Daily Twitter Digest

Mathematics

Study: Lean Verification of AI-Formalized Proofs Doesn't Guarantee the Original Proof Is Correct

The paper argues that autoformalisation—translating natural-language math proofs into formal languages like Lean for mechanical verification—cannot reliably confirm the original argument is correct, because resolving ambiguities in mathematical natural language is provably harder than any computable problem (the authors place it arbitrarily high in the Solvability Complexity Index/arithmetical hierarchy, i.e., SCI = ∞, exceeding even the Halting problem's SCI = 1). They back this with concrete examples of AI mistranslation, including OpenAI's publicized Navier-Stokes blow-up proof, where they claim the Lean-verified formalization does not actually match the natural-language proof it was supposed to encode. The authors frame this as a general warning against trusting Lean verification as a proxy for the validity of AI-generated mathematical reasoning in prose.

Discussion: 3 tweets from 3 authors · @ValerioCapraro, @gammaofzeta, @gro_tsen

Social Science

Age Gap Between Spouses Affects Survival Differently for Men and Women

Using longitudinal Danish register data covering the entire population's demographic history since 1990, this study applies hazard regression to nearly the full population of married couples, controlling for confounders like education and wealth. The findings show an asymmetric pattern: having a younger spouse is associated with better survival for men but worse survival for women, while having an older spouse is detrimental for both sexes, extending earlier work hampered by methodological and data limitations. Twitter discussion highlighted the striking specific numbers—men with wives 7-9 years younger have 11% lower mortality, while women with husbands 7-9 years younger see mortality rise 20%—framing it as a counterintuitive, sex-asymmetric finding worth sharing widely. The thread largely repeated the headline statistics rather than raising methodological critiques.

Discussion: 1 tweets from 1 authors · @jakozloski

Computer Science

Researchers Urge AI Labs to Protect Chain-of-Thought Monitoring as Safety Tool

This position paper argues that AI systems which reason in natural language offer a valuable but fragile safety opportunity: their chains of thought (CoT) can be monitored for signs of intent to misbehave. The authors acknowledge CoT monitoring is imperfect, like other oversight methods, but argue it shows enough promise to merit further research and investment, and call on frontier model developers to weigh how their design decisions might erode this monitorability over time. The paper is framed as a call to action rather than an empirical result. On Twitter, one of the authors emphasized that advancing this agenda required sustained coordination with third parties, and that the authors had been pushing internally for industry-wide commitments to preserve the ability to monitor AI reasoning before it potentially becomes obfuscated or optimized away.

Discussion: 1 tweets from 1 authors · @balesni

Computer Science

Meta Scales Robotic World Models to 8B Parameters, Finds Predictable Scaling Laws

The paper introduces RoboJEPA, a Joint Embedding Predictive Architecture (JEPA) world model trained on real robot data spanning 12 embodiments, scaled up to 8B parameters—reportedly the largest JEPA predictor built to date. The authors claim its latent rollout error follows a second-order power law in compute, which predicts model quality beyond the fitted scale, and that this "imagination error" correlates strongly with downstream planning performance, serving as a proxy for costly real-robot evaluation. They also show the model can be deployed zero-shot to plan toward a single goal image, solving long-horizon tasks on real hardware. Twitter commentary (via @arankomatsuzaki) highlighted the scale of the model and dataset and the novelty of establishing scaling laws for multi-embodiment robotic world models, with no substantive criticism surfaced in the discussion.

Discussion: 1 tweets from 1 authors · @arankomatsuzaki

Mathematics

DeepMind's AlphaProof Nexus Auto-Solves Open Erdős Problems via Lean Verification

The paper reports the first large-scale evaluation of LLM-generated formal proofs (in Lean) for tackling genuinely open math problems. Their most capable agent, combining LLM proof generation with formal verification, autonomously resolved 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures at a cost of a few hundred dollars per problem, and is reportedly being used in active research across combinatorics, optimization, graph theory, algebraic geometry, and quantum optics. A simpler baseline agent matched the Erdős results but was more expensive on the hardest cases, highlighting how agent design affects cost-efficiency. Twitter commentary noted this contrasts favorably with a recent controversial OpenAI preprint claiming mass problem-solving, since this work reportedly involved substantial engagement from mathematicians (per the paper's methodology section), lending it more credibility in the eyes of at least one commenter.

Discussion: 1 tweets from 1 authors · @hayamizu_lab

Biology

New Lineage-Tracing Map Charts Cell Fate Decisions in Mouse Embryo Development

No abstract is available for this paper, so this summary is based on discussion only. The study, published in Cell, reportedly uses comprehensive lineage tracing to reconstruct a detailed cell fate map of the mouse embryo, aiming to chart how cells commit to different fates during embryogenesis. The sole tweet discussing the work is from one of the authors announcing publication, expressing excitement about the project, with no independent commentary, criticism, or technical discussion yet visible on Twitter.

Discussion: 1 tweets from 1 authors · @JswLab

Computer Science

Researchers Build Trillion-Observation Dataset to Test Financial World Models

The paper argues that building a useful 'world model' for financial markets—one that can represent market-wide conditions, liquidity, volatility and cross-asset relationships for downstream planning—has been held back by narrow evaluations on single tasks and time periods. To fix this, the authors introduce Market-1T, a dataset of nearly one trillion 1Hz observations on US equities from 2008-2025, plus a standardized evaluation protocol, and use it to systematically compare 18 different representation-learning (encoder) strategies across almost two decades of market regimes. A key finding is that encoders with similar predictive accuracy can organize the underlying market state in very different ways, suggesting predictive performance alone is an insufficient proxy for representation quality. The tweet thread from one of the authors frames this as a NeurIPS paper filling a gap in financial representation learning, highlighting JEPA-style self-supervised approaches (referencing their own LeJEPA method) as well-suited to noisy, non-stationary financial data where ground-truth states are unknown. The commentary is promotional/author-driven with no independent critical pushback yet visible in the discussion.

Discussion: 1 tweets from 1 authors · @randall_balestr

Computer Science

DSReg Recovers Individual World Latents from JEPA Representations Without Decoders

The paper claims that joint-embedding predictive architectures (JEPAs) like LeJEPA identify the world's latent variables only up to a linear mixing, leaving individual factors entangled. The authors introduce DSReg, a post-hoc dependency-sparsity regularizer that, under a condition they call Structural Diversity (distinct latents leave distinct dependency 'footprints' on observations), provably disentangles these latents up to signed permutation—without reconstruction, a decoder, or labels. They argue this condition is weaker than prior identifiability assumptions (e.g., non-Gaussianity in ICA) and show across synthetic data, world-model probes, and visual encoders that DSReg improves individual-latent recovery while preserving dense prediction performance, and can be applied to existing trained checkpoints at no extra training cost. Twitter commentary (from a single highly-engaged post) framed this as a significant and elegant follow-up to LeJEPA, calling it a simple post-training fix that fully disentangles the latents LeJEPA leaves mixed—expressed with the quip 'what a time to be Gaussian.' No substantive criticism appeared in the visible discussion, which was largely enthusiastic and light on technical scrutiny.

Discussion: 1 tweets from 1 authors · @randall_balestr

Computer Science

PAMI Anchors Objects to Body Parts for Text-to-HOI Motion Generation

PAMI tackles text-conditioned human-object interaction generation, where prior methods treat human and object motion as separate trajectories and often produce object drift, missed contact, or penetration. Inspired by the Hough Transform, PAMI represents object motion as a weighted aggregation of votes from multiple body-part anchors, learned via PamiVAE, then generates coarse interactions from text before refining contact geometry with a hybrid long-range/short-range surface-sensing module. On the InterAct benchmark, the authors report more faithful interactions and 14.5% higher contact recall than prior state-of-the-art methods. The tweet thread is a single promotional announcement from the authors introducing the project page, paper, and code, with no independent technical discussion or critique yet visible.

Discussion: 1 tweets from 1 authors · @coralli04102979

Social Science

Archaeology Explores How Medieval 'Sensory Regimes' Shaped Material Culture

This Antiquity article argues that medieval European archaeology should move beyond cataloguing the five physical senses and instead study 'sensory regimes'—how artefacts, burial practices and urban spaces were used to enforce or resist social, political and religious authority from the sixth to seventeenth centuries. Drawing on varied case studies, the author contends that material culture actively shaped sensory experience in ways tied to power structures. One example highlighted is medieval spectacles, invented in late 13th-century northern Italy, which were valued not just for correcting eyesight but as symbols of 'moral vision,' reflecting the era's philosophical link between sight and perceiving divine light.

Discussion: 1 tweets from 1 authors · @AntiquityJ

Medicine

Study Reports COVID Spike Protein Lingers in Skull Marrow and Meninges

No abstract was available for this paper, so this summary is based solely on Twitter discussion. According to the tweet, the study reports that SARS-CoV-2 spike protein can persist in skull marrow, skull-meninges channels, meninges, and nearby brain tissue of people who had COVID-19—even after the virus itself is no longer detectable—and that spike protein alone appears sufficient to drive brain inflammation and worsen neurological injury in mice. The commentary highlights this as evidence for a mechanism by which viral remnants could contribute to long-term neurological effects of COVID-19, though the claims have not been independently scrutinized in the discussion beyond relaying the findings.

Discussion: 1 tweets from 1 authors · @CatsRule2023

Social Science

Historians Document Alleged Israeli Biological Warfare in 1948 War

This paper by historians Benny Morris and Benjamin Z. Kedar reportedly draws on newly available archival documents to argue that Israeli forces engaged in biological warfare during the 1948 war, allegedly including efforts to contaminate water sources as a military tactic against Arab towns and retreating Egyptian units. No abstract is publicly available, so this summary is based on secondary discussion rather than the paper's own stated methodology or findings. Twitter commentary focused on sharing corroborating sources, including a contemporaneous Haaretz report and UN Security Council records from 1948, treating the paper as a significant historical revelation rather than raising substantive methodological critique.

Discussion: 1 tweets from 1 authors · @academic_la

Computer Science

Predicting Post-Training Coding-Agent Skill from Base Model Checkpoints

The paper proposes a way to forecast whether a base LLM is worth expensive agentic post-training for coding tasks, without running a full agentic harness. By replaying successful post-trained agent trajectories and finding the 'decisive step' where a patch first flips a repo's tests from failing to passing, the authors build three lightweight screens (bits-per-byte on the certified action, a multiple-choice test against rejected alternatives, and prefix-conditioned pass@K) that measure a base checkpoint's latent potential at that step. Across ten base/post-trained model pairs, these screens closely track actual post-trained SWE-bench Verified pass@1 scores, suggesting they could let researchers convert existing agentic benchmarks into cheap base-model evaluations.

Discussion: 1 tweets from 1 authors · @rosinality

Biology

Study Links Preference for Group Hierarchy to Smaller Anterior Insula Volume

Using structural 3T MRI in 99 young adults, researchers conducted a conceptual replication examining neuroanatomical correlates of Social Dominance Orientation (SDO), the trait reflecting preference for group-based hierarchies. They found a negative correlation between SDO scores and gray matter volume in the bilateral anterior insula—a region tied to empathy and socio-emotional processing—and a positive link between SDO and Sensitivity to Reward (SR), with insula volume plus SR explaining 29% of variance in SDO scores. The authors frame this as supporting a prior, not-yet-replicated finding implicating the insula in SDO.

Discussion: 1 tweets from 1 authors · @metodogolo

Earth & Climate

African Mantle Blob May Be Dynamically Unstable, Study Suggests

This paper, published in Nature Geoscience, models the African Large Low-Shear-Velocity Province (LLSVP)—a massive seismic anomaly deep in Earth's lower mantle—and argues that its intrinsic density is low enough to make the structure gravitationally unstable rather than a static, long-lived pile. No abstract was available, so this summary is based on discussion of the paper's title and framing rather than its full methodology or findings. The authors' modeling approach reportedly uses density and shear-wave velocity constraints to assess whether such provinces can persist over geological time or are prone to breaking apart and reforming. Twitter discussion was limited to a single post linking the finding to the poster's own 'ECDO Theory' and noting that LLSVP centroid behavior appears consistent with it, without providing independent scientific critique or elaboration from other commenters.

Discussion: 1 tweets from 1 authors · @EthicalSkeptic

Engineering

500-Hour Human Touch Dataset Boosts Robot Manipulation Success

TouchScale introduces a 500-hour dataset of egocentric human interaction recorded with a single unified wearable setup, synchronizing RGB-D video, wrist RGB, and dense bimanual tactile sensing across ~2K task descriptions. The authors report that training on this data improves zero-shot contact prediction on unseen tactile sensors, boosts action recognition accuracy, and raises real-world robot success rates on four contact-rich manipulation tasks from 22.5% to 57.5%, with performance scaling upward as more data is used. The key claimed advantage is using a consistent sensor and collection protocol at scale, unlike prior datasets that merge heterogeneous sensors and procedures.

Discussion: 1 tweets from 1 authors · @EdisonLeeROC

Engineering

Survey Maps Neuromorphic Computing's Inroads into Radar and Radio Signal Processing

This review paper by Hamrell, Sjögren and Ovrén surveys recent work applying brain-inspired neuromorphic computing to radar and radio systems, arguing that such approaches can cut the energy and latency costs of digital signal processing while enabling AI methods in power-constrained settings. The authors cover applications in frequency-modulated continuous-wave radar, synthetic aperture radar, and RF signal classification, aiming to consolidate the field and connect researchers across radar and neuromorphic communities. Twitter commentary was limited to a note from BrainChip highlighting that the survey references their Akida chip among commercial neuromorphic hardware and cites AFRL-affiliated radar research, with no substantive critical discussion observed.

Discussion: 1 tweets from 1 authors · @BrainChip_inc

Biology

New Tokenizer Segments Neural Data by Behavior, Not by Neuron

The paper argues that neural foundation models fail to generalize across recording sessions because they tokenize individual neurons and sessions, which never repeat. Instead, the authors propose Tokenization with States (TWS), which segments each trial at behavioral events (like stimulus or movement onset) and tokenizes the resulting population-level geometric states, with no neuron- or session-specific embeddings. On held-out IBL sessions TWS reportedly decodes movement where a per-neuron foundation model performs at chance, and after training only on mice it transfers with a simple linear probe to macaque and Utah array recordings, achieving a Matthews correlation of 0.23 on reach direction decoding versus 0.01 from event timing alone.

Discussion: 1 tweets from 1 authors · @MillerLabMIT

Computer Science

Chunked KV-Cache Compression Creates Periodic 'Weak Spots' in Long-Context Retrieval

The paper identifies a systematic flaw in chunked KV-cache compression schemes used for efficient long-context inference: a token's retrievability depends heavily on its 'phase', i.e. its position relative to compression-window boundaries, causing retrieval accuracy to swing by up to 40 percentage points depending on phase even though average benchmark scores look fine. Through pretrained model families and causal interventions, the authors show attention components specialize unevenly across phases, and argue that gradient dynamics favor this sharp specialization, meaning standard evaluation can mask systematic positional failures. On Twitter, the main discussion point was that this research helps explain real-world behavior of commercial models: one commenter linked the findings to why DeepSeek's chunked/compressed attention mechanism (CSA) seemed flawed, and why a later version reportedly reduced its compression ratio significantly as a result.

Discussion: 1 tweets from 1 authors · @liuliu

Medicine

Meta-Analysis Revisits Claims of Sex-Differential Mortality Effects from DTP Vaccine

This paper (no abstract available, so the summary below is based on discussion and title alone) is a meta-analysis attributed to researchers including Peter Aaby that reportedly tests the hypothesis that the diphtheria-tetanus-pertussis (DTP) vaccine has 'non-specific effects' that increase mortality specifically among female recipients, compared to males. The DTP non-specific effects hypothesis has been a recurring and contested theme in Aaby's research program for years, arguing that some vaccines may have broader immunological effects beyond their target disease that differ by sex. The lone tweet discussing this DOI links it to Aaby's broader body of work (citing several other papers and PMIDs) and frames it through the lens of the 'Institute for Scientific Freedom,' an organization associated with vaccine-skeptical advocacy. Because the engagement consists of a single tweet with minimal elaboration, it's hard to gauge independent scientific reaction; readers should treat the framing as commentary from a specific advocacy-adjacent source rather than a neutral assessment of the paper's methodology or conclusions.

Discussion: 1 tweets from 1 authors · @BanounHelene