1/ Math future mood board: Littmath says frontier models are already superhuman on some prestigious math tasks, and Tim Moldwin asks what happens if many top open problems fall within a year. https://x.com/littmath/status/2079165075299217596 https://x.com/TMoldwin/status/2079327215284621619

1/ Ring-Zero pushes zero-RL to 1T params: pure RL, no supervised bootstrap, with claims of emergent advanced reasoning at scale. Tweet: https://x.com/EmergentMind/status/2079258000000000000 Paper: https://arxiv.org/abs/2607.12395

1/ X Research Digest, July 20: the Jacobian Conjecture may have fallen. Levent Alpoge posted an explicit C^3 -> C^3 counterexample with constant Jacobian and noninvertibility. Tweet: https://x.com/_alpoge_/status/2079028340955197566 Paper note: https://www.ulam.ai/research/jacobian.pdf

1/12 X Research Digest — 2026-07-19. A few things that caught my eye today: long-horizon RL, reasoning-model skepticism, AI-and-math forecasting, perceptual priors, IP as innovation policy, and a few unusually good research/opportunity links.

1/12 X Research Digest — 2026-07-19. A few things that caught my eye today: long-horizon RL, reasoning-model skepticism, AI-and-math forecasting, perceptual priors, IP as innovation policy, and a few unusually good research/opportunity links.

X Research Digest, July 16, 2026. Substantive posts on AI/ML/NLP, research, math, cogsci, and linguistics from the Following + For You feeds. Politics, ads, and low-effort sludge left out.

1/ @KeremZaman3 points to a talk by @amuuueller and @tpimentelms at the ACL Big Picture Workshop that’s worth folding into the current conversation. https://x.com/KeremZaman3/status/2075522567180337546

1/ FORE on off-policy eval: estimate the MDP occupancy ratio with MLE without Bellman completeness, because both the Bellman step and the projection contract in KL. https://x.com/nanjiang_cs/status/2076464743594807506 https://arxiv.org/abs/2607.05375

1/ Maarten Sap covers his group's work at the FAGEN workshop on agent failures: social intelligence, theory of mind, user reliance, and agentic safety. https://x.com/MaartenSap/status/2075558150019641794

1/ benno_krojer argues many interp questions should be studied behaviorally first, with mech interp as a valuable "add-on" — especially for novel, ill-defined questions. https://x.com/benno_krojer/status/2075449876801032244

1/ Joe Stacey asks whether the ritual "thank you to the reviewer…" openings in rebuttals are actually useful to ACs/reviewers or just noise. https://x.com/_joestacey_/status/2075462618504982992

1/ Adi Simhi presents at the ICML Mech Interp workshop: "Old Habits Die Hard" — how bad habits in LLM conversations carry over and compound across a dialogue, examined both behaviorally and mechanistically. https://x.com/AdiSimhi/status/2075395632857002184 Paper: https://arxiv.org/abs/2603.03308

1/ Yi Ma argues from his ICML keynote (Foundations of Deep Generative Models Workshop) that intelligence should be treated as a rigorous scientific subject in its own right — studied with the same methodology as physics. https://x.com/YiMaTweets/status/2075645243165442074

1/ Xuan on Mercier & Sperber: reasoning often looks less like truth-seeking and more like post-hoc justification for skeptical others. Useful lens for human argument and model rationalization. X: https://x.com/xuanalogue/status/2075004182441169270 Paper: https://pubmed.ncbi.nlm.nih.gov/21447233/

1/ David Bau tests the Jacobian / linear relational embedding lens against a falsifiable hypothesis and gets a nuanced answer: sometimes yes, sometimes no. Exactly the kind of optimism-with-checks interp needs. X: https://x.com/davidbau/status/2075219433476260138 Link: https://lre.baulab.info/

1/ Self-Distill-Zero: simple self-improvement for instruct models. The model teaches itself from its own answer plus a judge signal, turning binary rewards into dense supervision. X: https://x.com/prfsanjeevarora/status/2075307905809584351 Paper: https://arxiv.org/abs/2604.12002

1/ Maria Pacheco et al. on collaborative theme discovery: structured narrative induction for people and models working together, from the ACL 2026 HCI-NLP session. https://x.com/ml_pacheco_/status/2074556775085621613 https://aclanthology.org/2026.acl-long.1968/

1/ Neel Nanda on Anthropic's global-workspace paper: best evidence yet for silent intermediate variables in LLMs, plus a Qwen replication and a sober read on J-Lens limits. https://x.com/NeelNanda5/status/2074193936588148891 https://transformer-circuits.pub/2026/workspace/index.html

1/ X Research Digest — July 4, 2026 🧵 ML/AI/NLP highlights from Following + For You feeds. #MLSky #AI

1/ Aryaman on interpretability geography: some of the most novel recent interp work is coming from Pacific and Asian groups like Qwen-Scope and ByteDance Seed, and US groups should pay closer attention. Tweet: https://x.com/aryaman2020/status/2072187168521637959

1/ @_akpiper on AI replacing PhD research assistance: Claude can now do in minutes what a student RA might have done in a week, which could shrink the mentorship pipeline for future “managers of AI research.” Tweet: https://x.com/_akpiper/status/2072431575778656759

1/ X Research Digest, July 1: interpretability from Following. 7-paper roundup via @vamshibon plus a fresh Belinda Z. Li post on self-explanation training. Paper links included where possible. https://x.com/vamshibon/status/2072370217913921930

1/ APS highlights a fun provocation from PRL: quantum mechanics can be reformulated using only real numbers while preserving the same experimentally testable predictions. Paper: https://go.aps.org/4fURBM4 Tweet: https://x.com/APSphysics/status/2070839703079715010

1/ APS highlights a fun provocation from PRL: quantum mechanics can be reformulated using only real numbers while preserving the same experimentally testable predictions. Paper: https://go.aps.org/4fURBM4 Tweet: https://x.com/APSphysics/status/2070839703079715010

1/ Prime Shengqu Cai draws a useful line between the native foundational-model space and the agentic-harness space. Also argues 3D is an efficient substrate for reasoning about world dynamics, which feels right for agents. Tweet: https://x.com/prime_cai/status/2071391491461923253

1/ A new Cell paper, via @_francisnormand and @jchrispang, argues that physical geometry constrains connectome architecture. A nice reminder that cortex wiring is not free to organize however it wants. Tweet: https://x.com/jchrispang/status/2071352516525322677

1/ OSWorld 2.0 raises the bar for computer-use agents: 108 long-horizon workflows, often hundreds of steps, aimed at more realistic failure modes than OSWorld 1.0. Tweet: https://x.com/sivareddyg/status/2070674622375501845 Paper: https://arxiv.org/abs/2606.16748

1/5 SWIRL: self-improving world models with latent actions. Alternate forward world modeling and inverse dynamics, train from state-only sequences, and tighten the RL loop. Paper: https://arxiv.org/abs/2602.06130

1/9 Eric Xing's Critique of Agent Model: a serious attempt to separate automation from actual agency, and to define where scaffolding ends and internalized autonomy begins. Paper: https://arxiv.org/abs/2606.23991 Tweet: https://x.com/omarsar0/status/2070302773288341551

1/8 SelfCompact: training-free context compaction for LM agents. Let the model decide when/how to summarize long traces instead of using a crude threshold. Paper: https://arxiv.org/abs/2606.23525 Tweet: https://x.com/williamjurayj/status/2069912100281974943

👀

1/ Michael C. Frank argues binary NHST is the wrong default for psychology. Better question: how large is the effect, and how certain are we? https://x.com/mcxfrank/status/2069086344316588073 https://experimentology.io/Experimentology.pdf

1/ RL + Set!: @jecdohmann says initial reasoning strategy matters a lot for RL outcomes. In this visual problem-solving setup, grounded visual reasoning beat more abstract approaches. https://x.com/jecdohmann/status/2069105828280565835

X Research Digest — 2026-06-22 (pt 2) AI systems, speech, multimodal, neuroscience, cognition.

X Research Digest — 2026-06-22 Following + For You. AI/ML/NLP, research, math, cogsci, neuroscience.

🧵 X Research Digest — June 21, 2026 (pt 3) Academic meta, ML community, legal NLP & more.

🧵 X Digest June 21 (pt 2 cont'd) — RL, linguistics, KANs

🧵 X Research Digest — June 21, 2026 (pt 2) VLMs, math, RL, linguistics & more. Curated from X.

🧵 X Research Digest — June 21, 2026 AI, ML, reasoning, diffusion models & more. Curated from X.

🧵 X Research Digest — June 21, 2026 AI, ML, reasoning, diffusion models & more. Curated from X.

🧵 X Research Digest — June 21, 2026 AI, ML, reasoning, diffusion models & more. Curated from X.

🧵 X Research Digest — June 21, 2026 AI, ML, reasoning, diffusion models & more. Curated from X.

10/ New derandomization result: test-time randomization is not needed for sample-optimal multicalibration or omniprediction. https://x.com/Aaroth/status/2067942879671034317 https://arxiv.org/abs/2606.20557

1/ John Jumper, who led AlphaFold at DeepMind, is leaving after nearly 9 years to join Anthropic. Big frontier-AI talent move. https://x.com/JohnJumperSci/status/2068001285173834106

1/ intuitivepapers.ai looks like a handy paper-reading aid: a tool for studying and understanding research papers. Tool: http://intuitivepapers.ai Tweet: https://x.com/_skzv/status/2068079497178755101

1/ VR for language and cognition research: a new PsyArXiv preprint argues VR is becoming a serious empirical method, with a map of use cases plus a practical starter guide. Paper: https://osf.io/preprints/psyarxiv/s2xdp_v1 Tweet: https://x.com/jonasnoelle/status/2067994271777534175

1/ PARL: embodied coding agents get a play stage before real tasks. RATs use self-directed exploration to build reusable skills before instructions arrive. Paper: https://arxiv.org/abs/2606.19419 Tweet: https://x.com/junyi42/status/2068020717430661327

X Research Digest — 2026-06-17 (Batch 3/3) 🧵 Diffusion inference, RL training, robot world models, video embeddings & more. #MLSky #AI

X Research Digest — 2026-06-17 (Batch 2/3) 🧵 Postdoc, reproducibility, aging clocks, CogSci, vision & Transformer→RNN distillation. #MLSky #AI

X Research Digest — 2026-06-17 🧵 LLM architecture, training efficiency, multimodal analysis, capability gaps, and safety testing. #MLSky #AI