2026-09-17

TL;DR - Paper2Agent converts research papers and their associated code and data into interactive, tool-using AI agents. This could make published work easier to reproduce, query and extend through agent-driven scientific collaboration.

• Uses the Model Context Protocol to expose research artifacts as callable tools.
• Incorporates manuscripts, code and data into the agent’s context.
• Aims to reproduce original results and answer new scientific questions.
• Supports collaboration between paper-derived agents to generate novel insights.

TL;DR - ScienceIDE converts scientific code repositories into executable, verifiable environments for training and evaluating AI agents. Models trained on verified scientific interaction trajectories improved on held-out code repair and selected broader benchmarks, suggesting positive transfer.

• Expert-defined cases and acceptance criteria guide repository transformation, task generation, execution, and scientific verification.
• The environments support supervised fine-tuning, reinforcement learning, and evaluation through verified agent trajectories.
• The authors trained PhAI-IDE models at 72B, 9B, and 4B parameter scales.
• Reported gains span held-out scientific-code repair and selected general-purpose code, reasoning, and knowledge benchmarks.

TL;DR - Paper2Agent is an automated framework that converts static scientific papers into interactive AI agents capable of answering questions and applying published methods to new data. It could make research more reproducible, reusable, and extensible.

• Each agent acts as a “virtual corresponding author” for its source paper.
• Agents can explain the paper and operationalize its methods on new datasets.
• Multiple paper agents can collaborate, enabling research workflows that combine knowledge and methods across publications.
• The framework shifts scientific literature from passive documents toward executable, interactive resources.

TL;DR - ASLEval is an authorization-aware framework for measuring privacy exposure across all visible exits in multi-step LLM agent sessions. It shows that evaluations focused on a single expected output can substantially underestimate leakage.

• Expected-outlet-only evaluation missed 46.9% of exposure captured by the union of visible exits.
• Attacker self-reports exhibited both omissions and high false-discovery rates, making them unreliable privacy proxies.
• Schema-aligned internal evidence generally appeared before visible exposure at the request or probe level, helping diagnose leakage paths.
• Restricting model-visible tool returns altered exposure pathways but could also eliminate successful completion of legitimate tasks, highlighting a privacy–utility tradeoff.

TL;DR - Paper2Agent is an AI system that turns research papers into interactive agents capable of answering complex questions and supporting collaboration. Its authors say it could help researchers reproduce published work and navigate unfamiliar fields.

• The system converts papers into agents rather than treating them only as static documents.
• It is designed to answer complex queries about the represented research.
• Its proposed benefits include easier paper reproduction and faster understanding of unfamiliar work.
• The provided description does not specify the system’s architecture, evaluation methods, or measured results.

TL;DR - Andromeda 2 is an evidence-grounded agentic system that autonomously designs and tests drug formulations using computational tools and a miniaturized laboratory. For paclitaxel delivery, it substantially outperformed probabilistic optimization and design-of-experiments baselines under a matched budget.

• Achieved a 50% high-performance formulation hit rate, versus 17% for Andromeda 1 and 2% for wet-lab DoE.
• Identified 12 formulations satisfying all four target product profile objectives, compared with 6 and 0 for the baselines.
• Its median paclitaxel AUC was 70.1 mg·min/mL, versus 12.0 and 3.5 mg·min/mL.
• Structured in-house experimental evidence raised mean AUC by 34% in a controlled ablation, highlighting the value of evidence-grounded agentic reasoning.

TL;DR - Nature reports PreGame, a machine-learning algorithm that identifies broadly tumour-reactive γδ T cells in multiple myeloma from single-cell CITE-seq data. Expansion of this T-cell population may provide a biomarker of therapeutic response.

• PreGame analyzes combined single-cell transcriptomic and surface-protein measurements.
• The method focuses on γδ T-cell receptors with broad tumour reactivity.
• Changes in the identified cell population could help monitor treatment response in multiple myeloma.

TL;DR - This Nature study develops “xenocortication,” using human-derived neural organoids to introduce human neurons into developing mouse cortical circuits. The approach enables circuit- and behaviour-level investigation of human neurodevelopment in vivo.

• Human-derived organoids provide the neurons used for xenocortication in mice.
• The model connects cellular neurodevelopment with circuit- and behavioural analysis.
• The supplied abstract does not specify experimental outcomes, disease applications, or quantitative results.

TL;DR - A Nature study uses a rapid autopsy programme to investigate the molecular and genetic evolution of lethal metastatic bladder cancer. The resource may clarify how aggressive histological subtypes become heterogeneous and develop treatment resistance.

• Examines metastatic bladder cancer across its temporal evolution.
• Focuses on aggressive histological subtypes and tumour heterogeneity.
• Supports detailed molecular and genetic analyses of therapy resistance.
• The provided abstract does not report specific methods or findings.

TL;DR - A Nature paper introduces xvr, a rapid patient-specific neural-network approach for rigidly registering 2D X-rays with 3D volumes. It aims to make this capability available across anatomical regions to broader clinical and research communities.

• Performs patient-specific 2D/3D X-ray-to-volume registration.
• Targets rigid alignment across multiple anatomical regions.
• Emphasizes rapid operation and broad accessibility.
• The provided excerpt does not include quantitative results or methodological details.

TL;DR - Delphy is a scalable Bayesian phylogenetics method for analyzing expanding viral outbreaks in near real time. It aims to give public health organizations state-of-the-art analysis of their own outbreak data with minimal operational friction.

• Designed to scale as viral outbreak datasets grow.
• Supports near-real-time Bayesian phylogenetic analysis.
• Intended to help public health bodies analyze local data and respond rapidly.
• The provided excerpt does not include benchmarks, implementation details, or specific accuracy results.

TL;DR - A Nature study identifies mitochondrial fission factor (MFF) as a selective promoter of ferroptosis and introduces MFF–SPARK, a sensor for studying and screening this cell-death pathway. The platform could accelerate discovery of ferroptosis-sensitizing compounds and upstream MFF regulators.

• MFF is a mitochondrial–peroxisomal fission adaptor with a newly identified pro-ferroptotic role.
• MFF–SPARK enables high-content screening of compounds that sensitize cells to ferroptosis.
• The sensor can also help identify upstream biological regulators of MFF.

TL;DR - This paper identifies “Value Flattening,” a PPO failure mode where critics produce overly uniform predictions despite sharp changes in Monte Carlo state values. It proposes sparse critic supervision that improves learned policies while using only three supervised states per response.

• Value Flattening appears in both LLM reinforcement learning and controlled FrozenLake experiments, worsening as the state space grows.
• The authors link it to an implicit variance penalty in the critic loss and redundant updates from temporally correlated states with similar gradients.
• Their SParse Proximal Policy Optimization (SP³O) method applies value loss only to a few well-separated states in each response.
• Experiments on Qwen3-Base report consistent policy improvements across model sizes and evaluation suites.

TL;DR - This paper uses Fisher-Rao information geometry to derive theoretical guarantees for the minimum proportion of fresh human data needed to prevent model collapse during recursive training on synthetic data. Its bounds remain meaningful in high-dimensional categorical distributions, unlike prior Euclidean analyses.

• Models recursively trained on synthetic data can progressively lose fidelity to the underlying real-data distribution.
• The analysis models training dynamics on the probability simplex using the Fisher-Rao metric.
• It derives quantitative contraction and invariance bounds that do not become trivial as dimensionality increases.
• The results indicate that the human-to-synthetic data ratio required for stable training differs from previous estimates.

TL;DR - This paper shows that transformer model growth, recursive depth, and boundary operators can improve pre-training scaling exponents rather than merely deliver fixed efficiency gains. A 7.4B growth architecture matches GPT-3 13B on CORE with roughly 20× less compute, with gains increasing at larger scales.

• Growing models during training, with or without shared weights, produces the largest changes in scaling exponents.
• Boundary operators that normalize and inject an earlier block also yield increasing compute-efficiency gains, though smaller than model growth.
• In data-constrained multi-epoch training, looping acts as a regularizer, and the compute-optimal loop count increases with scale.
• The results suggest that increasing usable computational depth within a fixed compute budget can improve scaling behavior.

TL;DR - Researchers combined mass photometry with single-molecule trapping to monitor individual virus-like particles assembling in real time at molecular resolution. The approach enables direct measurement of self-assembly pathways and dynamics.

• Tracks the assembly of individual virus-like particles rather than ensemble averages.
• Combines mass-sensitive measurements with single-molecule trapping.
• Enables real-time observation and quantification of intermediate assembly states.
• Provides a tool for studying the molecular mechanisms and kinetics of viral-particle self-assembly.