Daily Twitter Digest

Computer Science

Sakana AI's Agentic Scientist Orchestrates Parallel Coding Agents for Research

Sakana AI presents The Agentic Scientist, a distributed pipeline that coordinates pools of coding agents (e.g., Codex, Claude Code) through the full research lifecycle—idea generation, planning, sandboxed experiments, ablations, analysis, and report writing—under a meta-controller that enforces rubric-based verification at each stage. The system is designed to work either as a human co-scientist or fully autonomously; the authors report a co-scientist paper accepted at ALIFE 2026 and autonomous runs on open ML problems, with ablations suggesting that the orchestration layer, not just base model capability, drives reliable scientific autonomy. This is the third iteration of Sakana's AI Scientist line, following earlier versions and a Nature-published result. On Twitter, the Sakana team and collaborators framed this as a shift from single-agent paper-writing to a multi-agent 'scaffold' with shared memory and evidence-based verification meant to counter myopic hill-climbing in autonomous research agents. Discussion was largely promotional/descriptive (thread explanations, GitHub/blog links) rather than critical, with no substantive independent skepticism visible in the tweets provided.

Discussion: 4 tweets from 3 authors · @hardmaru, @SakanaAILabs, @SakanaAILabs, @RobertTLange

Physics

AI Agents Reportedly Solve Research-Level Theoretical Physics Problems

The paper introduces a benchmark of 50 unpublished, research-level theoretical physics problems with verifiable answers, each normally solvable by a world expert in about a day. Testing off-the-shelf AI agents with tools like literature search and code execution, the authors report that their best agent achieves 98% accuracy for under $3 per problem, with only small accuracy drops when web search and background papers are withheld—suggesting the agent relies heavily on memorized literature rather than genuine search. The authors note this strong performance contrasts with the apparent absence of agent-led breakthroughs in theoretical physics so far. Twitter discussion was limited to identifying one of the authors, Praneeth Netrapalli of Google DeepMind, and sharing links to the preprint, without substantive critical engagement with the results.

Discussion: 2 tweets from 2 authors · @stellensatz, @Dr_Singularity

Chemistry

15-Year Effort Yields Coating-Free SEM for Native Organic Materials

This Accounts of Materials Research paper summarizes 15 years of work developing coating-free, low-voltage scanning electron microscopy to image organic and hybrid materials in their native state, avoiding the mass addition and structural distortion caused by conventional conductive metal coatings. By optimizing electron optics, landing voltage, specimen preparation, and conductive substrates like ITO, the authors resolved individual molecules (e.g., avidin, gold nanoparticles) and tracked the same specimen across scale, depth, and time—from fracture surfaces down to nanofibers, through buried composite layers, and during beam-driven structural changes like fullerene nanofilm folding. The authors frame this as part of a broader 'cinematic chemistry' philosophy for observing materials in their unaltered chemical states. The tweets, from the authors themselves, emphasize the long gestation of the technique and credit experimental execution, framing it as enabling observation of meso-hierarchical structure across multiple dimensions without the artifacts introduced by coatings; no independent critical commentary appears in the discussion.

Discussion: 2 tweets from 2 authors · @mesohierarchy, @khng1980

Earth & Climate

Foraminifera from Great Barrier Reef Cores Reveal Sea-Level Drop Before Last Glacial Maximum

No abstract was available, so this summary is based on discussion and press materials only. The paper reportedly uses large benthic foraminiferal assemblages from Great Barrier Reef drill cores (collected via the Integrated Ocean Drilling Program) to reconstruct global mean sea level between roughly 40,000 and 30,000 years ago, documenting a precise falling trend in sea level during the global cooling that preceded the Last Glacial Maximum. Twitter activity consisted mainly of institutional press-release announcements from Tohoku and Nagoya Universities framing this as a high-precision paleoclimate reconstruction relevant to understanding past global cooling and sea-level change, with no substantive scientific criticism raised in the discussion.

Discussion: 2 tweets from 2 authors · @tohoku_univ, @NagoyaUniv_info

Chemistry

SolvPred Tool Inverse-Designs Solvent Mixtures via Hansen Parameters

SolvPred is an open-source Python framework that inverse-designs solvent mixtures by targeting desired Hansen solubility parameters while respecting practical experimental constraints like measurement precision, boiling points, and miscibility. Applied to a polyimide condensation reaction, it screened 171,871 possible ternary mixtures and identified seven viable candidates; the best achieved a fivefold yield improvement (16% to 71%) at a milder reaction temperature compared to the conventional single-solvent approach. The authors position it as a general tool for property-guided solvent optimization in data-driven chemistry workflows. The tweet highlighting the paper emphasized the practical appeal of replacing intuition-based solvent selection with a systematic search, noting not just the yield boost but also a reported 97% reduction in solvent cost — a detail from the discussion not explicitly covered in the abstract.

Discussion: 1 tweets from 1 authors · @m_ichi1271

Earth & Climate

2026 State of the Climate Report: 7 of 16 Planetary Vital Signs at Record Extremes

Ripple and colleagues' annual update tracks 16 climate-related indicators—covering greenhouse gases, temperatures, oceans, ice, energy balance, and wildfire—alongside human drivers of change. The report finds 7 of 16 (44%) indicators at record extremes, with several showing accelerating trends or abrupt shifts that raise concerns about climate instability, feedback loops, and tipping points; fossil fuel combustion remains the dominant driver, though renewable electricity and battery storage expansion are noted as encouraging but insufficient progress. On Twitter, lead author William Ripple highlighted the report's headline statistic as a call to attention, framing it as evidence Earth's climate system is entering increasingly uncharted territory.

Discussion: 1 tweets from 1 authors · @WilliamJRipple

Biology

Enoki Mushroom Extract Shows Anti-Cancer, Immune Effects in Lab Study

This paper (no abstract available, so details are based on discussion) reports metabolic profiling and systems-biology analysis of an extract from Flammulina filiformis (enoki mushroom) strain Fv19, combining computational modeling with lab experiments to characterize compounds with both anti-cancer and immune-modulating activity—apparently tested in mouse models rather than humans. Twitter discussion centered on long-standing anecdotal associations in Japan between enoki mushroom consumption and lower cancer mortality, citing Nagano Prefecture's notably low cancer death rate (~97 per 100,000 vs. ~160 prefecture-wide) among enoki-farming communities, with commenters expressing hope the findings could eventually lead to human clinical trials.

Discussion: 1 tweets from 1 authors · @AaronOtsuka

Computer Science

AgentValve Bounds Sybil Attacks in AI Agent Payment Networks

The paper introduces AgentValve, a protocol for authorizing payments between autonomous AI agents in permissionless markets, where adversaries could otherwise fabricate reputation via cheap fake identities. It models reputation as finite capacity along trust paths in a directed graph, using a smart contract to atomically reserve capacity so that an honest agent's financial exposure is bounded by what it extends into adversarial regions, regardless of how many fake identities an attacker creates. The authors combine payment-backed capacity creation with bounded route length to prevent capacity depletion attacks, and introduce a witness-commit mechanism that cuts on-chain verification to linear work with constant state per payment. Evaluated on six months of Ethereum data plus synthetic attacks (up to 10,000 Sybil identities), the system kept exposure bounded while reducing admission costs by up to 40.8%, with per-payment costs around $0.04-$0.21. Twitter discussion was limited, mainly a promotional note highlighting the collaboration between Google, Mysten Labs, and Purdue University, with congratulations to one of the researchers — no substantive technical critique appeared in the available commentary.

Discussion: 1 tweets from 1 authors · @kostascrypto

Computer Science

Nvidia's Robota-Embed Fuses Vision, Language, Action for Robot Retrieval

Nvidia researchers introduce Robota-Embed, described as the first vision-language-action embedding model that jointly encodes vision, language, and action for both current robot states and their future continuations, trained via predictive retrieval alignment on a 1.7M-anchor corpus. The paper reports it outperforms action-free embedding baselines (like Gemini Embedding 2) substantially on their new Robota-Embed-Bench, and shows two downstream uses—Embed2ICL (retrieval-augmented in-context policy conditioning) and Embed2Evolve (task-local memory growth around a frozen policy)—boosting success rates across DROID-Sim, LIBERO, and ManiSkill benchmarks while cutting token usage. Ablations suggest removing the action modality hurts retrieval more than removing vision, underscoring the value of action-aware embeddings. Twitter discussion was limited to a single widely-shared announcement post highlighting the project as a new Nvidia release, with no substantive critical commentary surfacing in the engagement.

Discussion: 1 tweets from 1 authors · @arankomatsuzaki

Engineering

Humanoid Robots Play Table Tennis Doubles with Coordinated Footwork

PingpongDuo is a hierarchical control framework enabling two humanoid robots to play table tennis doubles, where teammates must strike in strict alternation while avoiding collisions in a shared workspace. The system combines a team-level planner that selects positioning targets while filtering risky moves, with individual footwork and striking skills trained from human motion references and refined via simulator feedback calibrated to real ball-flight data. In simulation it improves return success rate by 13.2% over prior methods, and in real-world tests a robot executes a 2.15m clearance-and-recovery cycle in 2.4s at 2.23 m/s peak speed, while two robots sustain 10+ consecutive alternating strikes.

Discussion: 1 tweets from 1 authors · @arankomatsuzaki

Computer Science

UniForge3D Unifies 3D Reconstruction and Generation to Fix Coordinate Mismatch

UniForge3D tackles building complete 3D assets from sparse, unposed images by jointly modeling reconstruction (which preserves observed geometry) and generation (which fills in unseen regions via learned shape priors) within a shared anchor-guided coordinate space, resolving the mismatch between camera-relative and canonical coordinate systems. The method exchanges features bidirectionally between branches, predicts per-view point maps alongside a complete sparse voxel structure, and selects the most informative input view per voxel to guide detailed shape generation. The authors report state-of-the-art appearance and geometry results on the Toys4K and GSO benchmarks. Twitter discussion was limited to a brief announcement sharing project page, code, paper, and model links, without substantive technical critique.

Discussion: 1 tweets from 1 authors · @camenduru

Computer Science

NVIDIA's Star Elastic Trains Multiple Nested LLMs in One Run

Star Elastic is a post-training method that embeds N nested submodels inside a single parent reasoning LLM via one training job, rather than training each model separately. Applied to NVIDIA's Nemotron Nano v3 MoE models, it produces smaller nested variants (e.g., 23B and 12B from a 30B parent) that match or beat independently trained baselines, claims a 360x compute reduction versus training from scratch, and supports dynamic per-phase model selection (different submodels for 'thinking' vs 'answering') to improve the accuracy-latency tradeoff, including quantized FP8/NVFP4 deployments. On Twitter, the main reaction concerned a specific derived example—an NVIDIA-made 74B-A9B 'Bolt' checkpoint compressed from a 120B-A12B 'Super VL' parent—with one commenter suggesting similar elastic-style compression could plausibly be replicated by users on any large MoE model today, implicitly questioning how novel or NVIDIA-specific the technique really is.

Discussion: 1 tweets from 1 authors · @LottoLabs

Medicine

ACC 2026 Statement Sets Exercise Guidelines for Cardiovascular Health

No abstract is available, so this summary is based on discussion only. The paper is described as a new American College of Cardiology scientific statement on exercise prescriptions for cardiovascular health, reportedly recommending 150–300 minutes per week of moderate aerobic activity plus strength training at least twice weekly. Commentary on Twitter highlighted the statement's emphasis that the single largest cardiovascular benefit comes from moving out of a sedentary lifestyle into any regular activity, rather than from optimizing among already-active regimens.

Discussion: 1 tweets from 1 authors · @a_garciahermoso

Medicine

Network Meta-Analysis Compares Antibiotic Regimens for Intra-Abdominal Infections

This systematic review and network meta-analysis compares different antibiotic prophylaxis and empirical therapy regimens used for intra-abdominal infections, aiming to rank their relative efficacy and safety; no abstract was available, so this description is based on the title alone. Twitter discussion around the paper consisted mainly of a single share highlighting the study with no substantive critique or debate attached.

Discussion: 1 tweets from 1 authors · @AliSMV7

Computer Science

Study Finds a Width-Driven Phase Transition in Neural Network Superposition

Using Anthropic's toy model of superposition, the authors show that increasing model width drives a continuous phase transition: below a critical width, only a subset of features get meaningful representation vectors while the rest vanish (partial-representation phase), and above it every feature is represented (full-representation phase). They derive via a 'partial random projection' approximation that this critical width scales linearly with the number of active features (up to log corrections), and show the loss scaling law changes qualitatively across the transition—linear in feature count below threshold, quadratic and inversely decaying with width above it. Non-uniform feature firing probabilities delay the transition and reduce loss, since frequently occurring features claim more representational space.

Discussion: 2 tweets from 1 authors · @ZimingLiu11, @ZimingLiu11

Computer Science

GRAM Turns Recursive Reasoning Models Into Probabilistic Multi-Trajectory Generators

The paper introduces Generative Recursive reAsoning Models (GRAM), which extends Recursive Reasoning Models—a deterministic alternative to autoregressive token generation that iteratively refines a latent state—into a stochastic framework. Instead of following one fixed latent trajectory to a single answer, GRAM samples multiple latent trajectories via amortized variational inference, letting the model explore different hypotheses and scale inference-time compute through both recursive depth and parallel sampling. The authors report it as a latent-variable generative model that supports both conditional reasoning (p(y|x)) and unconditional generation (p(x)), and show gains over deterministic recurrent/recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks. The tweet thread is a brief announcement that the paper was accepted to NeurIPS 2026, with code and a project page released; no technical debate appeared in the discussion shared.

Discussion: 1 tweets from 1 authors · @MingyuJo_

Computer Science

AdamW Weight Decay Should Scale With Model and Dataset Size, Study Finds

The paper reinterprets weights learned by AdamW as an exponential moving average (EMA) of recent gradient updates, where the key hyperparameter is the EMA timescale (effectively the number of recent iterations being averaged). The authors find this optimal timescale, measured in epochs, stays roughly constant across model and dataset sizes, which implies weight decay should decrease as dataset size grows and increase as model size grows (under muP learning-rate scaling). They validate these scaling rules on ResNet-18, Vision Transformers (CIFAR-10, ImageNet), and NanoGPT pre-training, also showing muP's learning-rate scaling breaks down late in training unless weight decay is adjusted accordingly. On Twitter, the main point of discussion was precedence and convergence of ideas: commenters noted that this EMA/timescale framing of weight decay times learning rate had been independently proposed earlier, and that a related paper extended the same perspective to LLM pre-training scaling, with the discussion mostly tracing the lineage of the idea across these works rather than raising substantive criticism.

Discussion: 1 tweets from 1 authors · @xidulu

Biology

Diffusion Models Infer Phylogenetic Trees Under Realistic Evolutionary Assumptions

The paper introduces DeepPhylo, a family of conditional generative models that build phylogenetic trees top-down, middle-out, or bottom-up using a discrete diffusion framework, trained via simulation-based inference on data from complex evolutionary models. Because likelihood-based methods like MCMC only work for simple, often misspecified models (e.g., assuming constant mutation rates, no indels), the authors argue this approach lets researchers infer trees under richer, more realistic processes. On datasets up to 65 taxa, the models reportedly recover MCMC posteriors more accurately than prior simulation-based methods like Phyloformer 2, and outperform the popular but misspecified conventional tool IQ-Tree on realistic data while matching it on simpler cases. The lead author's tweet thread (cross-posted with coauthors) frames the work as addressing a long-standing gap between flexible evolutionary modeling and tractable inference.

Discussion: 1 tweets from 1 authors · @AlanNawzadAmin

Computer Science

GLM Team Proposes Single-Rollout Async RL to Stabilize Agentic LLM Training

The paper introduces Single-rollout Asynchronous Optimization (SAO), which replaces GRPO's group-wise sampling with one rollout per prompt to reduce off-policy drift in asynchronous RL for LLMs, and adds value-model training tweaks plus a strict double-sided token-level clipping scheme for stability. The authors report stable training for a thousand steps and improvements over GRPO variants on agentic coding and reasoning benchmarks (SWE-Bench Verified, BeyondAIME, IMOAnswerBench), and note SAO was used in training the GLM-5.2 model. They also claim single-rollout RL helps in simulated online-learning settings where the environment keeps changing. On Twitter, the paper was flagged as part of a broader wave of work tackling PPO/GRPO stability in LLM RL, grouped alongside other critic-training and asynchronous-optimization approaches (e.g., Best Practice Critic Optimization, EasyPPO, TEMPO, AC2) rather than discussed in depth on its own merits.

Discussion: 1 tweets from 1 authors · @lu__jasper

Physics

ARES: A New Plane-Wave DFT Code Written Entirely in Rust

ARES is a plane-wave density-functional-theory code built as a 22-crate Rust workspace, supporting norm-conserving and PAW pseudopotentials, LDA/PBE functionals, k-point sampling, forces, stress, relaxation, and MPI parallelism. Its central feature is systematic validation: every numerical component is checked against Quantum ESPRESSO using regenerated reference fixtures, achieving band-energy agreement below 0.4 meV median and pseudopotential delta-factor accuracy of 0.696 meV/atom across 65 elements, though it currently lacks spin-polarized/spin-orbit support and GPU acceleration. Twitter commentary noted this is a rare DFT implementation in Rust, departing from the traditional Fortran/C lineage of such codes, with one widely shared summary emphasizing that Rust's compile-time safety checks—rather than raw speed—are the main selling point for avoiding parallel-computing bugs. Some speculated whether Rust could gain broader traction in computational materials science as a result.

Discussion: 1 tweets from 1 authors · @yoko_materialDX