A 17-year-old high school student has successfully turned common algae into a biological altimeter that reached the stratosphere. Andrew's StratoSpore project combines spectral sensing with machine learning to measure altitude through algae fluorescence???a world first that could transform how we mo...
Read Full Article →
A new arXiv paper argues that generated contact sound can supply the force information that video-only robot learning lacks. If the pipeline holds, contact-rich manipulation data becomes a generation problem rather than a hardware problem.
ScienceBuddy couples harness evolution with model reinforcement learning, turning researcher feedback into tasks and evaluation rubrics automatically. The operational consequence is that evaluation stops being a separate engineering function and becomes a byproduct of daily scientific work.
Bellman Policy Optimization (BPO) reformulates Policy Mirror Descent using the Bellman equations to avoid estimating intermediate state values, producing a critic-free RLVR method with a proven unique optimal solution. This analysis argues the theory is real but the empirical case is unproven, and names the labs and benchmarks that will decide whether BPO matters.
Chain-of-Self-Questioning (CoSQ) turns abstention into an explicit pre-commitment information check, evaluated on TruthfulQA's 817-item multiple-choice validation set across eleven model families. This analysis separates what the paper actually shows from what the truncated abstract leaves unproven.
A September 2026 arXiv paper describes plan injection, an attack that hides adversarial intent inside benign-sounding reasoning planted in an actor's context. The finding undercuts the central premise of chain-of-thought monitoring and forces a rethink of how safety cases are evidenced.
Duplex Cue introduces an evaluation for in-turn adaptation in full-duplex voice agents, separating listener intent (backchannel, collaboration, interruption) from speaker behavior. The framework exposes that current binary evaluations cannot measure whether an agent incorporates a mid-utterance contribution while keeping the floor.
The arXiv paper proposes a causal inference framework that combines repeated generated answers, question counts, cross-system share of use, and notice probabilities to estimate GEO and GEM effects. This analysis argues the framework's real value is not the model — it is the notice-probability data layer it forces into existence.
MindTopo is a new arXiv benchmark testing whether foundation models can reason about topological invariants — properties preserved under continuous deformation — rather than metric relations like distance and angle. The paper argues these relations are foundational to spatial understanding and largely absent from current model evaluations.
Real-SWE, published by Specific on September 12, 2026, benchmarks AI coding models against private enterprise codebases instead of public repositories. This analysis explains what changed for buyers, which vendors are exposed, and how to run a defensible evaluation before signing a coding-agent contract.
AWS introduced the Agent Evaluation Metric (AEM), a turn-level framework for multi-turn agent evaluation that separates the originating failure from inherited failures. This analysis examines what AEM actually measures, what it leaves unmeasured, and who wins if it becomes the default.
A new arXiv paper replaces answer-level self-voting with output-free probe execution, making test-time RL viable for code. This article breaks down who should adopt it, what it costs, and where it breaks.
IBIB argues that usable capability is a joint property of weights, serving route, precision, output contract, and harness, and that all 18 audited benchmarks miss this by scoring advertised model identifiers. The protocol adds a gold-blind capability-binding preflight and a reliability-inclusive first-pass scoring rule.
Append the next batch without leaving this page.
