Daily Twitter Digest

← 2026-09-242026-09-252026-09-26 →archive
Computer Science

AI Agents Co-Synthesize Code and Proofs for Verified Distributed Systems

The paper introduces Inductive Deductive Synthesis (IDS), an agentic LLM system that jointly and incrementally builds an implementation alongside its formal correctness proof, learning from failed attempts to guide future strategies. While SOTA coding agents (Codex/GPT-5.4, Claude Code/Opus 4.6) solve only 2/7 distributed key-value-store verification specs, IDS reportedly achieves 7/7 in about 6.8 hours and $106 per spec—roughly 200x faster than expert human effort and 17% cheaper than existing agents—while also optimizing implementations to run up to 3x faster than published verified systems. The authors frame this as a step toward giving AI agents formal correctness guarantees that testing alone cannot provide. On Twitter, the authors highlighted the NeurIPS oral acceptance and framed the work as a leap for combining agentic coding with formal verification, positioning it alongside related efforts on specifying human intent correctly (a follow-up challenge) and other lab outputs like SPECS and BenchEvolver. Discussion so far is mostly self-promotional from the authors, with no substantive external critique yet visible in the available tweets.

Discussion: 2 tweets from 2 authors · @shulynnliu, @mertcemri

Computer Science

LLMs Enable Scalable Deanonymization of Pseudonymous Online Users

The paper demonstrates that LLM-based agents can automatically deanonymize pseudonymous internet users, both by searching the open internet to re-identify Hacker News and Anthropic Interviewer participants at high precision, and via a closed-world pipeline that extracts identity features, embeds them for candidate matching, and uses LLM reasoning to verify matches across text corpora. Across three constructed benchmarks—Hacker News to LinkedIn, cross-platform Reddit matching, and same-user Reddit history splitting—the LLM approach achieves up to 68% recall at 90% precision, vastly outperforming classical deanonymization baselines that worked only on structured data like the Netflix prize dataset and here perform near 0%. The authors argue this undermines the 'practical obscurity' that has protected pseudonymous users and calls for revised privacy threat models. Twitter reaction was brief but pointed, with commenters framing the result starkly as automated doxxing of pseudonymous 'internet friends,' reflecting concern over the ease and scale at which LLMs can now unmask anonymous online identities.

Discussion: 2 tweets from 2 authors · @equacc, @silencexlm

Computer Science

Gated DeltaNet-2 Splits Erase and Write Gates in Linear Attention

The paper introduces Gated DeltaNet-2, a linear attention architecture that generalizes Gated DeltaNet and Kimi Delta Attention by replacing their shared scalar gate with separate channel-wise erase and write gates, allowing more precise editing of the fixed-size recurrent memory state. The authors derive a chunkwise WY algorithm and gate-aware backward pass to keep training efficient, and report that at 1.3B parameters trained on 100B FineWeb-Edu tokens, it outperforms Mamba-2, Gated DeltaNet, KDA, and Mamba-3 variants on language modeling, commonsense reasoning, and especially long-context retrieval (RULER needle-in-a-haystack). Twitter discussion, led by one of the authors, is limited to an announcement that the paper was accepted to NeurIPS 2026 and framed as new state-of-the-art for hybrid/linear attention models, with no independent critical commentary yet visible.

Discussion: 1 tweets from 1 authors · @ahatamiz1

Medicine

ERA Working Group Proposes Consensus Definition for Unexplained Chronic Kidney Disease

This consensus statement from the European Renal Association's Genes & Kidney Working Group addresses chronic kidney disease of unexplained cause (CKDx), a diagnosis of exclusion affecting an estimated 16-20% of CKD patients globally. The authors propose a formal definition of CKDx along with standardized diagnostic and reporting recommendations, emphasizing genetic testing (which yields a diagnosis in 11-30% of cases) as a key tool given the lack of prior consensus on evaluation standards. The stated goal is to reduce misdiagnosis, close diagnostic gaps, and enable a more cause-directed approach to nephrology care. Twitter discussion around this paper was limited to sharing the announcement from the ERA's official account, without substantive debate or critique visible in the available commentary.

Discussion: 1 tweets from 1 authors · @NDTsocial

Computer Science

RetNet Proposed as Transformer Alternative with O(1) Inference Cost

Microsoft researchers introduce Retentive Network (RetNet), an architecture that reformulates sequence modeling via a 'retention' mechanism supporting three equivalent computation modes: parallel (for training efficiency), recurrent (enabling constant-cost inference per token), and chunkwise recurrent (for efficient long-sequence handling with linear complexity). The authors report language modeling experiments showing RetNet matches Transformer performance while improving decoding throughput, latency, and GPU memory usage, positioning it as a potential successor architecture for LLMs. Twitter discussion centered on excitement that RetNet could end the Transformer era by combining parallel training with cheap O(1) inference, though the highlighted commentary was largely promotional rather than critical, with no substantive technical pushback surfaced in the discussion.

Discussion: 1 tweets from 1 authors · @hooshaaii

Biology

Study Links Sea Surface Temperature to Pink Abalone Recruitment Decline

No abstract is available, so this summary is based on discussion only. The paper reportedly compares stock–recruitment models for pink abalone (Haliotis corrugata) off Mexico's Pacific coast, a fishery that has been declining since 1975 amid climate-related environmental shifts. According to tweets, the analysis finds sea surface temperature useful for predicting recruitment variability, suggesting that monitoring this variable could help improve harvest management strategies. Japanese-language commentary on Twitter emphasized the strong sense of urgency running through the paper, noting that both pink abalone and a related species (H. fulgens) have suffered sustained fishery declines on the Baja California peninsula.

Discussion: 2 tweets from 1 authors · @SocStudMollDiv, @SocStudMollDiv

Medicine

Stanford Team Identifies Compound Targeting Cardiac Fibrosis via GPCR Pathway

Using an iPSC-based drug discovery platform, researchers screened for antifibrotic compounds and identified CGS15943, validating its effects in human cardiac fibroblasts, engineered heart tissues, and animal heart failure models. Mechanistically, they show that multiple adenosine receptor subtypes converge on Gβγ subunits to activate PI3K-AKT and YAP signaling, driving fibrotic gene expression; CGS15943 suppresses this atypical GPCR axis, pointing to a new therapeutic target for cardiac fibrosis, which currently lacks FDA-approved treatments. The work positions adenosine receptor-driven Gβγ signaling as a druggable pathway distinct from canonical GPCR signaling routes.

Discussion: 1 tweets from 1 authors · @Joseph_C_Wu

Medicine

Intermittent Hemodialysis Reframed as Repetitive Biophysical Stress on Organs

No abstract is available for this paper, so this summary is based on discussion only. The paper, published in Renal Failure (2026), reportedly argues that intermittent hemodialysis should not be viewed solely as a technique for clearing solutes from the blood, but also as a recurring hemodynamic and biophysical perturbation that places a cumulative mechanotransduction burden on cellular homeostasis, potentially contributing to "systemic organ stunning." Commentary on Twitter highlighted this reframing as a notable conceptual shift, emphasizing the idea that repeated dialysis sessions may impose mechanical stress on organs beyond their intended role in solute removal, though no substantive critical pushback appeared in the available discussion.

Discussion: 1 tweets from 1 authors · @JonathanNefro

Mathematics

New Infinitesimal Version of Morera's Theorem Proved

The paper strengthens the classical Morera theorem for holomorphic functions: instead of requiring circular contour integrals of a continuous function f around every point to vanish exactly, it suffices that these integrals over circles of radius r decay faster than r^2 as r→0. The authors show this weaker, purely local/asymptotic condition still forces f to be holomorphic throughout the domain, using a distributional ∂-primitive, a circular identity for weak ∂-derivatives, and an asymptotic mean-value criterion for harmonicity. This recovers the classical exact-vanishing case as a special instance.

Discussion: 1 tweets from 1 authors · @totomityann

Physics

English Translation of Yamazaki's Book on 3d-3d Correspondence Released

This is an English translation of a 2015 Japanese textbook introducing the 3d-3d correspondence, which connects 3d N=2 supersymmetric gauge theories to the geometry of 3-manifolds via compactification of the 6d N=(2,0) theory. Rather than starting from the 6d theory top-down, the book builds up from first principles—covering renormalization, gauging, Sp(2n,Z) dualities, supersymmetric localization, quantum Teichmuller theory, and complex Chern-Simons theory—to show how geometric structure emerges organically from the space of quantum field theories. It includes appendices on supersymmetry, dilogarithm functions, and cluster algebras, plus graded exercises, aimed at pedagogical use.

Discussion: 1 tweets from 1 authors · @pascalkwanten

Biology

netneurotools: An Open-Source Toolkit for Network Neuroscience Analyses

This Nature Protocols paper (no abstract available, so details are based on discussion only) introduces netneurotools, a Python toolkit developed by a network neuroscience lab to standardize and simplify common brain imaging analysis tasks. The tool is described as "trainee-oriented," suggesting an emphasis on accessibility and reproducibility for students and early-career researchers entering the field. On Twitter, the lead author framed it as their lab's internal toolkit for everyday brain imaging tasks, now shared publicly; the tweet offers a description rather than critical discussion.

Discussion: 1 tweets from 1 authors · @misicbata

Medicine

Dietary Fat Fuels Colorectal Cancer Metastasis via Ceramide-YAP Pathway

The study reports that dietary fat drives sustained ceramide biosynthesis in colorectal tumors via the enzyme Degs1, and this ceramide accumulation activates YAP (through PP2A-mediated dephosphorylation) to trigger a distinct regenerative gene program marked by Basp1. Using mouse models, the authors show that eliminating Basp1-high cancer cells or knocking out Degs1 blocks metastatic seeding without affecting primary tumor growth, while blocking ceramide degradation does the opposite—suggesting this pathway specifically governs metastatic initiation rather than tumor expansion. The findings point to a metabolic-signaling axis linking diet to cancer spread that is mechanistically separable from primary tumor biology. On Twitter, the lead author highlighted the paper as revealing how dietary fat–driven ceramide synthesis promotes metastasis independently of primary tumor growth, with congratulatory engagement from collaborators; broader independent critique was not present in the available discussion.

Discussion: 1 tweets from 1 authors · @UpasanaDA

Chemistry

New Open Benchmark Tests ML Models on Polymer Property Prediction

The paper introduces PolyBench26, an open benchmark of nearly 250,000 polymer-property datapoints spanning eight physical properties, covering not just homopolymers but also alternating, random, and block copolymers—architectures largely absent from prior benchmarks. Using this dataset, the authors compare language-model, graph-based, and descriptor-based ML approaches across four tasks (in-distribution prediction, dataset-size scaling, repeat-unit complexity, and transfer to new architectures), finding that graph-based models consistently achieve the lowest errors and remain robust as complexity increases. Twitter commentary highlighted the scale and openness of the new dataset and noted that graph-based models currently lead, while also pointing out that predicting copolymer properties remains a harder, less-solved problem.

Discussion: 1 tweets from 1 authors · @yoko_materialDX

Biology

Well-preserved Cretaceous foot fossil sheds light on early oviraptorosaur anatomy

Researchers describe a nearly complete fossilized foot (PMOL-AD00126) from China's Early Cretaceous Jehol Group, assigning it to Caudipteridae, an early-diverging oviraptorosaur clade. Using 3D reconstruction, they identify previously undocumented features—including a subarctometatarsalian metatarsus, distinct ridges on metatarsals, and elongated extensor ligament pits—while bone histology indicates the individual had not yet reached full skeletal maturity. The authors argue these findings clarify both the taxonomy of caudipterids and broader patterns in oviraptorosaur limb evolution. The discussion consisted mainly of the lead author announcing the paper's publication with enthusiasm, without substantive critique or debate visible in the available tweets.

Discussion: 1 tweets from 1 authors · @kohtapalaeo

Computer Science

RLVR Fine-Tuning Trajectories Are Nearly Rank-1, Enabling Cheap Extrapolation

The paper finds that during RLVR (reinforcement learning with verifiable rewards) fine-tuning of LLMs, the parameter changes across training steps are extremely low-rank—most of the performance gain is captured by a single rank-1 direction whose magnitude grows nearly linearly with training. Building on this, the authors propose RELEX, which observes just a short early window of training, fits a linear extrapolation along this rank-1 subspace, and predicts far-future checkpoints without further training. Tested on Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base, RELEX matches or beats full RLVR training using as little as 15% of the compute, and can extrapolate 10-20x beyond the observed steps; the authors attribute this to the rank-1 projection acting as a denoising filter that removes optimization noise. On Twitter, the main reaction was excitement from the lead author announcing NeurIPS 2026 acceptance, highlighting the striking claim of near-100% RLVR performance at roughly 20% of the training cost. Discussion so far is mostly promotional (paper/blog/code/checkpoint links) rather than critical, so independent scrutiny of the low-rank claim and its generality across models/tasks is still pending.

Discussion: 1 tweets from 1 authors · @weizhepei

Social Science

Romantic Kissing Isn't Universal—Found in Only 46% of Cultures Studied

Using cross-cultural databases (eHRAF World Cultures, SCCS) plus an ethnographer survey covering 168 cultures, researchers tested the common claim that romantic-sexual kissing (lip-to-lip contact) is a human universal. They found no support for that claim: only 46% of sampled cultures showed evidence of romantic-sexual kissing, and its presence correlated strongly with a society's social complexity—more complex societies kissed more often. Notably, no ethnographer reported the practice among Sub-Saharan African, Amazonian, or New Guinean forager societies. Twitter commentary focused on the surprising rarity of the behavior and its link to social complexity, treating the forager absence as a striking, discussion-worthy detail rather than raising direct methodological objections.

Discussion: 1 tweets from 1 authors · @jakozloski

Computer Science

LoRA+ Fixes a Learning-Rate Imbalance in Low-Rank Fine-Tuning

The paper shows that standard LoRA fine-tuning is suboptimal for large-width models because its two adapter matrices, A and B, are updated with a single shared learning rate. Using scaling arguments from large-width network theory, the authors argue this shared rate prevents efficient feature learning, and they propose LoRA+, which sets a well-chosen different learning rate ratio for A and B. Across extensive experiments, LoRA+ reportedly yields 1-2% performance gains and up to ~2x faster fine-tuning at the same computational cost as standard LoRA. Twitter discussion has focused on explaining this 'hidden mathematical imbalance' as the root cause of LoRA's limitations and framing LoRA+ as a simple, low-cost fix.

Discussion: 1 tweets from 1 authors · @hooshaaii

Medicine

Researchers Propose Reclassifying Paratonia as a Distinct Movement Disorder

This paper argues that paratonia—the involuntary, variable resistance to passive movement seen in up to 90-100% of end-stage dementia patients and common in Parkinson disease and atypical parkinsonisms—meets the criteria to be formally classified as a movement disorder rather than being dismissed as volitional refusal of care. The authors present a validated bedside instrument distinguishing paratonia from parkinsonian rigidity, spasticity, and dystonia, propose a staging schematic for when botulinum toxin A treatment is appropriate (before fixed contracture develops), and cite preliminary evidence from a randomized pilot trial supporting this intervention. They argue formal classification would enable dedicated research, education, and advocacy to improve clinical management, since paratonia is often currently mismanaged with psychotropic medications intended for behavioral refusal rather than an involuntary motor phenomenon. On Twitter, neurologists highlighted the clinical relevance of distinguishing paratonia from other rigidity-like conditions, with discussion focused on how to differentiate it from spasticity, parkinsonian rigidity, and dystonia at the bedside—reflecting interest in the diagnostic and treatment implications for dementia care rather than expressed skepticism.

Discussion: 1 tweets from 1 authors · @JackyGanguly

Medicine

New Pulsed Field Ablation Waveform Cuts Delivery Time, Preserves Safety in Pig Study

This preclinical study in 24 pigs compared two pulsed field ablation waveforms (Seq1 vs. a redesigned Seq2) across different irrigation rates and doses for pulmonary vein isolation, a key procedure for treating atrial fibrillation. All groups achieved 100% durable pulmonary vein isolation at 28 days, but the new Seq2 waveform reduced ablation delivery time from 20 to 3.8 seconds while producing zero catheter thrombus/char and few-to-no microbubbles on echocardiography, compared to more thrombus and microbubbles seen with the original waveform at lower irrigation rates. Higher irrigation flow also reduced thrombus formation with the original waveform but did not curb microbubble generation. Hemolysis and myocardial injury markers were similar across groups except for expected increases at higher (2x) doses.

Discussion: 1 tweets from 1 authors · @True_EP

Social Science

New Framework Ties SMD Denominator Choice to Statistical Estimand

This paper (no abstract available; summary based on discussion only) argues that the denominator chosen for a standardized mean difference (SMD) implicitly determines what population quantity is being estimated. The authors propose a new metric, LTG-SMD, which standardizes by the variability of true scores within an explicitly defined target population, aiming to make the estimand transparent and consistent. An accompanying R package (ltgsmd) implementing the method has been released on CRAN. The single tweet discussing this (in Japanese) simply summarizes the paper's core idea—that SMD denominators secretly encode estimand choices—and highlights the availability of the new R package, without raising notable critiques or skepticism.

Discussion: 1 tweets from 1 authors · @d_nakamuran