Daily Twitter Digest

Physics

Sub-Earth Planet Candidate Found at Nearby M-Dwarf, But Paper Includes Unusual Disclaimer

The paper reports a sub-Earth-sized transiting planet candidate (R≈0.70 Earth radii) around GJ 774 A, an M-dwarf just 12.86 pc away, based on repeated TESS transit signals across eight sectors spanning 2019-2026. The authors run extensive checks (odd/even transit tests, secondary-eclipse searches, PRF-deblended photometry, Gaia RV data) to rule out blended eclipsing binaries, concluding the signal most likely originates from the primary star A rather than its wide binary companion B, though the planet is not formally validated and some 2-3 sigma tension in the photometry remains unexplained. Twitter users highlighted and reacted with surprise to an unusual disclaimer in the text stating the author had not independently verified all analyses and lacked expertise to check every technical detail, with some questioning whether this violates arXiv's submission norms; this is being discussed as commentary rather than an established fact about the paper's validity.

Discussion: 3 tweets from 3 authors · @Dr_Dark_Matter_, @david_kipping, @AstroMikeMerri

Medicine

Small 4B-Parameter Model Rivals Larger LLMs on Clinical Record Tasks

Researchers introduce Clinical System One (CS1), a 4-billion-parameter 'decision model' trained on 12.4 million question-answer pairs from 3.6 million patients to make fast, typed judgments from clinical records without task-specific retraining. On nearly a million held-out patients, CS1-4B outperformed open decision models up to 27B parameters on 13 of 14 tasks while running 6.4x faster, and when paired with a generalist LLM agent, it screened patients for clinical trial eligibility far more effectively than keyword search while using 690-fold fewer LLM tokens. The authors frame this as evidence for a division of labor in clinical AI: small models for repeated record-reading, large generative models for synthesis and reasoning. Twitter discussion, led by the authors themselves announcing the preprint, emphasized the scale problem in healthcare data and framed CS1 as a solution for population-level questions (e.g., trial matching, follow-up tracking) that are currently too expensive to ask of large patient populations; the thread had not yet accumulated substantive outside critique at time of posting.

Discussion: 1 tweets from 1 authors · @AI4Pathology

Computer Science

QD-Rec Uses Quality-Diversity Search to Generate Diverse, Reasoned Recommendations

The paper proposes QD-Rec, a method for recommendation in high-stakes decisions (e.g., M&A targets, customer targeting, paper/book suggestions) where not just the choice but the reasoning behind it matters. It combines a contextual bandit for item selection with a foundation model that generates natural-language 'hypotheses' justifying each recommendation, using an archive of past hypotheses as stepping stones to encourage novelty. The authors report QD-Rec achieves the highest mean recommendation quality and builds the largest, most diverse archive of solutions across several tested domains compared to baselines. Twitter discussion, led by Sakana AI Labs (the paper's authors), framed this as extending open-ended evolutionary search—previously used for algorithm discovery and self-improving agents—into decision-support tools, emphasizing diverse well-reasoned options over single ranked lists; the engagement was primarily promotional with no independent critical commentary visible in the discussion.

Discussion: 1 tweets from 1 authors · @SakanaAILabs

Social Science

New Framework Proposes a 'Class Geography of Crisis' for Labour and Ecology

This paper argues that labour geography—which studies workers as active spatial agents shaping capitalism—needs reform to better address exploitation outside the Global North and capitalism's ecological contradictions. The author proposes a theoretical reorientation through three moves: reconstructing the field's foundations around class and exploitation, extending analysis to hyper-exploitation beyond advanced economies, and reframing ecological crisis as a 'class-ecological contradiction' that integrates social reproduction and expropriation into geographic analysis. The paper is conceptual/theoretical rather than empirical, aiming to set a new agenda for the subfield. Twitter discussion so far is limited to the author announcing the publication, with no substantive external critique yet visible.

Discussion: 1 tweets from 1 authors · @TherynDArnold

Computer Science

Comparing Reward Designs for Teaching LLM Agents to Respect Inference Budgets

The paper trains Qwen3.5 agents (4B and 9B) on AppWorld and BabyAI to allocate a fixed inference budget—tokens, API calls, or environment steps—across multi-turn tasks, comparing eleven reward formulations including length penalties used by labs like DeepSeek and Kimi. They find that reward shaping which penalizes overspending substantially beats training on task success alone, with a sigmoid overspend penalty maximizing success within a given budget and a fixed negative penalty best promoting 'natural' budget compliance across unseen budgets; combining frequency- and magnitude-based overspend penalties gives the best overall tradeoff. The authors argue a single trained policy can learn to self-allocate resources without needing an external budget controller. On Twitter, the author's thread framed the core finding as a practical comparison of how different labs' length-penalty schemes perform when adapted to multi-resource budgets, highlighting that no single reward dominates and there's a tradeoff between within-budget success and natural compliance under varying budgets.

Discussion: 1 tweets from 1 authors · @mkirchhof_

Medicine

One Resistance Workout Triggers Effects Far Beyond Muscle, Review Finds

This narrative review synthesizes human studies on the acute, single-session effects of resistance exercise across multiple physiological domains—not just muscle adaptation. The authors report that one bout can trigger myokine release, elevated post-exercise energy expenditure and protein synthesis, enhanced glucose uptake and insulin sensitivity, reduced blood pressure, pain reduction (exercise-induced hypoalgesia), lower state anxiety, improved executive function, and possible sleep benefits. The authors caution these overlapping acute responses vary in evidential strength and shouldn't be assumed to directly cause long-term adaptive or health benefits. Twitter discussion simply highlighted and summarized these diverse acute benefits—improved insulin sensitivity, protein synthesis, lower post-exercise blood pressure, reduced anxiety, better cognition, and sleep—without raising substantive criticism.

Discussion: 1 tweets from 1 authors · @a_garciahermoso

Computer Science

Theory Explains Which Features JEPA Models Keep in Their Latent Space

The paper develops a 'predictive closure' theory for joint-embedding predictive architectures (JEPA), arguing that feature selection in the learned representation depends on relationships between sets of features, not just each feature's individual predictability. This matters because prior theory suggested inherently unpredictable features (like a traffic light's state) would be dropped, but the authors show such a feature can still be retained if it helps predict other, more predictable features. They back this up with controlled experiments showing that representation quality for one feature can be improved indirectly—by giving the predictor side information about a related feature or adding an auxiliary objective on it. The single tweet thread shares a short explainer video and frames the paper as a practical guide to understanding what JEPA latents actually encode, with no substantive criticism raised in the available discussion.

Discussion: 1 tweets from 1 authors · @mosh_levy

Medicine

Review Maps Current Treatment Standards for IgA Nephropathy

This review summarizes the state of care for IgA nephropathy (IgAN), the most common primary glomerular disease, noting that kidney biopsy and MEST-C histological scoring remain the diagnostic and prognostic gold standard while proteinuria and blood pressure control are the key modifiable risk factors. The authors describe an expanding reno-protective drug toolkit beyond RAAS blockade—including SGLT2 inhibitors and endothelin receptor antagonists—alongside more targeted immunomodulatory approaches (mucosal immunity, B-cell cytokines, complement pathway) now in trials, while flagging that recent RCTs raised toxicity concerns with systemic corticosteroids. No validated IgAN-specific biomarker yet exists for diagnosis or monitoring treatment response. The tweet from the journal's account simply shared the open-access review without additional commentary or debate.

Discussion: 1 tweets from 1 authors · @NDTsocial

Other

Fourth-Down Analytics Models Hide Substantial Statistical Uncertainty

The paper argues that NFL fourth-down decisions are typically based on win-probability models fit via machine learning to just a few thousand noisy games, yet these models' uncertainty is rarely acknowledged in sports media. Using bootstrapping to quantify that uncertainty, the authors find the estimated optimal fourth-down decision is far less certain than commonly presented, suggesting analysts' confident prescriptions overstate the precision of the underlying models. On Twitter, commentary centered on reconciling this with real-world coaching debates: one commenter noted that a coach (Shanahan) can still be too conservative on fourth downs even if analytics communities are overconfident in their win-probability estimates, emphasizing that reporting uncertainty matters.

Discussion: 1 tweets from 1 authors · @Stat_Ron

Mathematics

What Is an 'Inductive Mean'? A Primer on Limits of Iterated Means

This note traces the concept of an inductive (or compound) mean—a mean defined as the limit of a convergent sequence of other means—back to Lagrange and Gauss's construction of the arithmetic-geometric mean. The author then surveys how the scalar geometric mean generalizes to symmetric positive-definite matrices, and presents several inductive mean mechanisms for sets of such matrices. The Twitter discussion simply flagged the paper's topic with minimal added commentary, so no substantive critique has emerged.

Discussion: 1 tweets from 1 authors · @FrnkNlsn

Medicine

Review Tackles Peripheral Artery Disease Management in Kidney Disease Patients

This review from the European Renal Association's CKD-MBD Working Group synthesizes evidence on peripheral artery disease (PAD) in chronic kidney disease, where it occurs 4-6 times more often than in the general population and carries elevated mortality risk, especially with chronic limb-threatening ischaemia. The authors cover PAD's multifactorial pathophysiology in CKD—including endothelial dysfunction, inflammation, and mineral bone disorder disturbances like vitamin D deficiency—alongside diagnostic challenges (symptom overlap complicating use of tools like ankle-brachial index) and the limitations of applying general-population treatment guidelines to CKD patients. It concludes by outlining ongoing studies and future therapeutic directions. The discussion on Twitter simply shared the review without added commentary or critique.

Discussion: 1 tweets from 1 authors · @CKJsocial

Mathematics

Mathematicians Prove the m=2 Amplituhedron Is a Positive Geometry

The paper resolves a long-standing conjecture in Positive Geometry, a field spawned by the discovery of amplituhedra in particle physics—semi-algebraic sets that encode scattering amplitudes. The authors prove that the m=2 amplituhedron satisfies the formal definition of a positive geometry, using a new Hodge-theoretic criterion that links work by Brown–Dupont to the original Arkani-Hamed–Bai–Lam framework. Twitter commentary was sparse but enthusiastic, with one widely-shared post calling it a 'mathematical breakthrough,' though no substantive technical critique appeared in the discussion.

Discussion: 1 tweets from 1 authors · @DeivonDrago

Computer Science

AEON Merges Video and 3D World Models into 4D Gaussian Splats

AEON tackles a known tradeoff in world models: video generators (like Sora) produce realistic but 3D-inconsistent scenes, while 3D generators (like Marble) produce geometrically stable but static scenes. The paper proposes a unified framework that operates in a text/image-conditioned video latent space but outputs 4D Gaussian Splatting as an explicit, renderable world representation, trained via a 4D reconstruction model and a domain adapter that distills generative video latents into reconstructive 4D latents. The authors report state-of-the-art results on standard benchmarks, enabling persistent, dynamic 4D environments that can be rendered along predicted camera trajectories. Twitter discussion was limited to a single enthusiastic framing of the paper's core contribution—bridging 'dreamy' video models and 'solid but static' 3D models into something realistic, consistent, and viewable—with no substantive technical pushback yet visible in the thread.

Discussion: 1 tweets from 1 authors · @vincieye

Physics

Particle Theorists Call for Community-Wide Push to Integrate AI

This paper argues that particle theory needs a coordinated, community-scale effort to harness AI's growing capabilities, framing AI not just as a computational tool but as an emerging research partner that is reshaping theoretical physics. The authors outline how particle theory's unique epistemic position complements broader AI-for-physics efforts and issue recommendations for building an AI-enabled theory community tied to the U.S. 'Genesis Mission' initiative. The summary is based on the abstract, as Twitter discussion offered little additional detail. Twitter engagement was minimal and mostly procedural: one commenter simply noted they had endorsed the paper on arXiv, without substantive critique or discussion of its claims.

Discussion: 1 tweets from 1 authors · @WKCosmo

Medicine

Review Maps Treatment Gaps in C3 Glomerulopathy and Ig-MPGN

This review examines C3 glomerulopathy (C3G) and immunoglobulin-associated MPGN (Ig-MPGN), rare kidney diseases distinguished by immunofluorescence patterns but sharing overlapping complement abnormalities and unpredictable clinical courses. The authors note there is no specific approved therapy—only renin-angiotensin inhibitors, lipid-lowering drugs, and renoprotective measures are recommended, while corticosteroids and mycophenolate mofetil are often ineffective and anti-C5 therapy eculizumab shows mixed results. They highlight an unsupervised clustering analysis identifying four distinct pathogenic subtypes at disease onset, suggesting this could guide more targeted complement-inhibiting therapies currently in trials. The Twitter discussion simply flagged the paper as a useful open-access summary of current treatment standards for these conditions, with no substantive critique offered.

Discussion: 1 tweets from 1 authors · @NDTsocial

Biology

Jungle Babblers Shrink Territories 78% During Breeding, Defying Resource Models

Researchers studying cooperatively breeding jungle babblers found that territory size shrinks by roughly 78% during breeding, a change not explained by resource availability but instead reflecting central place foraging around the nest, consistent with a "prudent parent" strategy. Using field observations, movement tracking, and predator-simulation experiments, the authors show lower displacement, longer patch residence, and fewer revisits near nests during breeding, while larger groups counterintuitively showed increased patch residence time, deviating from predictions of existing social patch foraging models. The study argues that predation risk and group size jointly shape space-use decisions in social animals.

Discussion: 1 tweets from 1 authors · @minivets

Computer Science

LLMs Show Hidden Bias Toward Japanese Culture in Open-Ended Questions

Researchers built CROQ, a dataset of culture-related open questions across 24 languages, and prompted LLMs to answer with an implied location. They find models disproportionately default to countries like Japan when answering generic cultural questions, contradicting prior assumptions about Western-centric bias. The study also shows English and other high-resource language prompts yield more diverse answers, while low-resource language prompts skew toward countries where that language is official, and that this Japan-leaning bias emerges during supervised fine-tuning rather than pretraining. Twitter commentary largely highlighted the counterintuitive finding itself—that LLMs seem 'obsessed' with Japan—as a surprising and amusing twist on known cultural bias research, with the tweet's engagement suggesting curiosity rather than deep technical pushback.

Discussion: 1 tweets from 1 authors · @thesupermannx

Computer Science

Scaling study says full backprop beats truncation in looped transformers

Looped transformers reuse the same block of layers multiple times to add effective depth without adding parameters, and are trained via backpropagation through time (BPTT), which is often truncated to the last few loops to save compute. The authors ran nearly 600 training runs varying model size, token budget, and forward/backward loop counts, fitting a unified scaling law in terms of 'effective parameters.' They find that when loop counts are jointly optimized, full (untruncated) BPTT through fewer loops consistently beats truncated BPTT at matched compute, and that the compute-optimal configuration is just two fully-trained loops, which outperforms a non-looped model of equal compute while using about 27% fewer parameters. The tweet thread (from one of the authors) frames the core finding as a practical rule of thumb: truncation only pays off if you're locked into a fixed loop count, otherwise it's better to reduce the number of loops and train all of them with full backpropagation. Engagement centered on the scale of the study (~600 trained models) as a way of empirically settling a training-efficiency tradeoff rather than relying on intuition.

Discussion: 1 tweets from 1 authors · @siyuc3141

Computer Science

AI Learns to Ask Better Questions to Distill Knowledge from Teacher Models

The paper introduces Ask-to-Distill (A2D), a framework for on-policy distillation that trains a question-generating model to synthesize prompts maximizing useful teacher supervision, without needing correct answers, correctness rewards, or per-question student updates. It works by optimizing a reward combining teacher-student disagreement and teacher confidence over the student's own response space, plus diversity controls, to build a bank of synthetic prompts. Across math, code, and agentic tasks, the authors report A2D-generated questions outperform unoptimized student prompts and match or beat real-question banks for closing the teacher-student performance gap. Twitter discussion so far is limited to the authors' own thread introducing the method, framing it as answering 'what should a student ask to learn most from its teacher,' with no independent critical commentary yet surfaced.

Discussion: 1 tweets from 1 authors · @SunJiashuo36