Open list · 2025–2026 edition · updated July 2026
The 2025–2026 AI papers that most need code.
This is an early demo showing a sample of the list — the full version will be more comprehensive and structured. I'd rather show a handful I've verified than a big list I haven't.
How it's ranked: one list, by demand. Fresh 2026 papers (citations still ~0) are ranked by HuggingFace daily-paper upvotes; 2025 papers with real citation history are ranked by citation velocity (citations ÷ recency). Every entry is hand-verified to currently lack a public repo; proprietary / closed-lab reports are excluded.
Candidates come from LIVE 2025–2026 sources: a Semantic Scholar bulk-search across ~48 AI/ML subfields (year 2025–2026, sorted by citations) plus the HuggingFace daily-papers API harvested day-by-day across 2026 (capturing per-paper upvotes and any linked GitHub repo). Anything HuggingFace already links a repo for was dropped up front; the rest were hand-verified by parallel search agents — each paper checked on its arXiv page, GitHub, project pages, and Papers-with-Code, and classified has-code / no-code / exclude (survey, benchmark-only, or closed proprietary model report). Only papers confirmed to currently lack any official or community public repo were kept. Proprietary / closed-lab reports that the community cannot implement are excluded (e.g. GAIA-2 by Wayve, Tstars-Tryon by Alibaba, π*0.6 by Physical Intelligence). LIMITATIONS: 2026 citation counts are near-zero and noisy; HuggingFace upvotes are a hype-weighted, not quality-weighted, signal; "no public code" means none found via genuine search at verification time (2026-07-09) — several of the freshest entries are flagged "code may appear later," and some will.
Signal used: HuggingFace daily-paper upvotes for near-zero-citation 2026 work; citation velocity (Semantic Scholar citations ÷ recency) for 2025 papers with real citations. No fabricated or aggregate stats. Every entry hand-verified to lack public code.
The single most-upvoted 2026 daily paper without a repo: a looped/recurrent world-model architecture people want to reproduce, but no code is out.
Verified 2026-07: searched arXiv abs page, HF daily papers, GitHub and loop-model awesome-lists; no official/community repo found. HF upvotes=477 (top uncited-2026 demand signal). Uncertain: very new (Jun 2026), code may appear.
A high-citation CVPR 2025 method that injects visual chain-of-thought into VLA robot policies; a released implementation would let the community reproduce its planning gains.
Verified 2026-07: project page cot-vla.github.io has no code link; no repo on GitHub/Papers-with-Code (the similarly-named ACoT-VLA repo is a different paper). Real S2 citations (472).
Shows reasoning models internally 'know' when to stop, promising big inference savings, but the SAGE-RL training code is only shared on request.
Verified 2026-07: project page hzx122.github.io/sage-rl shows a placeholder 'Code' link; authors state code is available on request by email; not public. HF upvotes=266.
A collaborative multi-agent RL method for heterogeneous agents that drew heavy attention, but no training code exists.
Verified 2026-07: genuine Mar-2026 method paper (not an S2 merge); no repo findable on GitHub/HF/PwC. HF upvotes=198. Uncertain (recent).
A self-distillation twist on RLVR that reportedly stabilizes reasoning RL; no repo to verify the recipe.
Verified 2026-07: genuine Apr-2026 paper (not an S2 merge error); no repo findable. HF upvotes=179. Uncertain (recent).
Reframes RL training objectives to match the real (monotonic) inference policy; a reference implementation would let others test the claim.
Verified 2026-07: project page anitaleungxx.github.io/MIPU exists but links no code; none found on GitHub/PwC. HF upvotes=160. Uncertain (recent, may release later).
A SOTA value-based RL framework for long chain-of-thought reasoning that beats GRPO/DAPO, but ByteDance never open-sourced VAPO's value-model tricks.
Verified 2026-07: ByteDance Seed published the method but did not open-source VAPO; built atop verl/DAPO with no standalone official repo. Real S2 citations (229).
Extends RLVR to non-verifiable domains via LLM-graded rubrics; only the rubric datasets ship, not the training code practitioners want.
Verified 2026-07: Scale AI released only rubric datasets on HF; no official training-code repo (third-party RuscaRL merely converts the datasets). Real S2 citations (227).
A hierarchical VLA (ICML 2025) that follows open-ended language instructions in real time; released only as a project page and demos.
Verified 2026-07: only project page hi-robot-vla.github.io; no official or community code repository found. ICML 2025. Real S2 citations (211).
Adaptive test-time scaling that trades quality for latency in image editing; a fresh CVPR 2026 method with no repo yet.
Verified 2026-07: genuine CVPR-2026 paper (ADE-CoT), resolves correctly (not a merge error); no public repo found. HF upvotes=138.
A variance-adaptive advantage estimator for multi-reward RL that generalizes GRPO; no implementation released.
Verified 2026-07: genuine GRPO-variant paper, resolves correctly (not a merge error); no public repo found. HF upvotes=138.
A striking result that skipping explicit 'thinking' can match reasoning traces; the exact NoThinking setup has no official code.
Verified 2026-07: Papers-with-Code lists 'no code available'; no official repo for the NoThinking method found. Real S2 citations (~172).
Multimodal autoformalization of math 'in the wild'; a referenced project page but no confirmed code repo.
Verified 2026-07: MMFormalizer.github.io project page is referenced but no code repo confirmed on GitHub/HF/PwC. HF upvotes~=106. Uncertain.
A dual on-policy distillation objective for reasoning models; appears only in awesome-lists with no code.
Verified 2026-07: no repo on arXiv/HF/GitHub; appears only in awesome-lists. HF upvotes=103. Uncertain (recent).
Claims full attention can be transferred into sparse attention in ~100 training steps, promising cheap long-context; no code to reproduce it.
Verified 2026-07: sparse-attention transfer method (Nanjing Univ./Alibaba); no repo found on GitHub/HF/PwC. HF upvotes=98. Uncertain (recent).
Routes samples between group-relative (GRPO-style) and self-distillation policy optimization; a promising RL-unification idea with no released code.
Verified 2026-07 (user-suggested): SRPO/sample-routed policy optimization; no official repo found. HF upvotes=32. Uncertain (recent).
No papers match “”.