Daily Twitter Digest

Computer Science

Study Dissects Which Coding-Agent Harness Components Actually Help

The paper runs a controlled, component-level study of coding-agent harnesses rather than treating them as black boxes, varying planning, action space, and context management independently across 176 settings, four models, and two benchmarks (SWE-Bench Verified, Terminal-Bench 2.1). Key findings: context management mainly helps by preventing context-overflow failures and matters more as context budgets shrink; rule-based elision before LLM summarization is the most efficient context strategy; planning helps weaker models' accuracy but mainly saves cost for stronger ones; and predefined tools aid models with weak bash skills, while bash-capable models do fine (and cheaper) with just a bash interface. Trajectory analysis ties these effects to concrete behavioral mechanisms—trajectory length, stopping points, and code-edit granularity.

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

Biology

New Oviraptorosaur Species Found in Hell Creek Formation

A new paper reports Infernogallus jamesi, described as one of the most complete oviraptorosaur specimens found in North America, recovered from the Hell Creek Formation. Based on tweet discussion alone (no abstract available), the specimen is said to show dental and anatomical features suggesting fruit-eating adaptations, offering potential insight into dietary diversity among these feathered, beaked theropods near the end of the Cretaceous. No formal abstract text was accessible to verify these details. Twitter commentary was limited to a single highly-engaged post highlighting the discovery as notable, with no substantive critical discussion yet visible.

Discussion: 1 tweets from 1 authors · @Fang83768513

Computer Science

Pallas's TPU Top-k Bug Fixed Without the Performance Cost It Was Blamed On

The paper examines JAX/Pallas's top_k kernel on TPU v4, which officially returns duplicate indices for inputs containing -inf — a known limitation its source comments justify as too costly to fix. The authors show this trade-off is unnecessary: using IEEE 754 totalOrder as the correctness standard, they find official Pallas errors extend well beyond duplicates (wrong on over a third of tested special-value inputs), while a bitwise-correct version costs only 4.7% more cycles. By analyzing bottlenecks in the cross-lane unit and vector instructions, they design shape-specific algorithms that beat official Pallas on 20-24 of 25 tested shapes and always beat native XLA, in one case patching the closed-source libtpu compiler in-process to unlock needed instructions.</br>On Twitter, the author highlighted that they achieved a bitwise-correct and faster implementation despite Pallas's official stance that fixing the bug was too expensive, framing the work as a practical rebuttal to that assumption.

Discussion: 1 tweets from 1 authors · @ayaka14732

Computer Science

GitSwarm Lets AI Agents Accumulate Work via Git-Based Shared Memory

The paper introduces "compounding inference," a paradigm where inference-time computation persists across attempts rather than being discarded trajectory-by-trajectory. GitSwarm instantiates this with asynchronous, homogeneous agents that explore, verify, and refine solutions collaboratively through a shared, branchable Git repository, with atomic commits and explicit semantic dependencies tracking how work builds on prior contributions. The authors report GitSwarm solving all 30 problems on IMOProofBench-Advanced with GPT-5.5, beating baselines on ProgramBench (79.4% vs 65.1%), improving neural architectures on three research tasks, and showing that 94.7% of contributions get built upon later, evidencing genuine accumulation of computation. Twitter discussion, including a promotional post from a researcher highlighting the work, focused on the novelty of treating inference as a persistent, collaborative multi-agent process akin to version-controlled software development rather than isolated sampling, pointing to a referenced thread for deeper explanation. No substantive critical pushback appeared in the available tweets, which were largely descriptive/promotional rather than skeptical.

Discussion: 1 tweets from 1 authors · @rsalakhu

Computer Science

GATOR Turns Casual Photos into Simulation-Ready 3D Objects

GATOR is a generative and agentic framework that reconstructs complete, textured 3D object assets and their scene-relative pose from one or more casual images, even when parts of the object are occluded. It combines a modality mixer fusing RGB, mask, and pointmap features with text-guided semantic conditioning, then refines the generated asset using a multimodal agent that runs an edit-render-review loop to improve geometry and texture. The authors report strong fidelity on synthetic objects, cluttered tabletops, and indoor scenes, plus efficiency gains that make reconstructions usable in simulation. On Twitter, NVIDIA researcher Bowen Wen highlighted the system's ability to turn ordinary photos into simulation-ready assets, showcasing demos of humanoids interacting with GATOR-reconstructed furniture, though the discussion was limited to promotional framing without independent critical commentary.

Discussion: 1 tweets from 1 authors · @bowenwen_me

Computer Science

Self-Supervised Method Splits Representations into Invariant and Equivariant Parts

The paper introduces SIE, a self-supervised learning approach that splits learned representations into invariant and equivariant components, using a hypernetwork-based predictor designed to avoid collapse into pure invariance. To support evaluation, the authors release 3DIEBench, a dataset of over 2.5 million renderings across 55 object classes with full control over applied transformations, and report significant gains on equivariance-related tasks compared to existing methods. On Twitter, Yann LeCun highlighted the work approvingly while noting related prior research by himself and collaborators (Garrido and Najman) exploring similar invariant/equivariant representation learning ideas, pointing to convergent efforts in the field rather than raising substantive criticism.

Discussion: 1 tweets from 1 authors · @ylecun

Computer Science

Shared 'Rosetta' Neurons Follow a Scaling Law as Models Grow

The paper studies Rosetta Neurons—neurons whose activation patterns recur across independently trained models—in language models up to 30B and vision models up to 5B parameters. The authors find these neurons grow in absolute number but shrink as a fraction of total neurons, following a sublinear power law, while becoming increasingly selective and monosemantic with scale (a 'Neuron Polarization Effect'). They propose an analytical model balancing feature utility against limited neuron capacity to explain both trends, and show Rosetta Neurons become more domain-specialized, demonstrating this via a data-filtering case study for continued pretraining. On Twitter, the main reaction was celebratory: the authors announced the paper won the Best Paper award at the COLM workshop on Scientific Understanding of Foundation Models, with limited independent discussion or critique visible beyond that announcement.

Discussion: 1 tweets from 1 authors · @yasamanbb

Computer Science

New Book Covers Foundational Concepts of Large Language Models

This book-length arxiv submission presents a structured introduction to large language models, organized into six chapters covering pre-training, generative models, prompting, alignment, inference, and reasoning. The authors explicitly aim to cover foundational concepts rather than the latest cutting-edge techniques, positioning it as a reference for students, researchers, and practitioners in NLP. The tweet discussing it simply lists this alongside two other machine learning systems textbooks without substantive commentary or critique, so there's no notable debate to report—just a pointer to a resource bundle for people studying the field.

Discussion: 1 tweets from 1 authors · @caglar_ee

Physics

Lecture Notes on Renormalization Group and SUSY QFT Resurface Online

This is a pedagogical lecture series by Timothy J. Hollowood, originally delivered at BUSSTEPP, introducing the renormalization group (RG) in quantum field theories relevant to particle physics. The notes aim to build physical intuition for the running of couplings, connecting the Wilsonian RG picture to standard perturbative calculations done via dimensional regularization and minimal subtraction, before introducing RG properties specific to supersymmetric theories. The tweet sharing it offers no additional commentary beyond the title, so the summary here is based on the abstract alone.

Discussion: 1 tweets from 1 authors · @Aobeka11

Other

FactorBench Benchmarks 5,000 AI-Mined Trading Factors, Finds No Clear Winner

FactorBench evaluates roughly 5,000 automatically mined financial factors from nine methods—spanning genetic programming, reinforcement learning, generative models, and LLM agents—across five equity markets. It uses a shared evaluation framework to assess factor validity, temporal generalization, distinctness from each other and from the classic Alpha101 benchmark, and ultimately after-cost portfolio performance in long-only and long-short strategies. The paper's headline finding is that no single paradigm consistently dominates, raising questions about whether more sophisticated discovery methods actually translate into better real-world trading signals. Twitter discussion was brief, mainly sharing the paper title and link without extended critique, though the framing implicitly invites skepticism about hype around LLM- and RL-based factor discovery outperforming simpler methods.

Discussion: 1 tweets from 1 authors · @PtrPomorski

Other

Bronze Age Copper Cymbals from Oman Point to Indus Musical Traditions

Archaeologists report a pair of well-contextualized third-millennium BC copper-alloy cymbals excavated at Dahwa, Oman, the only such securely dated find from Arabia. The authors argue these cymbals closely parallel contemporaneous examples from the Indus Valley and depictions in Mesopotamian iconography, supporting evidence for Indus migrant communities in Early Bronze Age Oman and suggesting shared musical, and possibly ritual, practices across the Arabian Gulf region. Twitter commentary from the publishing journal highlighted the rarity of the find given how poorly instruments typically survive archaeologically, framing it as a unique window into prehistoric musical traditions; no substantive criticism appeared in the discussion.

Discussion: 1 tweets from 1 authors · @AntiquityJ

Medicine

Review Stresses Continued Need for Histological Typing in Cardiac Amyloidosis

This JACC review, led by Esther González, addresses the role of histological amyloid typing in diagnosing cardiac amyloidosis (ATTR-CM and related forms), covering indications, biopsy techniques, and interpretive pitfalls. No abstract was available, so this summary is based on discussion only. On Twitter, a cardiologist highlighted the piece as a timely reminder that, despite growing enthusiasm for non-invasive diagnostic pathways (e.g., scintigraphy-based approaches), tissue-based typing remains an essential and still unmet diagnostic need at many centers.

Discussion: 1 tweets from 1 authors · @MSBBrandao

Computer Science

New Benchmark Tests Whether AI Judges Can Verify Long Computer-Use Tasks

AgentHorizon introduces a benchmark of 1,373 computer-use tasks built from 166 hours of human-recorded trajectories across three operating systems, designed to test whether LLM-based judges can correctly determine if a long agentic task (100-300 steps) was actually completed as instructed. By pairing similar instructions and swapping them to create 'negative' trajectories, the benchmark checks whether judges can distinguish genuine success from a plausible-looking but incorrect completion; the best model tested, GPT-5.5, reached only 80.9% balanced accuracy, with notable differences in how models handle acceptance versus rejection of valid/failed runs. Twitter commentary largely just relayed the paper's core framing and results, highlighting the benchmark as a useful tool for probing the reliability of automatic judges used in agent training and evaluation, with no substantive critical pushback evident in the discussion.

Discussion: 1 tweets from 1 authors · @xhluca

Biology

New 3D Super-Resolution Method Speeds Up Whole-Cell Protein Mapping

The paper introduces Intrinsic3D, a method combining oblique plane microscopy with single molecule localization microscopy (SMLM) to image entire cell volumes in 3D at nanometre resolution more than ten times faster than standard DNA-PAINT, by scanning a light sheet and merging localization data from adjacent planes without needing to stitch separate image stacks. The authors demonstrate the technique on nuclear pore complexes in U2OS cells and on Drosophila egg chambers, mapping E-cadherin and Gliotactin distributions to reveal that 'spot' adherens junctions are enriched near tricellular junctions and may play a broader adhesion role than previously thought. The authors suggest the method could be combined with multiplexing approaches like exchange-PAINT to enable 3D spatial proteomics in tissue. Twitter discussion was limited to a single share noting the paper's title and highlighting the fast 3D SMLM imaging approach, without substantive critical commentary.

Discussion: 1 tweets from 1 authors · @vnzloy

Mathematics

New 'Duo' Divergence Yields Closed-Form KL Divergence for Truncated Distributions

The paper generalizes the Fenchel-Young and Bregman divergence frameworks to handle comparisons between probability measures from different (nested) exponential families, introducing a 'duo' Fenchel-Young divergence and proving a majorization condition that guarantees its non-negativity. Using this machinery, the authors derive a closed-form expression for the Kullback-Leibler divergence between truncated normal distributions and show that a skewed Bhattacharyya distance between nested exponential families is equivalent to a skewed duo Jensen divergence. Twitter discussion, from the author himself, highlighted the truncated Gaussian KL divergence result as a notable concrete application of the new theoretical construct.

Discussion: 1 tweets from 1 authors · @FrnkNlsn

Social Science

Brazilian Soap Operas Linked to Lower Fertility Rates

This paper estimates television's effect on fertility in Brazil by exploiting the staggered rollout of Globo, the dominant producer of soap operas (novelas) that typically depict small, affluent families. The authors find that women in areas newly covered by Globo signal had significantly lower fertility, with effects concentrated among lower-socioeconomic-status women and those in mid-to-late childbearing years, consistent with couples stopping after fewer children. They argue the effect isn't just TV ownership but specifically novela content, citing evidence that parents began naming children after soap opera characters. Commentary on Twitter highlighted these striking findings—72% of lead actresses under 50 had no children on-screen, and births fell about 5% as coverage expanded—framing it as a vivid example of media shaping real-world demographic behavior, though the tweet itself offered no critical pushback.

Discussion: 1 tweets from 1 authors · @jakozloski

Biology

Ancient DNA Reveals Genetic Continuity in Yellow River Basin Over 4,500 Years

This study analyzes genomes from 61 ancient individuals (roughly 4,500–200 years before present) excavated in Henan and Shandong provinces, tracing population history in the middle-lower Yellow River basin. Based on the paper's title, the researchers report evidence of long-term genetic continuity among local populations alongside episodes of admixture with groups from other regions, though the full abstract was not available to verify specific methods or findings. The work comes from a team at Xiamen University. Twitter discussion was limited to a brief announcement noting the paper's publication and the size of the ancient DNA dataset, with no substantive technical critique offered in the available commentary.

Discussion: 1 tweets from 1 authors · @liluoiiluoli

Chemistry

AstraZeneca Details Improved Scale-Up Route to ATR Inhibitor Ceralasertib

This Organic Process Research & Development paper (no abstract available, so details are based on discussion) describes a process chemistry route developed by AstraZeneca researchers for manufacturing ceralasertib, an orally bioavailable ATR kinase inhibitor being investigated as an anticancer agent, with improvements aimed at scale-up suitable for larger production. The Twitter mention simply highlights the paper as the basis for a ChemDev poster presentation by Violet YJ Chen, crediting Mark Graham and colleagues at AstraZeneca, without additional technical discussion or critique.

Discussion: 1 tweets from 1 authors · @MatthewHorwitz1

Medicine

Digital training tied to fewer anastomotic leaks after right colectomy

EAGLE-2 is an international prospective cohort study of 2,875 right colectomy patients across 332 hospitals in 60 countries, comparing outcomes between surgeons who had completed an open-access digital safe-anastomosis training module and those who had not. Anastomotic leak or collection occurred in 16.4% of patients operated on by untrained surgeons versus 9.1% for trained surgeons, and after adjustment for case-mix (ASA grade, urgency, contamination) training was associated with roughly half the odds of leak, plus lower reoperation and readmission rates. The authors argue this real-world data supports embedding the brief digital module into colorectal quality-improvement programmes globally, building on an earlier randomized trial (EAGLE) that found similar benefits.

Discussion: 1 tweets from 1 authors · @BJSurgery

Computer Science

A Single Orthogonal Map Aligns Independently Trained Multimodal Models

The paper shows that independently trained multimodal contrastive models (CLIP, SigLIP, FLAVA variants), despite differing architectures and training data, induce embedding spaces related by a single orthogonal transformation Q (plus a mean shift) that simultaneously aligns both the image encoder and the text encoder across models. The authors back this empirical finding with a theoretical argument: if two models' image-text similarity kernels agree on a small anchor set, they must be related by exactly this kind of unified orthogonal map, with practical implications for backward-compatible model upgrades without re-embedding and for representation privacy. On Twitter, a researcher who previously worked on a related problem for a narrower class of contrastive models called the result 'incredibly cool,' highlighting that it extends unsupervised representation alignment to a broader, more general setting than their own prior work could establish.

Discussion: 1 tweets from 1 authors · @sharut_gupta