The IBM Telco churn benchmark can report 99.01% accuracy while producing misleading model comparisons and costly retention decisions. The audit finds that pre-split SMOTE, a synthetic-data technique, boosts churn F1 by 13.1 points, with about 36% of synthetic training examples interpolating test instances. Under the paper’s $50 offer and 24-month customer-value proxy, the cost-optimal threshold saves about $77,000 per 1,000 customers versus an F1-focused approach.
A new attention mechanism lets a query change not only how strongly a language model reads a token, but which parts of that token’s value it extracts, giving researchers a small, editable control surface for steering outputs. The learned 2D B-spline surfaces add 0.3% parameter overhead and cut WikiText-2 perplexity by 2 to 2.5 points on LLaMA 3.2 models. Editing one layer changed greedy output on 9 of 10 prompts, while setting coefficients to -1 created “attention walls” that redirected an explosive-device prompt toward general education.
AIM gives an autonomous research system a way to search ideas, check whether experiments tested those ideas, and spend its limited run budget where evidence is most useful. It groups ideas, ranks promising directions, dispatches parallel solvers, then audits the resulting implementations before updating the search. The page says AIM leads the two reported task-group averages, and on Flash Attention it reaches ScientistOne’s best score up to 3.1× faster.
A new audit says generative-model memorization claims can change dramatically once similarity scores are tested against a real null distribution, which matters for anyone measuring data copying or mitigation quality. Using matched controls and false-discovery control, the paper certifies 36 of 61 MemBench images, while held-out controls appear in just 0.01% of calibration splits. Its stricter permutation test finds near-duplicate copying in none of the twenty-four generators examined.
Google’s new Trusted Execution Environment based federated learning system lets third parties verify how training data is anonymized, while moving computation to servers to improve speed, accuracy, and device coverage. The protected code can be remotely checked, while its internal state stays confidential and its operation stays intact, subject to current TEE limits. Gboard already uses the system and reports substantially faster computation than its previous setup.
Replacing native sigmoid, tanh, and SiLU calculations with short polynomial programs can raise complete training-step throughput by up to 8.0% on GB200 systems, giving LLM builders a way to ease special-function bottlenecks without changing the model checkpoint. The method swaps each function for packed multiply-add arithmetic inside the kernel, with gains of 2.7% for dense SiLU, 2.9% for tanh-softcapped attention, and 8.0% for routed-expert SwiGLU. Across four tasks, final smoothed training-loss differences ranged from -0.107 to +0.079 near 100 billion tokens.
LLMs’ verbal confidence markers do not map to uncertainty the way human phrases such as “possible” and “likely” do, making natural-language confidence harder to interpret safely. The paper benchmarks models against a corpus of human uncertainty markers, then fits a marker-to-probability mapping directly to model outputs rather than relying on repeated sampling. That exposes systematic gaps between model confidence semantics and human expectations.
AI tools that remove the awkward work of dealing with other people may also remove the interactions that help research happen. The article calls this the “Waymo effect,” using driverless rides as the example: convenience makes solitude feel like a pure gain, while concealing what conversation and shared practice contributed. For research leaders, the risk is choosing efficient individual work by default, even when difficult ideas need people in the same room.