2026-09-02

TL;DR - This paper formulates inference-time configuration tuning on live, noisy requests as Online Hyperparameter Optimization and introduces IMABO, an infinitely many-armed bandit framework. It enables production systems, including LLM agents, to optimize choices without representative validation data.

• IMABO separates selecting previously sampled configurations via a bandit policy from proposing new configurations via an interchangeable oracle.
• Its IMOSS policy grows the active configuration set as (t^\beta) and provides an expected cumulative quantile-regret bound of (O(p_\rho^{-1/\beta} + T^{(1+\beta)/2})).
• Tree-structured Parzen Estimation, incumbent mutation, and a pretrained tabular foundation model all improve over uniform-random configuration proposals.
• IMABO achieves the lowest cumulative regret across the evaluated settings, spanning classical ML tuning and LLM-agent configuration.

TL;DR - WorldBench is a multilingual benchmark of 1,600 culturally and persona-grounded workflows for evaluating agents in sandboxed environments. Frontier models achieve only 49.2% Constrained Task Success, exposing brittleness in long-horizon tasks and preserving environment state.

• Covers seven languages and eight cultures, with tasks refined by language- and culture-specific human annotators.
• Evaluates realistic multi-step workflows through structured actions in sandboxed environments.
• Introduces Constrained Task Success (CTS), combining task completion, minimal modification, and complementary metrics via deterministic and LLM-based judging.
• All evaluated models show substantial gaps between task correctness and environment preservation.

TL;DR - This Nature article surveys research into why some infections trigger long COVID and other post-viral syndromes, and how those insights might guide treatments. The provided excerpt does not specify particular mechanisms, findings, or therapies.

• Hundreds of millions of people are reported to have long-term conditions following infection.
• Researchers are investigating the triggers that determine when an acute infection becomes a persistent illness.
• The work aims to identify effective treatments for long COVID and related post-viral conditions.

TL;DR - This is an author correction to a Nature study on signalling thresholds and negative B-cell selection in acute lymphoblastic leukaemia. The provided content does not specify what was corrected or whether the study’s conclusions changed.

• Published online in Nature on 2 September 2026.
• The correction concerns research on B-cell selection mechanisms in acute lymphoblastic leukaemia.
• No corrected data, methods, results, or conclusions are included in the provided text.
• Despite the requested AI framing, the available content does not establish an AI or machine-learning component.

TL;DR - A genome-wide experiment that mutated every DNA letter in a virus found unexpected biological effects that leading AI models could not reliably predict. The work highlights both the value of exhaustive mutation mapping and current limits of AI-based genomic prediction.

• Researchers systematically rewrote individual DNA positions across an entire viral genome.
• The mutations produced biological consequences that were sometimes surprising.
• Top AI models failed to predict these effects accurately.
• The brief provided does not specify the virus, experimental methods, models tested, or quantitative results.

TL;DR - SMELT is a compute-matched Mixture-of-Experts Transformer that reuses its middle layers to gain effective depth without increasing parameter, FLOP, or KV-cache budgets. Its scaling laws indicate 6.8–18.0% training-FLOP savings on the compute-optimal frontier, with especially strong downstream gains for code.

• SMELT loops the middle half of its layers twice while matching an unlooped baseline on per-token FLOPs, non-embedding parameters, and KV cache.
• Scaling experiments cover four model sizes up to 54B non-embedding parameters, with separate Chinchilla-style scaling laws fitted for looped and baseline architectures.
• Benefits grow with sample length and the number of in-context examples, and downstream improvements exceed what validation loss alone predicts.
• Mechanistic analysis suggests the second pass reduces attention sinks and shifts attention toward content-relevant tokens.

TL;DR - IntroConformal is a training-free conformal risk control framework that uses a vision-language model’s internal signals to provide finite-sample, distribution-free factuality guarantees. It reduces reliance on external verifiers and better handles confidently incorrect outputs.

• Derives conformity scores from layer-wise semantic stability in hidden-state representations.
• Introduces verification probability, based on the model’s self-assessment of claim factuality, as a stronger introspective score.
• Satisfies conformal risk guarantees across multiple large vision-language model architectures.
• Reduces abstention while matching or exceeding external-verifier baselines in claim-level discrimination.

TL;DR - This paper finds that cost-saving LLM cascades can hide severe reliability failures because the verifier used to route and evaluate answers has substantial blind spots. Fine-tuning students on verifier rejections may worsen or collapse performance while internal metrics remain deceptively stable.

• Verifier blind-spot rates increased from 0.12 to 0.55 as student size scaled from 0.5B to 32B, but decreased with stronger verifiers.
• A frontier verifier reduced the blind-spot rate to about 0.05 but escalated 46% of hard-MATH queries, eroding the cascade’s cost advantage.
• Corrective fine-tuning on rejected answers degraded the small student across both same-family and cross-family teachers.
• Verifier-based monitoring reported a steady 3% error even as true delivered error reached 32%, showing that in-loop metrics cannot reliably detect degradation.

TL;DR - Autonomous agents discovered eight explainable rules for rapidly assessing inorganic crystal plausibility. The rules can identify damaged or difficult-to-synthesize structures and substantially reduce costly DFT validation while preserving nearly all high-performing candidates.

• Agents generated, tested, and actively refuted two million candidate laws to derive eight Plausibility Rules for Inorganic Structures (PRIS) spanning five physicochemical mechanisms.
• Experimental structures satisfy PRIS rule sets at rates of 82–99%, compared with 6.5% satisfying Pauling’s rules 2–5 together.
• The strictest PRIS set detects 87.9% of damaged structures, versus 1.6–3.2% for distance-cutoff screening.
• In inverse design, PRIS and its synthesis score reduced the DFT queue by up to 67.3% while retaining 99.2% of candidates meeting the DFT-validated bulk-modulus target.

TL;DR - Satellite imagery showed accelerated movement of a glacier–rock mass days before its collapse in Nepal, which triggered a deadly flash flood. The observation suggests remote sensing could help identify warning signs for similar hazards.

• Pre-disaster images captured increasing movement in the unstable glacier–rock mass.
• The acceleration was visible only days before the collapse.
• The collapse caused a deadly flash flood.
• The limited content does not specify the satellite system, analysis method, or predictive reliability.

TL;DR - BenchMIRT examines what LLM benchmarks actually measure. Because only the title and source are provided, its methods and findings cannot be summarized reliably.

• The work focuses on interpreting the capabilities or constructs captured by LLM benchmarks.
• Its framing questions whether benchmark scores reflect the abilities they are commonly assumed to measure.
• No specific datasets, methodology, experiments, or results are available in the provided content.