Daily Digest

Jiatong Li, Jiatong Li, Weida Wang, Yunqing Liu · openalex

A new benchmark (S²-Bench) and instruction-tuning dataset (OpenMolIns) shift molecule generation evaluation from one-to-one retrieval to open-ended, one-to-many creativity. Fine-tuning a Llama3.1-8B on OpenMolIns outperforms GPT-4o and Claude-3.5, suggesting domain-specific small models can surpass general-purpose giants on molecular design tasks. For Isomorphic Labs, this means smaller, cheaper models fine-tuned on curated molecule-language data could become viable for daily molecular optimization and editing workflows, reducing reliance on costly API calls while improving creative candidate generation. The benchmark's one-to-many framing more closely mirrors real drug discovery needs than existing datasets.

Ariana Peck, Joshua Hutchings, Jonathan Schwartz, Yue Yu · openalex

A Kaggle challenge on particle picking in cryo-electron tomography drew over 1,000 participants and produced ML models that beat existing state-of-the-art methods across five molecular species. The winning solutions highlight that data augmentation is critical when training data is scarce, and that subtomogram averaging is surprisingly tolerant to moderate over-picking but breaks down under severe over-picking. All tomograms and annotations are now public as a benchmark. This is relevant because reliable particle picking remains a major bottleneck for in situ structure determination, and better ML approaches directly accelerate the structural biology pipeline that underpins AI-driven drug discovery.

Canyu Chen, Jian Zhen Yu, Shan Chen, Che Liu · openalex

A new benchmark, ClinicalBench, systematically compares LLMs against traditional ML models (SVM, XGBoost) on clinical prediction tasks. Across 14 general-purpose and 8 medical LLMs, with various prompting and fine-tuning strategies, none outperformed the traditional models. This suggests LLMs currently lack the structured reasoning needed for clinical decision-making, despite excelling at text-based tasks and exams. For anyone building AI for biomedical prediction—including drug discovery—this is a grounding result: don't assume LLMs will replace established ML workflows for regression or classification without careful evaluation.

Haoyu Han, Li Ma, Yu Wang, Harry Shomer · openalex

A comprehensive benchmark reveals that GraphRAG excels over standard RAG on relational reasoning and query-based summarization, while RAG remains superior for simple factoid QA. The key insight is that hybrid strategies combining both paradigms consistently outperform either alone. This directly informs design decisions for retrieval-augmented systems in drug discovery—where knowledge graphs of molecular interactions benefit from graph-structured retrieval and text-heavy literature favors traditional RAG. Efficiency trade-offs and evaluation biases are also characterized, making this a practical guide for productionizing such systems.

Junru Zhou, Cai Zhou, Xiyuan Wang, Pan Li · openalex

Laplacian eigenvectors consistently boost GNN expressivity, but naive handling of their orthogonal-group symmetry causes instability and lost power. This work introduces learnable O(p)-invariant encoders per eigenspace and smooth interpolation between nearby eigenvalues, yielding globally expressive graph representations that are more robust to perturbation. For molecular graph modeling, that means better capture of global structural properties without sacrificing stability—especially relevant if you're building models that need to generalize across conformers or noisy 3D coordinates. The approach leverages existing equivariant layers, so it's practical to integrate into current architectures. Worth a look if you're pushing graph representations beyond local message passing in drug discovery pipelines.

Yan Liu, Chuan Zhou, Shuai Zhang, Xiaotong Wu · openalex

Crystal structure prediction gets a boost from hypergraph-based diffusion models that capture multi-way atomic interactions better than pairwise graphs. The EH-Diff model enforces permutation and translation equivariance, outperforming SOTA methods on benchmarks with just one diffusion sample. For you, this is a potential cross-pollination: hypergraph representations could improve modeling of protein-ligand complexes or polymorph screening, where higher-order interactions (e.g., solvent effects, allostery) matter. Worth watching if the approach transfers to molecular systems.

Thomas Lund Koch, Giada Ferrari, Sunita B. Sumanam, Paula Flórez Salcedo · openalex

Cone snail venom toxins evolved from neuropeptide genes through multiple independent recruitment events, driven by exon shuffling, alternative splicing, and transposable elements—not whole-genome duplication. This reveals a general mechanism for how gene innovation yields complex molecular repertoires, with direct implications for understanding how novel therapeutic candidates can arise from evolutionary tinkering. For drug discovery, it highlights a rich, underexplored source of peptide scaffolds that could be engineered for new targets.

Bo Ni, Zheyuan Liu, Yongjia Lei, Leyao Wang · openalex

A new survey maps out the key trustworthiness challenges in Retrieval-Augmented Generation (RAG) systems, organizing them across six dimensions (reliability, safety, privacy, fairness, explainability, accountability) and their interactions with retrieval and generation stages. The work highlights that improving one dimension often comes at the cost of another—for example, better privacy might reduce explainability. For anyone building or deploying LLM-based tools, especially in high-stakes domains like drug discovery, this provides a structured way to evaluate trade-offs and design more trustworthy systems. The survey also identifies downstream applications where trustworthy RAG can have significant impact, though it doesn't dive deep into any specific domain.

Guibin Zhang, Jiayang Lyu, Ran Sun, Xinlei Yu · hf_daily_papers

A new self-distillation method, LOPD, removes the hand-crafted privileged context (answers, feedback, skills) that limits current OPSD approaches. Instead, the teacher's privileged context is learned end-to-end as latent tokens retrieved from experience, giving dense token-level supervision to the student. Results show it beats RLVR and representative OPSD variants on tool use and code generation, with over 3x sample efficiency versus GRPO and Skill-SD. The key conceptual shift is making the self-improvement loop fully learnable rather than relying on designer-specified scaffolds—a step toward truly self-evolving agents. For your work, this matters as a potentially generalizable training paradigm for sequential decision-making and agentic systems, and the efficiency gains directly speak to production constraints you'd care about in ML infra.

Kai Chen, Jifeng Ding, Ning Ding, Jiaye Ge · hf_daily_papers

Mobius-v0 introduces a clean separation between knowledge storage (FFN memory) and compositional reasoning (self-attention reasoners), using hidden states as a cache to shuttle knowledge between the two during iterative inference. This design yields a 7B model that matches a standard Transformer with 37% less training data, and a 35B variant (Intern-S2-Mobius) that delivers the same downstream performance as Qwen3.5-35B at roughly 4x faster inference. For anyone building or deploying large foundation models—especially in cost-sensitive production settings or iterative scientific workflows—this architecture suggests a path to significantly cheaper training and faster serving without sacrificing quality. Worth a close look if you're tracking model efficiency trends or thinking about how to compress knowledge without losing reasoning capability.

Sen Xu, Wei Wang, Shixi Liu, Jixin Min · hf_daily_papers

Standard test-time compute scaling adds more solution samples, but this paper instead reallocates compute to a verification step that checks each reasoning claim for falsification. The insight: refuting a wrong claim is fundamentally easier than building a correct solution from scratch. By isolating decision-critical claims and searching for a single flaw, the method outperforms both pass@1 and self-consistency, often using fewer tokens. For your ML work, this is a clean, training-free technique for improving reasoning on LLM outputs—directly applicable to any system where you're sampling multiple candidate solutions (e.g., drug target predictions, molecular generation) and want to spend compute where it actually resolves ambiguity.

Xinle Deng, Yida Xue, Xiangyuan Ru, Haoming Xu · hf_daily_papers

MobileMem provides a new benchmark and framework for on-device long-term memory in AI agents, using a year-scale synthesis pipeline of mobile experiences to test temporal reasoning, knowledge updating, and implicit preference inference. It moves beyond static QA to continuous, experiential learning from heterogeneous, multimodal user data. For anyone building persistent, personalized AI systems—whether in drug discovery (e.g., learning from iterative experimental workflows) or geospatial AI (e.g., accumulating environmental patterns over time)—this framework addresses a key gap: enabling agents to remember and adapt across long horizons rather than treating each interaction in isolation.

Shu Wan, Miles Ma, Hank Zhu, Guangqi Liu · hf_daily_papers

A new paper identifies 'forecast collapse' in time-series foundation models: when predicting hourly equity returns, forecasts go flat and lose cross-sectional ranking ability, even though the same models work fine for volume. The root cause is a calibration–ranking tradeoff: optimizing squared error for per-series goals suppresses amplitude, while ranking objectives inflate it. The authors propose CalibRank, a simple loss that nearly triples cross-sectional correlation without blowing up forecast amplitude. This matters because it exposes a blind spot in standard time-series evaluation—per-series metrics can hide failures in cross-series structure that downstream decisions depend on. For an ML engineer, this is a clean demonstration of how objective design can silently break a model's practical utility, and the fix is elegant enough to apply broadly.

Zhensu Sun, Chengran Yang, Yunbo Lyu, Jieke Shi · hf_daily_papers

Second Thought reduces LLM agent inference cost by exploiting the idle window between thought and observation to fork parallel reasoning branches, cutting sequential decoding up to 43% while maintaining or improving accuracy. This training-free technique is immediately applicable to any ReAct-style agent and could lower latency/cost in production systems. For you, it's a clean, practical contribution to inference efficiency—no retraining, just a smarter scheduling trick—directly relevant to any deployment of reasoning agents.

Jingwei Li, Xinran Gu, Rui Dai, Xintong Hao · hf_daily_papers

Scaling high-quality domain data via repetition is a known pretraining trick, but this paper provides a practical scaling law: at a fixed tokens-per-parameter ratio, the optimal repetition count grows mildly with model size and strongly correlates with domain loss (not unique data volume). This means you can tune repetition on smaller proxy models to set budgets for larger ones, directly relevant to efficient pretraining for domain-specific LLMs like those used in drug discovery or geospatial AI. For Nathan, this is a concrete engineering insight that could inform how Isomorphic Labs might train domain-tuned models without wasting compute on over-repeated data, especially given the scarcity of high-quality biomedical text.

Hunar Batra, Lachin Naghashyar, Ashkan Khakzar, Philip Torr · hf_daily_papers

A new method, MMDiff, applies model diffing to multimodal LLMs by training sparse autoencoders on both base LMs and their multimodal-tuned versions, then comparing feature directions. This isolates exactly which internal features change during multimodal training, enabling causal removal or steering of those features. Results are concrete: selectively removing vision-related features degrades spatial and OCR accuracy by 12–17%, steering improves them by ~3%, and attack success rate on multimodal safety benchmarks drops by 24%. For you, this is directly relevant to interpretability and control work at Isomorphic Labs — especially if you're thinking about auditing or steering model behavior in drug discovery contexts (e.g., preventing hallucinations in protein-ligand interaction predictions). The approach is clean: sparse autoencoders aren't new, but diffing them across training checkpoints to find causal features is a well-executed twist.

Noah Amsel, Jack Zhang, Kwangjun Ahn, Ali Naeimi · hf_daily_papers

Dion3 tackles the computational overhead of the Muon optimizer—a known pain point for training large models—by orthogonalizing only a subset of momentum matrix rows per step, using efficient Gram Newton-Schulz kernels, and batching to cut communication costs. It matches Muon's loss but slashes optimizer step time by up to 6x, and is available as a drop-in replacement. For anyone training large models (like at Isomorphic Labs), this is a practical speedup without sacrificing quality, making it worth testing on your current workloads.

Fanfei Li, Jana Zeller, Manuel Prada-Corral, Thaddäus Wiedemer · hf_daily_papers

LITTLECURRICULUM is a curated 88B-token corpus limited to Grade 5-and-below material, used to train a 5B-parameter model (LITTLELEARNER) from scratch. The key result: post-training and in-context learning help the model better use knowledge within its scope, but cannot raise capabilities beyond what was in the training data—posing a hard ceiling on out-of-distribution generalization. This provides empirical evidence that simply layering fine-tuning or few-shot prompting on top of a fixed pretraining corpus cannot create emergent skills outside the training distribution. For ML engineers, this is a clean experimental design to study knowledge boundaries, with implications for data curation strategies and the reliability of capability evaluations (e.g., for drug discovery models trained on restricted domain data).

Graeme Wearden · guardian

UK house prices saw their sharpest August decline since 2018 (down 2% month-on-month), while investor behavior shows 56% of offers from landlords are at least 10% below asking — signaling a buyer's market emerging. Jamie Dimon warned the UK chancellor against higher bank taxes, citing job flight risks. For your portfolio, the housing slowdown could ease BoE rate hike pressure, potentially benefiting UK equities and bond yields, while Dimon's warning suggests UK bank stocks face headwinds from potential tax changes.

bbc_world

The strike on a Kyiv civilian market—six injured—signals Russia's continued willingness to hit urban economic life, not just front-line targets. For European security and macro stability, it reinforces that the war remains a persistent tail-risk into the winter, with knock-on effects for energy markets, NATO cohesion, and broader investor sentiment.

Mabel Banfield-Nwachi · guardian

Netanyahu publicly rejected Trump's 15-point Gaza peace plan, insisting on full Hamas disarmament before any Israeli withdrawal, while Hamas demands an end to Israeli aggression first. Ceasefire talks continue with Trump envoys meeting regional mediators, but Israeli airstrikes persist. Stalled US-Iran talks also risk disrupting Hormuz shipping, which could affect oil prices — relevant for your portfolio if macro risk materializes.

Fiona Harvey Environment editor · guardian

The UN desertification COP is gaining surprising momentum and real money ($12bn) as food security concerns spike. Saudi Arabia is an unlikely lead financier, positioning itself as a climate adaptation player while blocking emissions progress elsewhere. For you: this signals increased global spending on land restoration and drought resilience — a macro trend that could affect commodity prices, food inflation (and thus your personal portfolio), and open new investment opportunities in agtech or land-management startups.

bbc_world

Kushner's direct engagement with Hamas signals a major shift in US Middle East diplomacy, potentially reviving a two-state framework that has been dormant for years. For global markets, any credible peace track could ease regional risk premiums, though the odds of near-term breakthrough remain low.

wealth_common_sense

The bull market is finally broadening beyond the S&P 500 and Mag 7. After years of concentration risk, the AI capex wave is driving broad-based regional equity performance — diversification is working again. For your index-heavy portfolio (especially if tilted toward US large caps), this means the long-standing caveat of 'it's just the S&P 500' is fading. The fundamentals are now keeping pace with prices, making the rally more sustainable. If you're rebalancing or considering factor tilts, this reduces the urgency to overweight mega-cap tech and supports a more balanced global equity allocation.

monevator

Monevator's latest Weekend Reading is paywalled for members only. The teaser content reveals no substantive insights, only a generic 'what caught my eye this week' intro and a prompt to subscribe. Without access, the substance—likely curated investing links and commentary—remains opaque. For someone following UK-centric FIRE and index investing, this is a reminder that Monevator's best content is gated; the real value is in the curated links, not this blurb.

abnormal_returns

The latest Abnormal Returns linkfest is titled 'fostering overconfidence' and delivers exactly that: a parade of market euphoria signals—healthcare breaking out, European stocks at all-time highs, and a 'degenerate economy' meeting a capex boom—alongside institutional caution. Vanguard’s belated ETF push and the rise of issuers like Corgi reflect the passive indexing shift you already ride, while yet another data point shows active funds getting trounced. On the AI side, SpaceX’s $60B Cursor acquisition and the industrial-equipment buildout for data centers underscore the breadth of the AI capex cycle, but a podcast from Paul Kedrosky warns the AI bubble will burst. The 'chaotic' US military AI policy adds regulatory uncertainty. For your portfolio, the macro takeaway is that overconfidence is broad and cheap; your indexed approach remains appropriate, but the AI-specific froth merits watching for sector rotations that could affect both your holdings and the drug-discovery funding environment.

abnormal_returns

The market continues to climb despite a decade of shocks—real yields are rising, valuation debates are intensifying, and leveraged ETFs are now significant market movers. Private credit default rates are climbing, while a majority of stocks are outperforming year-to-date. For your portfolio, this suggests staying the course with index investing remains prudent, but keep an eye on leveraged ETF flows and private credit stress as potential volatility triggers. Housing recession signs easing could affect macro rates, which matter for ISA and ETF returns.

hacker_news

The Hacker News community is buzzing about a piece arguing that frontier AI models are being deliberately degraded—either through heavy-handed safety alignment, competitive moat strategies, or both. The insight is that the most capable models are increasingly less useful for nuanced, open-ended tasks because lab policies suppress harmless but high-entropy outputs. For someone building on top of foundation models—whether for drug discovery or mapping—this means the API you call today may be worse than the one you trained on six months ago. It also raises questions about whether the alignment tax is too high, and whether open-weight models will eventually surpass closed ones by default. Directly relevant to Nathan's work at Isomorphic Labs and his interest in AI alignment and inference efficiency.

hacker_news

The NIH is canceling grants based on politically charged keyword lists rather than scientific merit, upending billions in biomedical research overnight. This destabilizes the foundational science pipeline—basic biology, datasets, and trained talent—that AI-driven drug discovery depends on. For Isomorphic, that means fewer healthy academic collaborators, potential gaps in emerging research areas, and a risk that federally funded spinout talent dries up. Watch for secondary effects on partnerships and data access, and consider how this reshapes the competitive landscape if academic researchers migrate to industry or leave science entirely.

hacker_news

Nvidia is scaling back its financial guarantee for a $250 billion data center project with OpenAI, signaling a more cautious stance on AI infrastructure spending. This could indicate Nvidia's concerns about near-term demand, a shift in partnership dynamics, or a desire to diversify risk. For the AI ecosystem, it may slow OpenAI's compute expansion and ripple through cloud and hardware pricing. As an ML engineer, this affects the availability and cost of compute for frontier models, which indirectly influences the competitive landscape for AI-driven drug discovery.

hacker_news

Qwen 3.8 27B matches or beats larger models on benchmarks, but at inference time it tends to "overthink" — emitting lengthy chain-of-thought even for trivial prompts. That's not just an oddity; it directly impacts token spend and latency, especially at scale. The lesson for ML engineers: new reasoning-capable models need careful configuration of reasoning effort, temperature, or max tokens before they're production-ready. Defaulting to full reasoning is wasteful for high-throughput or real-time use cases. This is a familiar pattern with reasoning models, but Qwen's popularity means the default behavior will shape many deployments. It also underscores the value of dynamic adaptive thinking — letting the model decide when to think briefly. For work on inference optimization and platform engineering, this is a concrete reminder that capability and efficiency are now distinct axes.

venturebeat

A cascade architecture for RAG systems slashes inference costs by ~6x while improving auditability: route all cases through a deterministic stage first (exact matches, rule-based logic), then a retrieval-only stage that pulls evidence for ambiguous cases, and only then—for the 10–15% residue—call an LLM. This design ensures full explainability on the majority of decisions, eliminates LLM inconsistency on easy cases, and cuts cost by keeping the expensive model out of the fast path. For any ML engineer pushing RAG into production, especially in regulated settings, this pattern is the pragmatic alternative to the naive “LLM front line” approach.

hacker_news

Cloudflare quietly injected a JavaScript analytics snippet into a pure HTML site upon nameserver switch, requiring manual opt-out via a non-obvious dashboard path. This illustrates a growing pattern where infrastructure providers lever control plane access to enable revenue-adjacent features (analytics, observability) by default, eroding the implicit trust in neutral DNS/platform layers. For someone managing production systems and evaluating infrastructure dependencies—whether at Lyft or now Isomorphic Labs—this is a concrete reminder to audit post-migration defaults thoroughly, especially with providers like Cloudflare that span DNS, CDN, compute, and observability. The bait-and-switch on traffic integrity is the kind of subtle platform lock-in that can complicate debugging and data governance in ML/data pipelines, and reflects a broader tension between convenience and user sovereignty in platform engineering.