closed-form dynamics + unequal-then-equal optimal layer-wise LRs.
MDS Robust NPE - ICML2026 repro
Independent CPU repro of Minimum Distance Summaries for Robust NPE (arXiv 2602.09161); 4/4 claims.
Success-Conditioning as Policy Improvement — repro
Independent CPU reproduction (arXiv 2601.18175, orid FEmXFeqYNZ): trust-region equivalence + three-way equality, exact to machine precision.
ICML-2026 Repro: Session-2 Inference Benchmarks
Reproductions of 5 inference-only ICML-2026 benchmark papers: VCG-Bench, MEMO, MedMosaic, CATArena, MORE.
- Running
Repro - CATArena
🎯View and collaborate on a project logbook with an AI agent
- Running
Repro - MORE
🌐Explore and share your project logbook with an AI agent
- Running
Repro - VCG-Bench
🎯Collaborate on a coding logbook with your AI agent
- Running
Repro - MedMosaic
🎯Collaborate with an AI agent to update your logbook
Provable Grokking in Ridge Regression - repro
Independent NumPy reproduction of arXiv 2601.19791 grokking claims (C1/C2/C3).
LM-CC Repro (2602.07882)
Partial reproduction of LM-CC (orid tI5CFbRhmV): metric implemented; correlation eval incomplete.
Repro - VLM-RobustBench (paper HwXyyvK7ZJ)
Independent reproduction of VLM-RobustBench (arXiv 2603.06148): logbook, code+results bundle.
RACO Reproduction (arXiv 2602.02495) — ICML-2026-agent-repro
Reproduction of RACO (Reward-free Alignment for Conflicting Objectives): logbook, code bundle, BeaverTails Pareto results (Qwen3-4B/1.7B).
Fair & Robust Regression - Reproduction
Reproduction of 'A General Framework for Fair and Robust Regression' (OpenReview R4W0oOjCKG): 3/3 claims verified.
Balancing Learning Rates Across Layers — repro
Independent CPU reproduction (arXiv 2606.00340, orid 4vztmTrGhd): exact 2-step closed-form dynamics + unequal-then-equal optimal layer-wise LRs.
Optimal Unconstrained Self-Distillation - Repro
Independent CPU reproduction of ridge self-distillation (arXiv 2602.17565): strict improvement, negative optimal weight, one-shot GCV tuning.
MDS Robust NPE - ICML2026 repro
Independent CPU repro of Minimum Distance Summaries for Robust NPE (arXiv 2602.09161); 4/4 claims.
Attention & Frank-Wolfe — repro
Independent CPU reproduction (arXiv 2508.09628, orid zrn7rRuvhW): hardmax=Frank-Wolfe step, Voronoi exponential rate, finite-beta metastability.
Success-Conditioning as Policy Improvement — repro
Independent CPU reproduction (arXiv 2601.18175, orid FEmXFeqYNZ): trust-region equivalence + three-way equality, exact to machine precision.
SAM Stability Analysis — repro
Independent CPU reproduction of Stability Analysis of Sharpness-Aware Minimization (arXiv 2301.06308, orid tNvmMTsRqh).
ICML-2026 Repro: Session-2 Inference Benchmarks
Reproductions of 5 inference-only ICML-2026 benchmark papers: VCG-Bench, MEMO, MedMosaic, CATArena, MORE.
- Running
Repro - CATArena
🎯View and collaborate on a project logbook with an AI agent
- Running
Repro - MORE
🌐Explore and share your project logbook with an AI agent
- Running
Repro - VCG-Bench
🎯Collaborate on a coding logbook with your AI agent
- Running
Repro - MedMosaic
🎯Collaborate with an AI agent to update your logbook
Flat Minima and Generalization SCO repro
Independent reproduction of arXiv:2511.03548 (flat minima, SA-GD, SAM in smooth SCO).
Provable Grokking in Ridge Regression - repro
Independent NumPy reproduction of arXiv 2601.19791 grokking claims (C1/C2/C3).
Adam Convergence on Degenerate Polynomials repro
Independent numpy/mpmath reproduction of arXiv:2603.09581 (Adam local linear convergence + decoupling).
LM-CC Repro (2602.07882)
Partial reproduction of LM-CC (orid tI5CFbRhmV): metric implemented; correlation eval incomplete.
LatentMAS Repro (2511.20639)
Reproduction of LatentMAS (orid syG9I9ofd8): honest negative on accuracy + partial efficiency.
Repro - VLM-RobustBench (paper HwXyyvK7ZJ)
Independent reproduction of VLM-RobustBench (arXiv 2603.06148): logbook, code+results bundle.
tau2-Bench Repro (2506.07982)
Reproduction of tau2-Bench dual-control claims (orids 71171 + OC2z7iSQKa).
RACO Reproduction (arXiv 2602.02495) — ICML-2026-agent-repro
Reproduction of RACO (Reward-free Alignment for Conflicting Objectives): logbook, code bundle, BeaverTails Pareto results (Qwen3-4B/1.7B).