1/ Akari Asai received the 2025 AAAI/ACM SIGAI Doctoral Dissertation Award for "Beyond Scaling: Frontiers of Retrieval-Augmented Language Models." X: https://x.com/AkariAsai/status/2084278883554033741

1/ Daniel Yamins: v2 of "Theory of Contravariance" adds transformer contravariance plus RSA/CKA treatment. X: https://x.com/dyamins/status/2084608953183302082 Paper: https://arxiv.org/pdf/2607.08561

23/ Chanwoo Park on NeurIPS review norms: don't expect authors to answer reviewer follow-ups on their own rebuttal if you haven't engaged the rebuttals on papers you reviewed. https://x.com/chanwoopark20/status/2084063990745809047

15/ Prompt language and cultural framing matter: this COLI paper looks at how response patterns and value alignment shift across countries and languages. https://x.com/CompLingJournal/status/2084279443753976108 https://doi.org/10.1162/COLI.a.583

7/ Jesse Hoogland reading group on Singular Learning Theory: tying loss-landscape geometry to internal model structure as a principled story for why neural nets generalize. https://x.com/industriaalist/status/2084365807547781584 https://arxiv.org/abs/2402.02364

1/ ICER: a black-box attacker can recover "erased" concepts from safety-tuned text-to-image models via in-context experience replay, with no gradients, agents, or aux VLM. https://x.com/zhiyichin/status/2083675263695556958 https://arxiv.org/abs/2411.16769

1/ ICER: a black-box attacker can recover "erased" concepts from safety-tuned text-to-image models via in-context experience replay, with no gradients, agents, or aux VLM. https://x.com/zhiyichin/status/2083675263695556958 https://arxiv.org/abs/2411.16769

1/ Dimakis on ICLR authorship quotas: like the stacked-collar polo era, a vivid signal of what a field decides is “cool” versus sustainable. X: https://x.com/AlexGDimakis/status/2083858827196199354

1/ TOPL reframes post-training as token-level correctness classification rather than next-token prediction, with better faithful generation under shift. X: https://x.com/ZitongHuang12/status/2083237028892655781 Paper: https://arxiv.org/abs/2607.17524

1/ Agent-assisted debugging datapoint: one month, bugs found across PyTorch, torch/ao quantization, slime, Tinker, vLLM, and FlashQLA GDN kernels. X: https://x.com/jxmnop/status/2083582177082794174

1/ Lovely math-memory post: at Olympiad camp, Oveis Gharan glanced at a combinatorics problem and outlined the full proof in about a minute. X: https://x.com/abeirami/status/2083270288632881276

1/ Tunguz with a spicy historical frame: neural nets may be vestiges of the era when humans still hand-designed ML algorithms, when composability mattered more because human insight was the bottleneck. https://x.com/tunguz/status/2082268293147365592

1/ Proof of Gaffke's conjecture, built by combining recent results and fully formalized in Lean 4. Rare treat: real math plus end-to-end formalization. https://x.com/GuanyangW/status/2081939934529917089 https://arxiv.org/abs/2607.23980

1/ Horace He with a joking theory of ML impact: if your work causes infra pain, it probably works, and if it really works, infra will feel it. X: https://x.com/cHHillee/status/2082601096250589620

1/ Chenhao Tan: mathematicians shouldn't assume AI disruption will just go away; adapt hard and fast instead of waiting it out. X: https://x.com/ChenhaoTan/status/2082469833577570457

1/ Decoder-only LMs may be injective: hidden states can in principle preserve enough info to recover the input exactly. X: https://x.com/s_scardapane/status/2082436329741160825 Paper: https://arxiv.org/abs/2510.15511

1/ Michaelzsguo revisits Lilian Weng's 2019 OpenAI talk on dexterous in-hand robotic manipulation after news that she's stepping back from Thinking Machines Lab for her health. X: https://x.com/Michaelzsguo/status/2081848350501167159

12/ @StatMLPapers on simulation-based inference with ML: a guide to Bayesian and frequentist SBI for inverse problems. X: https://x.com/StatMLPapers/status/2081591298063003896 Paper: https://arxiv.org/abs/2607.21702

1/ @AndrewLampinen on recent LLM-math progress through the old symbols-vs-neural-nets debate. X: https://x.com/AndrewLampinen/status/2081876412869615803 Paper: https://arxiv.org/abs/2102.03406

1/ Hebbian learning as a unifying lens: recent ICML work argues many learning algorithms look Hebbian under regularization and anti-Hebbian under noise, hinting at shared learning dynamics. https://x.com/LiuZiyin10/status/2080348442564472916

1/ Vitushkin's construction and rational maps: a note that the recent construction likely builds on Vitushkin's work on how a hypothetical 2D counterexample behaves at infinity, plus the striking detail that Vitushkin was blind. https://x.com/_onionesque/status/2079719860016427288

1/ DeepSearch-World + DeepSearch-Evolve: a 9B self-evolving web agent for long-horizon search, with verified self-improvement plus released code/data. https://x.com/LiangZhenwen/status/2079710767885279513 https://arxiv.org/abs/2607.07820

1/ Neel Nanda reports a wild interp finding: when a model is confused about a sentence’s meaning, Chinese “what does this mean?” characters can surface in latent J-Space. https://x.com/NeelNanda5/status/2079676406053961754

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