Daily AI Briefing — September 21, 2026

AI SAFETY & ALIGNMENT

ExpBoN — Exponential-Noise Best-of-$n$ — introduces a soft BoN variant with exponentially fast convergence in total variation, expected reward, and both directions of KL divergence, providing the first tractable finite-$n$ decomposition for a noise-based BoN mechanism and a theoretical foundation for test-time alignment with finer-grained control over the reward–distribution-shift trade-off. Standard Best-of-$n$ (BoN) sampling draws $n$ candidate completions and selects the one with the highest reward-model score — simple and effective, but hard maximisation offers only coarse control. Soft BoN (Verdun et al., 2025) provides smoother control via a Boltzmann-like scheme and converges to the optimal KL-regularised distribution. ExpBoN takes a different route, using the exponential-noise report-noisy-max mechanism. Its key theoretical result is an exact decomposition that holds for any finite $n$, yielding exponential convergence rates not just in expected reward but also in total variation distance and both forward and reverse KL divergences. The authors provide comprehensive convergence and regret analyses and position ExpBoN as a bridge between computationally cheap inference-time methods and the stronger guarantees of provably convergent alignment. For practitioners, the significance is that ExpBoN may allow alignment to be adjusted at inference time with predictable, quantifiable trade-offs — a step toward engineering-grade control over model behaviour without retraining. [arXiv:2609.21899](https://arxiv.org/abs/2609.21899)

AI EVALUATION

EnterpriseVal, accepted at MLSys 2027, frames the disconnect between frontier model capability and enterprise GenAI outcomes as fundamentally a measurement problem — and proposes a complete evaluation system with formal use-case specifications, a six-family metric catalogue, calibrated LLM-as-judge scoring via prediction-powered inference, and a two-tier executable gate algorithm that maps metric vectors to REJECT/CONDITIONAL/SCALE decisions. The paper’s starting observation is striking: on GDPval, frontier models in late 2025 produced deliverables rated as good as or better than experienced professionals in nearly half of blinded comparisons, yet an MIT study found roughly 95% of ~300 enterprise AI deployments showed no measurable P&L impact, and Gartner projects over 40% of agentic AI projects will be cancelled by 2027. EnterpriseVal argues that public capability benchmarks answer “what can the model do?” while deployment decisions require “is this workflow fit, reliable, safe, and worth scaling — here, on our data, under our controls?” The system comprises (i) a formal object for the use case and frozen socio-technical configuration (model, prompts, retrieval, tools, guardrails, human oversight — all versioned), with an autonomy×consequence matrix determining evaluation intensity; (ii) metrics spanning fidelity, utility, efficiency, reliability (pass@k), assurance, and oversight; (iii) a grading protocol that scales blinded expert judgement with calibrated LLM-as-judge through prediction-powered inference; (iv) a two-tier gate algorithm that maps metric vectors with confidence bounds to clear decisions; and (v) a value-and-risk model in which the reviewer catch rate is an explicit measured parameter. A pilot across three workflows in a global bank reports human-graded citation precision of 88% and a hallucination rate of 1.6% for the best model on credit-memo drafting, and analyst refinement effort falling from an estimated 27.4 to 2.9 hours per document on procedure transformation. For anyone building or governing enterprise AI systems, EnterpriseVal formalises what a defensible deployment evaluation actually looks like. [arXiv:2609.21841](https://arxiv.org/abs/2609.21841)

RecreationWorld introduces a five-platform evaluation framework for hybrid computer-use agents — agents that autonomously decide when to explore a graphical interface, write code, and visually verify their own artifacts — addressing the fact that real digital work interleaves GUI interaction and software development rather than stacking them end-to-end. Current computer-use agents have advanced along separate lines: one line focuses on graphical interaction (clicking, scrolling, typing in browser or OS interfaces), the other on code and command-line development. RecreationWorld argues that meaningful digital work requires both, dynamically interleaved by the agent itself. The framework spans Ubuntu, macOS, Windows, Android, and Web, providing a unified harness with native GUI control and coding tools. The core task is “recreation”: given a running reference application, the agent must discover its behaviour and build a faithful implementation with no prescribed workflow. The reference serves as an oracle for hidden behavioural tests, providing execution-grounded rewards — a significant improvement over static evaluation that relies on text-based descriptions or predetermined success criteria. By scaling trajectory generation with high-quality open-source applications and providing reproducible environments across all five platforms, RecreationWorld offers infrastructure for systematically benchmarking hybrid CUAs — exactly the kind of evaluation environment needed as agentic systems begin to combine software engineering and interface manipulation in production. [arXiv:2609.22000](https://arxiv.org/abs/2609.22000)

GLOBAL & GEOPOLITICAL AI

The US is actively considering expanding its technology blockade against China to include cloud computing services — a move that industry analysts and SCMP reporting frame as potentially the most severe choke point for Chinese AI development since the chip export bans, because cloud access provides the only remaining legal route to frontier-scale compute for Chinese firms cut off from advanced semiconductors. Chinese AI developers currently cannot legally buy the world’s most powerful AI chips under US export controls, yet the country’s tech giants and start-ups have managed to keep pace with global frontier development in part through cloud services — which provide access to the same advanced hardware hosted remotely, often through third-party providers. If Washington restricts cloud computing access, Chinese firms would lose that workaround, facing what the article describes as “the most severe choke point since the chip bans.” The timing is consequential: cloud-based training and inference have become critical infrastructure for AI development, and a cloud blockade would target not just training runs but the deployment of inference workloads that underpin commercial AI products. The report underscores that the US-China AI competition is shifting from hardware supply chains to the broader compute-access layer — a front with potentially deeper implications than the chip controls alone, since cloud restrictions would also affect inference deployment and ongoing model serving, not just initial training. SCMP