Daily Twitter Digest

Computer Science

Sakana AI's QD-Rec Pairs Evolutionary Search with LLMs for Open-Ended Recommendations

QD-Rec reframes recommendation as building a diverse, high-quality set of item-plus-reason pairs rather than a single ranked list, targeting high-stakes decisions like M&A targets or customer selection where the "why" matters as much as the "what." It combines a contextual bandit for item selection with a foundation model that generates natural-language hypotheses, using quality-diversity evolutionary search and an archive of past hypotheses as stepping stones; new entries are kept only if novel or better than their nearest neighbor. Across M&A, customer targeting, and paper/book recommendation benchmarks, the authors report QD-Rec achieves the highest mean quality and builds the largest archive of viable recommendations in most settings. Discussion so far is mostly Sakana AI's own announcement thread introducing the method, blog post, and code, framing it as an extension of open-ended foundation-model search into practical decision-support recommendation.

Discussion: 1 tweets from 1 authors · @SakanaAILabs

Medicine

Ecological Study Links Glyphosate-Resistant Soybean Expansion to Cancer Mortality in Argentina

This paper is a small-area ecological study examining the relationship between soybean-cultivated area (a proxy for glyphosate herbicide exposure, used with glyphosate-resistant GM soy) and cancer mortality rates across Argentine provinces and departments from 2012 to 2022. No abstract was available, so this summary is based on discussion of the paper rather than its stated findings; according to a tweet summarizing the research, the authors report that regions with greater soybean cultivation—and thus presumably higher glyphosate exposure—show higher cancer death rates. As an ecological study, such designs can identify population-level correlations but cannot establish individual-level causation, a limitation inherent to this type of analysis.

Discussion: 1 tweets from 1 authors · @OrganicNewsClip

Medicine

ADHD Decision-Making Deficits Linked to Suboptimal Choices, Not Risk-Seeking

This paper argues that decision-making problems in ADHD are not driven by a preference for risk per se, but by a tendency to choose options with lower expected value (suboptimal choices) that happen to be confounded with risk in most experimental tasks. A meta-analysis (48 studies, ~1,100 ADHD and 1,100 control participants) found ADHD-control differences only when the risky option was also the suboptimal one, and a new experiment (40 ADHD adults vs. 40 controls) confirmed that ADHD participants made more suboptimal, but not more risky, choices. The authors conclude that interventions and theories framing ADHD decision-making deficits as 'risk-seeking' may be mischaracterizing the underlying cognitive issue. A Japanese-language tweet summarizing the paper highlighted the finding as explaining why people (self-identifying with ADHD traits) favor immediate results and struggle to follow through on long-term goals, framing it as a brain-based decision-making mechanism rather than impulsive risk-taking.

Discussion: 1 tweets from 1 authors · @rin_rinnak

Computer Science

RemoteANN Cuts Object Storage Requests for Vector Search by 16x

RemoteANN is a vector search engine designed specifically for cloud object storage (like S3), where network latency and per-request costs make conventional vector indexes inefficient. It pairs an IVF index with RaBitQ quantization and data reordering to improve locality, then uses cluster pruning, range-GET requests, request merging, and NVMe caching to minimize discrete storage calls; a new clustering method (QGKMeans) also speeds up index construction. The authors report up to 5.9x throughput gains and 16.1x fewer object storage requests at 97% recall on real-world datasets compared to existing baselines. The tweet discussing this paper was brief but enthusiastic, calling it a strong read with broad applicability—suggesting its techniques could be iterated on by other projects building disaggregated storage-compute vector search systems.

Discussion: 1 tweets from 1 authors · @Hi_Mrinal

Computer Science

HCI Paper Proposes "Creative Reading" as Alternative to AI That Reads For You

The paper argues that AI reading-augmentation tools often frame reading as mere information transmission—"reading to discard"—outsourcing interpretation to machines, which changes what reading accomplishes, especially in scholarly contexts where forming one's own interpretation is the actual work. Drawing on literary/narrative theory and creativity-support research, the authors propose "creative reading" as a design goal: tools that help readers construct both interpretations and their own growth as readers, rather than just extracting summaries. A design space is sketched around supporting plural readings and reader transformation over time rather than efficiency alone. On Twitter, a widely shared commentary highlighted the paper's framing of "AI reads for me" vs. "AI reads with me" as a useful new axis layered onto the classic reader-centered vs. author-centered interpretive divide, creating a four-quadrant lens for categorizing current AI reading tools. The commenter found the paper's taxonomy of "AI reads with me" tools especially interesting, using analogies like running shoes (reducing friction without changing the activity) versus dumbbells (adding beneficial resistance, e.g., forcing readers to paraphrase or predict content) to distinguish assistive from augmentative design approaches.

Discussion: 1 tweets from 1 authors · @lijigang

Physics

New Rotating Black Hole Solutions with Scalar Hair Break Zeroth Law

The paper presents new stationary, axisymmetric black hole solutions in a subclass of Horndeski scalar-tensor gravity (no non-minimal curvature coupling), computed numerically via spectral methods without assuming an explicit metric ansatz. These black holes differ from Kerr, carry a primary hair parameter tied to linear time dependence of the scalar field, and possess a Killing horizon that is non-circular—a geometric feature leading them to violate the zeroth law of black hole thermodynamics (i.e., surface gravity is not constant over the horizon). Twitter discussion was limited to the authors' own announcement, highlighting the novelty of non-circular hairy rotating solutions and their thermodynamic law violation as the key takeaway, with no independent critical commentary yet visible.

Discussion: 1 tweets from 1 authors · @lorel_sg

Computer Science

SteerAct Boosts Robot Policy Success via Single-Feature Test-Time Steering

SteerAct proposes a few-shot method to improve pretrained robot action models at test time without updating weights, using a sparse autoencoder to find candidate features and a 'temporal descriptor' of activation-order patterns to select one task-specific feature and decide when/how strongly to steer it during execution. Using only 24 outcome-labeled rollouts per task, the authors report roughly 10% relative improvement across benchmarks (π0.5, Fast-WAM, RoboTwin 2.0, DexJoCo) and real-world manipulation tests, with the frozen VLA/WAM model unchanged. Twitter commentary highlighted the real-robot result of a 31.1%→52.2% success rate jump (about 68% relative improvement) from steering just a single feature with 24 test-time rollouts per task, framing it as a notable example of lightweight, interpretable test-time intervention from a Stanford/MIT/Harvard collaboration; no substantive critical pushback was present in the visible discussion.

Discussion: 1 tweets from 1 authors · @alex_kai2020

Biology

New Math Framework Models DNA's Codon Code on Both Strands at Once

This paper reframes the genetic code using double-stranded DNA rather than treating one strand as a passive complement. By encoding nucleobases in Z/4Z, the authors represent all 64 codons as points on a discrete torus and show the Watson-Crick complement operation is a distance-preserving involution that splits codon space into 32 'mirror' pairs and two 32-dimensional subspaces capturing static duplex symmetry versus strand-specific coding bias. Using information-theoretic decompositions tested across nine reference genomes (nearly 10 million codon transitions) and a human chromosome 17 locus, they claim 74–89% of the sense strand's synonymous codon usage pattern is actually explained by the antisense strand's amino acid identity, and that genomes operate far from the simple 'lumped' Markov behavior standard models assume.</br></br>The tweets are from the paper's own authors announcing the preprint, so there is no independent commentary yet—engagement is limited to promotional posts rather than critical discussion from outside readers.

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

Computer Science

Study Finds 96% of ChatGPT's 'Memories' Are System-Generated, Not User-Chosen

Researchers analyzed 2,050 memory entries from 80 real ChatGPT users to understand how the platform's 'Memory' feature builds an algorithmic self-portrait from private conversations. They found that 96% of memories were created unilaterally by the system rather than explicitly by users, 28% contained GDPR-defined personal data and 52% included inferred psychological insights, while 84% remained faithfully grounded in actual conversation content. The authors propose 'Attribution Shield,' a framework to flag sensitive inferences and suggest query rewrites that preserve utility while limiting unwanted personal data capture. The single tweet referencing the paper (159 engagements) simply links to it without added commentary, so there is no additional discussion to summarize beyond the paper itself.

Discussion: 1 tweets from 1 authors · @thesupermannx

Other

Open-Access Book Introduces Inquisitive Semantics Framework

This book presents inquisitive semantics, a logical framework that moves beyond treating sentence meaning purely as truth-conditions, instead modeling meaning as a process of raising and resolving issues during information exchange. The authors argue this better captures how language is actually used in conversation, and apply the framework to questions, coordination, modals, conditionals, and intonation. The Twitter discussion is a simple pointer to the free, open-access availability of the book rather than a critical discussion of its contents.

Discussion: 1 tweets from 1 authors · @adammcroom

Computer Science

Mechanistic Study Finds Transformers Build World Models Despite Apparent Failures

The paper examines TaxiGPT, a transformer trained on random walks through a simulated Manhattan, whose navigation failures had previously been read as evidence it lacks a coherent internal map. Using mechanistic analysis and causal interventions, the authors show the model actually represents intersections, streets, and its own position, and uses a 'goal compass' to navigate—but suffers from interference between superposed intersection features that disrupts localization. They introduce 'affordance packing' (grouping intersections sharing legal moves) as a partial fix and propose mechanistic indicators to track when world-modeling capacities emerge during training, arguing the field should move from binary 'does it have a world model' questions to studying the interacting mechanisms involved.

Discussion: 1 tweets from 1 authors · @MikeTamir

Computer Science

Muse Glimmer: Open-Weight 30B Agentic Model for Local Devices

The technical report introduces Muse Glimmer, a 30B-parameter open-source (Apache 2.0) causal language model with a perception encoder, designed to run agentic workloads—multi-step reasoning, tool use, and failure recovery—entirely on consumer hardware via 4-bit quantization. It accepts text and image inputs, was trained on over 100 languages, and reportedly outperforms similarly sized models like Gemma4-31B and Qwen3.6-27B on agentic benchmarks. Twitter discussion was limited to the announcement itself, with the author framing it as a step toward capable, local, cloud-independent AI agents and inviting community use; no independent technical scrutiny or skepticism was visible in the available commentary.

Discussion: 1 tweets from 1 authors · @yorambac

Mathematics

50-Year-Old Problem Solved: Ergodicity of Nonpositively Curved Geodesic Flows

The authors prove that for any closed genus-≥2 surface equipped with a C^{3+Lip} metric of nonpositive curvature, the geodesic flow is ergodic with respect to Liouville measure, resolving a longstanding open question in dynamical systems. They note a forthcoming follow-up extending the result to smooth rank-one manifolds in all dimensions. Twitter commentary from one of the authors frames the result as closing a 50-year-old open problem, with no substantive criticism or skepticism voiced in the discussion.

Discussion: 1 tweets from 1 authors · @janarhertz

Mathematics

New Simple Symmetric Venn Diagrams Found for 17, 19, and 23 Curves

The paper reports new simple, rotationally symmetric Venn diagrams with 17, 19, and 23 curves—extending a result previously known only up to 13 curves—found via a Metropolis random walk over rotation-invariant quadrangulations of the sphere, starting from a classical construction with multiple crossings resolved. Each diagram comes with a machine-checkable certificate, and representative cases at each size were formally verified in Lean 4; the authors note all the new diagrams are non-monotone, explaining why earlier crossing-sequence search methods missed them. A single tweet highlighted the find, framing it as a result of "two AI models" collaborating with a human, though the paper's abstract describes the discovery method purely as an algorithmic Metropolis walk with no mention of AI models communicating.

Discussion: 1 tweets from 1 authors · @pickover

Medicine

Two Echo Criteria Flag Significant Mitral Regurgitation with High Specificity

This study compared echocardiography (Echo) and cardiac MRI (CMR) performed near-simultaneously in 177 patients to evaluate how well standard ASE echocardiographic criteria identify significant chronic mitral regurgitation (MR). The authors report that just two of the ASE criteria achieved 91–95% specificity for detecting significant MR when benchmarked against CMR, suggesting a simplified diagnostic approach could reduce reliance on extensive multiparametric quantification. No abstract was available, so this summary is based on the paper's title and discussion. Twitter commentary framed the finding as a 'less is more' result, highlighting that smarter integration of a minimal set of echo criteria could streamline diagnostic workflows and increase confidence in significant MR calls without added complexity; no substantive criticism of the methodology was raised in the discussion.

Discussion: 1 tweets from 1 authors · @HeartToProve

Biology

Critically Endangered Jamaican Cave Velvet Worm Rediscovered After 50 Years

Researchers report five specimens of Speleoperipatus spelaeus, a troglobitic velvet worm known only from Pedro Great Cave in Jamaica, found at its type locality for the first time since the species was described from four specimens collected in the early 1970s. The paper documents the find along with behavioral observations, and notes that only a handful of possibly related specimens have otherwise been seen (in a different cave) in the intervening decades. The authors call for future phylogenetic work to clarify the species' evolutionary relationships and its status as one of the rarest known onychophorans. A co-author shared the paper on Twitter, highlighting it as the first documented sighting at the type locality since 1974, framing it as a notable rediscovery for Jamaican cave biodiversity.

Discussion: 1 tweets from 1 authors · @RoostersWorldja

Mathematics

Theory explains why cooldown phases boost SGD scaling laws

The paper analyzes the last-iterate risk of SGD on streaming least-squares regression under a discretely distributed feature model (Slock's model), which allows exact dynamics to be derived eigendirection by eigendirection. Under power-law source/capacity conditions, the authors compute exact finite-time risk and asymptotics for constant and Warmup-Stable-Decay (WSD) schedules, showing that any nonzero cooldown fraction—however brief—improves the scaling exponent over a constant learning rate, with linear decay minimizing the resulting constant; they also show optimal risk depends on batch size and horizon only through their product, up to a critical batch size. On Twitter, the author (posting as his final paper from postdoc) framed it as rigorous theoretical backing for the popular WSD schedule, crediting intern Noé Weeks and Francis Bach as collaborators; engagement was modest and mostly congratulatory with no substantive critical pushback visible in the discussion.

Discussion: 1 tweets from 1 authors · @FSchaipp

Computer Science

KronLSTM Proposes a Gentler Update Rule for Linear Attention Memory

Researchers from Sepp Hochreiter's lab introduce KronLSTM, a linear attention architecture built on a new 'Kronecker Delta Rule' for updating associative matrix memory. Unlike standard delta-rule updates, this revision modifies memory along only a single direction in matrix space, reducing interference with previously stored associations. The authors show mathematically that this yields bounded, stable memory dynamics without extra normalization, and report that KronLSTM is competitive with or outperforms DeltaNet-family models, especially on long-context benchmarks. Twitter discussion was limited but flagged it as a notable new linear attention variant from Hochreiter's group aimed at reducing update interference, extending the xLSTM line of work.

Discussion: 1 tweets from 1 authors · @cosminnegruseri

Biology

DiscoveryRSI: Self-Improving AI Agent for Generating Scientific Hypotheses

The paper introduces DiscoveryRSI, a framework that formalizes 'valuable hypotheses' as evidence-grounded claims with feasible investigations, explicit decision criteria, and alternative explanations, then uses multi-agent evolutionary search to build hypothesis trees and recursively improve itself via credit attribution, workflow replay, and rediscovery. Alongside a new benchmark (DFM-Bench), the authors report that hypothesis quality scores rose from 4.86 to 5.39 across two self-improvement rounds in four life-science domains, and that their 27B model outperformed other research agents in both automated and human expert evaluations, including a higher rate of matching hypotheses to subsequently published research. The single tweet highlighted announces the open-sourcing of the DiscoveryRSI-27B model, code, and paper, along with a live benchmarking arena where users can watch hypothesis generation happen in real time.</br>Twitter discussion was limited to a promotional announcement rather than substantive critique, pointing followers to the live arena, code, and model releases without independent commentary on the claims.

Discussion: 1 tweets from 1 authors · @abbyxuu

Physics

New Free-Field Realizations for Supersymmetric W-Algebras

The paper extends a recently proposed recursive construction method (due to Ide, Ito, and Miyazaki) to the N=1 supersymmetric W-algebras associated with Lie superalgebras B(n-1,n) and B(n,n). The authors derive explicit free-field realizations for several cases, including previously unknown ones for SW B(1,2) and SW B(2,2), and use these to compute central charges, conformal weights, and local integrals of motion (I1, I3, I5). This is a technical contribution to the mathematical physics of W-algebras and conformal field theory.

Discussion: 1 tweets from 1 authors · @nao_hepth