Daily Twitter Digest

Physics

AI Agent Solves 49/50 Unpublished Research-Level Theoretical Physics Problems

The paper introduces TheoPhysEval, a benchmark of 50 unpublished, research-level theoretical physics problems with verifiable answers, each solvable by a world expert in about a day or less. Testing off-the-shelf AI agents without human hints, the best performer—GPT-6-Astra run in Codex with tool access (literature search, scripts for symbolic/numerical work)—achieved 98% accuracy at under $3 per problem, and accuracy barely dropped when web search was disabled, suggesting heavy reliance on memorized literature. The authors note this raises the question of why such strong problem-solving hasn't yet translated into AI-led research breakthroughs in theoretical physics. Twitter discussion centered on astonishment at the near-perfect score and reports that some answers initially marked wrong were actually correct—meaning the human-generated answer key itself contained errors. Commenters (including a theoretical physicist reportedly losing sleep over the implications) framed this as an early warning sign that humans may soon struggle to independently verify AI-generated research output, while skeptics pushed back against hype framing like "physics is solved."

Discussion: 10 tweets from 10 authors · @curiouswavefn, @TricksterKallis, @rad1_study, @WKCosmo, @tmcosmos_org +5

Biology

New Jurassic Fossil Suggests Early Birds Lost Flight Sooner Than Thought

A newly described Jurassic avialan, Yuezhuornis yangyuanensis, is reported by Xu, Wang, Lin and colleagues in Nature as a flightless bird-line dinosaur, suggesting that early avialans had already diversified ecologically—including losing flight—much earlier than previously appreciated. No abstract was available, so this summary is based on the paper's title and discussion rather than detailed methods or findings. Twitter commentary from paleontologists simply flagged the new species name and citation, noting its significance for understanding early bird evolution without raising substantive criticism.

Discussion: 2 tweets from 2 authors · @Brontops_, @TomHoltzPaleo

Mathematics

Kane and Tao Pin Down Sharp Exponent for Counting Cluster Partitions

The paper bounds, for a finite collection X of sets, the number of triples (A1, A2, A) in X×X×X such that A splits as the disjoint union A1⊎A2. The authors show this count is at most |X|^(3/p), where p = log_3(27/4) ≈ 1.73814, giving an exponent of about 1.726 — a substantial improvement over the trivial |X|^2 bound — and prove the exponent is optimal. The result, motivated by run-time analysis of the ASTRAL phylogenetics algorithm, is proved via a one-dimensional convolution inequality verified through calculus and numerical checks, and the authors derive an analogous sharp bound (exponent log_2 6) for additive energy of subsets of the discrete cube {0,1}^n.

Discussion: 1 tweets from 1 authors · @PI010101

Computer Science

Agaram 2.5B: Tamil-First LLM Trained from Scratch Beats Sarvam-1 on Translation

The paper describes Agaram 2.5B, a 2.5B-parameter Tamil-English model trained from scratch on ~50B tokens (58% Tamil) using eight H100 GPUs over six days for about $13,300 total. After supervised fine-tuning on ~134k conversations, it outperforms Sarvam-1 on translation benchmarks (FLORES, IN22) and nearly doubles it on native-Tamil instruction-following (IndicIFEval), can perform basic tool-calling, but scores near chance on knowledge and reading-comprehension tasks, with only ~46% of everyday-question replies judged usable due to frequent factual errors. The authors frame this as a low-cost playbook for building language models for under-represented languages. Twitter discussion centered on the model's public preview launch, with the lab emphasizing Tamil/Tanglish/English chat capability and tool use, while noting that model weights await licensing clearance before release.

Discussion: 1 tweets from 1 authors · @Mechramc

Computer Science

LLM-jp-4: Japan's Open Sovereign LLM Family Detailed in New Report

The technical report introduces LLM-jp-4, a family of Japanese sovereign large language models built with emphasis on transparency and reproducibility, including an 8B dense model and a 32B-A3B mixture-of-experts model, both trained from scratch on roughly 12 trillion tokens. The authors document explicit criteria for selecting training data, external models, and software, and report evaluation results showing substantial gains over prior LLM-jp models and competitive performance against similarly-sized open-weight models. Twitter discussion simply flagged the release with minimal added commentary, mostly serving to spread awareness of the report within the open-source LLM community.

Discussion: 1 tweets from 1 authors · @odashi_t

Medicine

Review Charts 16 Years of Progress in HIV Treatment and Prevention

This Clinical Microbiology Reviews article surveys how HIV therapeutics have evolved since 2010, highlighting the shift from first-generation antiretrovirals to integrase inhibitor-based regimens anchored by dolutegravir and bictegravir, along with two-drug and NRTI-sparing strategies. It emphasizes long-acting injectables—cabotegravir-rilpivirine, standalone cabotegravir PrEP, and the capsid inhibitor lenacapavir—as the most consequential recent advances, alongside new options like fostemsavir and ibalizumab for multidrug-resistant HIV. The review also previews an emerging pipeline (third-generation integrase inhibitors, maturation inhibitors, broadly neutralizing antibodies) while flagging challenges including resistance to long-acting agents, metabolic effects of lifelong therapy, global access inequities, and gaps in pregnancy/breastfeeding safety data. Twitter discussion consisted of a single share framing the paper as a forward-looking snapshot of "HIV therapeutics in 2026," with no substantive critique or debate offered in the available commentary.

Discussion: 1 tweets from 1 authors · @AliSMV7

Computer Science

Study Traces Diffusion Policies' Edge Over Regression to Gradient Imbalance, Not Multimodality

The paper investigates why diffusion/flow-matching policies typically outperform simple MSE-trained regression policies in robot learning from demonstrations, a gap usually attributed to multimodal action distributions. The authors analyze real robot demonstration data and find heavy-tailed, state-dependent residual scales, and show that MSE training disproportionately weights gradients toward observations with large residuals, degrading optimization. They propose heteroscedastic Student-t regression policies (HT-Policies) that model input-dependent residual scales, achieving success rates competitive with generative baselines across four simulation benchmarks and real-robot tests while being faster to train and run, since they require only a single feed-forward pass without denoising. On Twitter, the author's thread framed the core contribution as showing the regression-vs-generative performance gap stems from imbalanced gradients under MSE rather than multimodality or manifold adherence, highlighting that a simple loss reformulation closes the gap without sacrificing inference speed.

Discussion: 1 tweets from 1 authors · @yuchen010807

Computer Science

Training-Free Attention Lets Diffusion Transformers Scale to 6x Native Resolution

The paper proposes AReA, a training-free attention mechanism for Diffusion Transformers that addresses the quadratic cost and out-of-distribution distortions that occur when generating images beyond a model's training resolution. The authors observe that attention's effective receptive field naturally narrows during denoising—from global early on (forming structure) to local later (refining detail)—and design AReA to mirror this: coarse cross-scale attention over downsampled tokens in early steps, shrinking to local windowed attention later. Without extra parameters or retraining, they report it outperforms existing training-free high-resolution baselines on quality while running up to 3.2x faster and reaching up to 6x native resolution.

Discussion: 1 tweets from 1 authors · @rajabi2001

Computer Science

Nvidia's TacitVLA Decouples Robot Reasoning from Action Latency

TacitVLA is a vision-language-action model that generates textual reasoning to improve its internal representations, but masks that reasoning from the action-generation module so actions don't have to wait for autoregressive text decoding to finish. The authors argue the benefit of reasoning comes mainly from training supervision shaping the backbone, not from actions conditioning on the generated text at inference time. Across four benchmarks (RoboMME, RMBench, LIBERO, PhysicalAI-AV) spanning manipulation and driving, it reportedly matches explicit chain-of-thought performance at up to 3.4x lower latency, and beats no-reasoning baselines by up to +17.1% success while surpassing prior published results on RoboMME Counting and RMBench. The tweet (from Aran Komatsuzaki) simply highlights the core finding—matching CoT performance at much lower latency—with no substantive critical discussion yet visible in the thread.

Discussion: 1 tweets from 1 authors · @arankomatsuzaki

Engineering

RootSLAM Speeds Up Visual-Inertial SLAM Up to 20x via Square-Root Bundle Adjustment

RootSLAM is a visual-inertial SLAM system for multi-camera setups, built on Basalt, that adapts square-root bundle adjustment from structure-from-motion to the SLAM pose-graph setting. By reformulating the global optimizer to run in single-precision floating point, the authors claim accuracy competitive with state-of-the-art learning-based methods while running up to 20x faster, with benchmarks and an ablation study provided; the implementation is to be open-sourced. Twitter discussion, led by one of the authors, highlighted the speed gains (3–14x faster in the tweet, up to 20x per the paper) and framed the work as an international academic collaboration, with no substantive critical pushback visible in the discussion shown.

Discussion: 1 tweets from 1 authors · @mateodemayo

Engineering

LLM Agents Autonomously Design, Verify, and Tape Out an Analog SAR ADC

The paper tests how far 'vibe tapeout'—LLM agents writing their own schematic/layout generators, running EDA tools, and iterating on results—has progressed for analog circuits. On an open SKY130 PDK, a cloud LLM agent autonomously designed an 11-bit, 5.12-MS/s SAR ADC, producing DRC/LVS-clean layouts that met spec after parasitic extraction, with a reference-design ('vibe porting') condition reaching full signoff without human intervention and yielding 1.6–2.8x smaller area; the agents also diagnosed post-extraction issues like CDAC coupling and comparator offset on their own. On a commercial 65nm PDK (under NDA, so tested with local LLMs), Kimi K3 passed 5 of 6 benchmark tasks while MiniMax M3 passed 2, and neither could solve a debugging task requiring inference from symptoms to cause; the authors note agents sometimes reported false passes, making independent verification essential, and conclude this is a 'golden age' specifically for designers who know what to verify.

Discussion: 1 tweets from 1 authors · @arutema47

Biology

New Nursery Pollination Mutualism Found Between Japanese Plants and Sawflies

Researchers describe a previously unknown nursery pollination mutualism between Japanese endemic Chrysosplenium ser. Macrostemon plants and Nipponorhynchus sawflies—the first such case documented involving sawflies rather than moths or other insects. Based on surveys of 29 populations, larval parasitism checks, pollinator-exclusion experiments, and proboscis measurements across 154 specimens, the authors show that these sawflies exclusively pollinate Macrostemon flowers while their larvae develop inside the resulting fruits feeding on seeds, with exclusion experiments confirming the plants benefit from visitation despite having backup self-pollination. The study also finds geographic variation in sawfly proboscis length that tracks local floral morphology, suggesting fine-scale coevolution between the partners.

Discussion: 1 tweets from 1 authors · @yokuyama

Biology

Review Charts Re-emergence of Brucella Species Across Animals and Humans

This review synthesizes current knowledge on Brucella species, the intracellular bacteria causing brucellosis, a major zoonotic disease transmitted from livestock, dogs, wildlife, and marine mammals to humans. Drawing on genomic surveillance and One Health epidemiology, the authors describe expanding species diversity (including emerging canine and marine lineages), mechanisms of intracellular pathogenesis such as vacuole maturation and type IV secretion system-mediated immune evasion, and the state of next-generation vaccine development given the lack of effective human or canine vaccines. The review frames itself as distinct from prior livestock-focused reviews by integrating these newer lineages and mechanistic advances into a unified One Health perspective.

Discussion: 1 tweets from 1 authors · @AliSMV7

Mathematics

Central Limit Theorem Proved for High-Temperature SYK Model, With AI-Assisted Proof

The paper proves a central limit theorem for the Gaussian q-body SYK model: at sufficiently small inverse temperatures β, the fluctuations of log Z_N(β) around its mean, rescaled by N^(q/2-1), converge to a Gaussian distribution with variance ρ(β)². Beyond the math result, the authors document how the proof was developed through iterative collaboration between human researchers and AI tools, and describe formalizing the main arguments in Lean 4. On Twitter, the discussion centered less on the SYK result itself and more on the paper's transparency about its AI-assisted research workflow — the authors detail upfront how AI was used throughout proof development, which commenters noted as reflecting a notable shift in how such collaboration is now openly disclosed in papers.

Discussion: 1 tweets from 1 authors · @ryoichiro_noda

Computer Science

Amazon Team Open-Sources Full Post-Training Recipe for GLM-4.5-Air

Rufus-Air documents an eight-stage post-training pipeline (SFT, Reasoning RL, Coding RL, Instruction-Following RL, General/Coding/Search Agent stages, and RLHF) applied to GLM-4.5-Air-Base, using only public data and open-source components without new human annotation or a distillation teacher. The authors report that strong SFT builds a capability floor, difficulty filtering keeps RL prompts productive, reward reliability guides stage ordering, and infrastructure choices matter as much as algorithmic ones—yielding a model that beats the official GLM-4.5-Air release and rivals similarly sized open models. Twitter discussion highlighted the project's scrappy, reproducible nature: a small Amazon Search team built this over roughly eight months using modest compute (rarely more than 400 H200s), including a two-month production pause.

Discussion: 1 tweets from 1 authors · @tourzhao

Social Science

Top Job Promotions Double Divorce Odds for Women, Not Men

The study compares Swedish political candidates who narrowly won versus lost races for mayor or parliament, as well as newly promoted CEOs, to isolate the causal effect of career promotion on marriage stability. It finds that a promotion to a top job roughly doubles women's baseline probability of divorce, while men's divorce rates are unaffected; the effect is concentrated among more gender-traditional couples, with gender-equal couples showing no such increase. The authors interpret this as evidence that status reversals strain marriages operating under traditional gender norms more than egalitarian ones. Twitter discussion focused mainly on relaying the striking headline finding—that winning (not just running for) a top office doubles women's divorce odds—without much pushback, treating it as a notable illustration of how career success interacts with gender dynamics in marriage.

Discussion: 1 tweets from 1 authors · @jakozloski

Biology

Zoo-Field Pipeline Rescues Abandoned Brown Shrike Eggs for Hatching

Researchers developed a two-phase salvage incubation protocol for the endangered Japanese Brown Shrike: abandoned eggs are first given 'bridging' incubation in portable field incubators, then transferred to Toyohashi Zoo for controlled second-phase incubation. Across 20 nests over 2023–2025, 83 eggs were salvaged, 73 showed embryonic development, and 41 hatched (56.2% hatching success among developing eggs); GLM analysis identified longer prior parental incubation, warmer field temperatures during abandonment, and a specific 37.8°C regime with one daily 5-minute cooling event as key predictors of success. The authors present this as a practical, replicable field-to-facility pipeline linking in situ monitoring with ex situ conservation for passerines.

Discussion: 1 tweets from 1 authors · @Nnipponia

Computer Science

Category Theory Unifies 382 Parameter-Efficient Fine-Tuning Methods

The paper proposes a categorical framework for comparing parameter-efficient fine-tuning (PEFT) methods, treating each as a smooth map from a small parameter space to weight space, with the frozen model as initial object and full fine-tuning as terminal. Using invariants like image geometry, fibre structure under the optimizer's metric, and base-change operations, the authors derive concrete theorems — e.g., bounds on LoRA's achievable rank, equivalence of LoRA+ trajectories to scaled plain LoRA under certain Adam settings, and limits on how orthogonal methods affect singular values — then classify 382 PEFT methods from 138 papers along seven formal coordinates. Every result is stated as a proved theorem, distinguishing original contributions from re-derivations of prior work. The lone tweet highlighted praises the paper's unusual rigor and polish, while noting (with apparent approval rather than complaint) that AI assistance — specifically Claude — appears to have contributed to the writing.

Discussion: 1 tweets from 1 authors · @aryaman2020

Biology

Genetic basis of leopard gecko and corn snake albino morphs mapped

Researchers assembled a chromosome-level leopard gecko genome and combined genomic, transcriptomic, and electron-microscopy data to identify candidate mutations behind hypomelanistic color morphs in corn snakes and leopard geckos. They propose that the three classic leopard gecko 'hypo' morphs trace to mutations in TYR (Tremper), SLC24A5/NCKX5 (Rainwater), and OCA2 (Bell), while additional pigmentation genes are dysregulated in mutant embryonic skin, suggesting a multigenic basis with morph-specific alterations to melanophore substructure. The authors argue this establishes reptiles as useful comparative models for pigment cell biology, showing conserved melanogenesis genes can produce species-specific effects. The tweet thread (in Japanese, from a snake/gene genetics account) simply relayed the paper's key finding, mapping each named leopard gecko morph—Tremper, Rainwater, and Bell albino—to its respective candidate gene (TYR, SLC24A5, OCA2) without added critique.

Discussion: 1 tweets from 1 authors · @gradation_Delta

Physics

Molecular Dynamics Confirms 'Effective Gravity' Can Flip Liquid-Gas Layering

Using 2D Lennard-Jones molecular dynamics simulations, the authors test a prediction of 'global thermodynamics': that under heat conduction (hotter bottom, cooler top), the usual liquid-below/gas-above arrangement can invert at some finite downward gravity rather than at zero gravity. They show the transition point matches calculations based on independently measured equilibrium saturation pressures, and that hysteresis behavior under heat flow collapses onto the equilibrium response when plotted against this 'effective gravity,' supporting the theory's predictive power for nonequilibrium structure formation.

Discussion: 1 tweets from 1 authors · @NAKAGAWANaoko1