Replication shows less-skilled, well‑calibrated radiologists gain the most from AI assistance

: Daniel Martin

replication, medium evidence suggestive

Applying a prior framework to 11,420 radiologist–case observations, the study finds that lower baseline ability and better belief calibration are associated with larger accuracy gains from assistance, guiding targeted rollout and training in clinical settings.

Deployed DoorDash RL shifts dispatch weights, cuts courier time costs and raises batching without harming delivery quality

: Haochen Wu, Yi Hou, Shiguang Xie

field experiment, medium evidence suggestive

A store-level policy trained offline to select a simple multiplier that reweights an existing dispatch optimizer increased batching and reduced courier time while customer metrics held steady in a production switchback test, illustrating a safe path to marketplace efficiency in that context.

Human‑in‑the‑loop workflow cuts AI‑assisted research failure rate from 72% to 16%

: Chen Zhu, Xiaolu Wang, Weilong Zhang

RCT, high evidence established

In a preregistered factorial trial, keeping models out of raw data execution, enforcing deterministic computation, and inserting three human decision gates reduces critical failures by 56 percentage points versus an unconstrained multi-agent baseline.

Geometric conditions determine when LLLM non‑local proposals actually aid discovery : Li Xia, Baoxun Wang (theoretical, medium evidence)

A formal “search compression” theory characterizes when non-local jumps from LLMs may help discovery, pointing leaders to domains where agentic exploration is more likely to pay off.

Nearly half of AI‑authored PR fixes are rejected; failure modes cluster into 14 reasons : Mahmoud Abujadallah, Ali Arabat, Mohammed Sayagh (descriptive, high quality)

Analyzing 306 non-merged agentic pull requests, the paper documents a 46.41% rejection rate and maps common failure modes, quantifying review burden and reliability gaps in agentic coding.

Provider‑hosted precomputed KV caches reproduce prefill exactly and cut prefill compute by an order of magnitude : Luoyuan Zhang (descriptive, high quality)

Measurements on tested models replicate prefill token-exactly and indicate 9–50x prefill compute reductions, which could enable content delivery network (CDN)-like economics for repeated reads and shift costs to storage and egress.

Graduates from Scheduled Castes and Tribes have substantially lower exposure to higher‑paying AI‑exposed jobs : Kaibalyapati Mishra (correlational, medium evidence)

Using India’s PLFS 2025, SC/ST graduates show 0.24–0.37 standard deviations (SD) lower occupational AI exposure, signaling a risk of widened caste earnings gaps without targeted access and training.

Agentic AI adoption raises code size but not total architectural smells, lowering smell density via size effects : Oliver Aleksander Larsen, Mahyar T. Moghaddam (quasi-experimental, medium evidence)

A staggered difference-in-differences on 151 Java repos associates adoption with larger codebases and no change in total architectural smells, so smell density falls by dilution rather than better design.

Frontier agents on WorkBench sharply raise task completion and cut harmful actions, with capability and safety improving together : Olly Styles (descriptive, high quality)

Two years of benchmarking reports task completion rising from 43% to 89% and unintended harmful actions falling from 26% to 2.5%, suggesting, in these tests, simultaneous gains in capability and safety on evaluated tasks.

Proprietary curated evidence dramatically boosts AI valuation agents' coverage and decision utility : Yinan Wang (quasi-experimental, medium evidence)

A stratified ablation indicates curated proprietary corpora contribute more to performance on drug-asset valuation than public tools or scaffolding, emphasizing evidence access over orchestration.

Agentic multi‑agent architecture autonomously resolves over 90% of common network incidents in production with layered safety : Arun Malik (descriptive, high quality)

A production system reportedly resolves most routine incidents via hierarchical agents, runbooks, progressive autonomy, and closed-loop checks, showing feasible autonomy when tightly bounded.

Including LLM prognostic predictions as covariates reduces variance and has a 'do no harm' guarantee : David Arbour, Eli Ben-Michael, Avi Feller, Apoorva Lal, Lo-Hua Yuan (theoretical, medium evidence)

Theory and simulations show LLM-generated prognostic scores can improve precision in randomized trials while defaulting to unadjusted estimates when uninformative.

Large scientist‑in‑the‑loop study finds LLMs produce plausible but conservative ideas and rarely propose null hypotheses : Honglin Bao, Siyang Wu, Xiao Liu, Sida Li, Shiyun Cao, James A. Evans (descriptive, high quality)

With 6,749 scientists and 25,139 ratings, models cluster around safe ideas and avoid negation, especially in pluralistic fields; human-trained reward models partially improve alignment.

Autonomous 'Computer' agent automates far more work, cuts completion time and user dissatisfaction versus search : Jeremy Yang, Kate Zyskowski, Noah Yonack, Jerry Ma (quasi-experimental, medium evidence)

Matched sessions on Perplexity are associated with much higher autonomous work time, faster completions, and roughly half the dissatisfaction, suggesting sizable user efficiency gains from autonomy in that product.

GenAI adoption improves liquidity and stabilizes volatility where governance is strong, but can amplify informational imbalances in weak markets : Sepideh Khalafi, Ali Salari (quasi-experimental, medium evidence)

Cross-market evidence associates GenAI with better liquidity and market quality in well-governed settings, with potential information asymmetry in weaker ones.

DI‑style assistants increase information elaboration but raise cognitive load; strategy training mitigates costs : Shuqing Liu, Kerr Manson, Thomas Ware, Dennis Galletta, Narayan Ramasubbu (RCT, medium evidence)

Randomized studies find dialectical inquiry bots spur deeper analysis but increase cognitive load; user strategy training recovers net benefits.

Open‑source contributor policies for AI agents are fragmented and misaligned with emerging AI governance instruments : Jassem Manita, Aziz Amari (descriptive, high quality)

A comparative audit shows uneven policies on disclosure, liability, and oversight for machine contributors, offering a taxonomy and maturity score to align with governance frameworks.

Environment engineering (permissions, artifacts, budgets, human gates) unlocks stronger autonomous discovery with low cost : Amy Xin, Jiening Siow, Junjie Wang, Zijun Yao, Fanjin Zhang, Jian Song, Lei Hou, Juanzi Li (descriptive, high quality)

Carefully engineered agent environments with permissions, artifact stores, budgets, and human gates report better benchmark discovery at low API cost, pointing to scaffolding over raw model swaps.

Foundation time‑series models improve zero‑shot forecasting but decision utility depends on calibrated quantiles : Xiaobin Zhang, Lefei Shen, Mouxiang Chen, Zhuo Li, Hongkai Li, Han Fu, Jianling Sun, Xiaoxue Ren, Chenghao Liu (descriptive, high quality)

In these tests, strong forecasting did not guarantee better cloud consolidation; calibrated quantiles were needed to balance utilization and reliability.

Instruction files have mixed effects on agentic PR outcomes; quality and length of instructions matter : Ali Arabat, Mohammed Sayagh (quasi-experimental, medium evidence)

Across 15,549 agentic PRs, instruction files improve merge rates in about 28% of projects and worsen them in about 26%, with effects driven by instruction structure and quality.

Standardized framework shows 19 LLMs achieve autonomous penetration success rates from ~11% to ~69% in simulated targets : Jiaqi Luo, Jiarun Dai, Zhile Chen, Jia Xu, Weibing Wang, Yawen Duan, Brian Tse, Geng Hong, Xudong Pan, Yuan Zhang, Min Yang (descriptive, high quality)

Wide variation in simulated autonomous penetration success highlights capability heterogeneity and the case for standardized red-team testing.

Systematic review: ML boosts prediction‑based productivity; DL drives automation in high‑skill industries; generative AI restructures knowledge work : Lavanya Singla, Sushil Laddhu, Surbhi Tiwari, Shweta Goel (review, medium evidence)

A synthesis across studies links ML to predictive gains, deep learning to automation and capital deepening, and generative AI to shifts in cognitive work and employment structures.

Automatically generated multi‑agent systems are costlier and often underperform single‑agent chain‑of‑thought baselines : Prathyusha Jwalapuram, Hehai Lin, Chuyuan Li, Fangkai Jiao, Sudong Wang, Yifei Ming, Zixuan Ke, Chengwei Qin, Giuseppe Carenini, Shafiq Joty (descriptive, high quality)

Multi-agent systems produced automatically can underperform well-tuned single-agent chain-of-thought with self-consistency and cost up to 10x more; expert-designed MAS excel only on tasks suited to parallelization.

Special issue synthesis: AI reshapes Chinese management but success depends on culture, leadership and governance : Tachia Chin, Chien-Liang Lin, Chris Rowley, Lei Huang (review, medium evidence)

Organizational culture, leadership, and governance are associated with AI payoffs in Chinese management contexts, constraining over-delegation to algorithms.

Critical synthesis argues AI adoption often reproduces global power asymmetries and governance gaps in the Global South : Itoro Abraham (review, medium evidence)

A review of 50 articles links AI infrastructures and platforms to accountability gaps, epistemic injustice, and precarious labor in postcolonial settings, calling for sovereignty-aware governance.

Coding agents match or exceed human methodological diversity but verdicts remain fragile to prompt framing : Meysam Alizadeh, Fabrizio Gilardi, Mohsen Mosleh, Enkelejda Kasneci (quasi-experimental, medium evidence)

Coding agents produce effect estimates aligned with consensus across repeated runs yet flip judgments with prompt changes, emphasizing interpretive fragility.

Structured LLM pipeline matches professional mediators on short‑term prep outcomes and cuts preference‑inference error : Jamie Bergen, Sarit Kraus (RCT, medium evidence)

Two controlled studies show a modular pre-mediation pipeline performs on par with professionals for preparation and reduces preference-inference error by 36%, after tuning to curb over-affirmation.

Benchmark finds LLM agents can outperform median expert humans on several bio tasks and executes wet‑lab validation for one model : Andrew Bo Liu, Samira Nedungadi, Bryce Cai, Alex Kleinman, Harmon Bhasin, Seth Donoughe (descriptive, high quality)

Agents exceed median expert performance on three bio tasks and generate executable lab protocols in one case, raising biosecurity concerns.

Lightweight grounding adapter lets coding agents produce expert‑quality GEOS simulation decks in minutes : Matthew Ho, Brian Liu, Jixuan Chen, Audrey Wang, Lianhui Qin (descriptive, high quality)

Providing simulator contracts enables agents to generate full GEOS decks in about five minutes at expert quality, indicating the value of domain-grounding adapters.

Human-in-the-loop architectures and reliable productivity

Across controlled and production settings, constraining where models can act is associated with reliability and gains. When LLMs reason but do not execute code, and when humans hold decision gates, error rates fall without obvious speed loss in the studies reviewed. In operations, scoped autonomy that adjusts an existing optimizer’s weights or encodes runbooks achieves efficiency while respecting safety boundaries. The open question is how far to loosen those constraints: bespoke, engineered multi-agent systems look promising in narrow domains, while off-the-shelf multi-agent orchestration often adds cost with little benefit. Editorially, the trajectory favors bounded autonomy plus oversight as the default, with autonomy expanded only where measurement shows net utility.

Operational infrastructure and cost economics for agent workloads

The economics of agents increasingly sit in infrastructure, not just models. Provider-hosted KV caches bring order-of-magnitude prefill savings for repeated reads, effectively moving costs from compute to storage and network. End-to-end decision benchmarks suggest that even excellent forecasts do not yield better outcomes without calibrated decision thresholds in those tests, so firms must optimize for decision loss, not prediction error. Environment engineering that bakes in permissions, artifact stores, and budgets can lower failure and spend. The editorial read: serving stacks and calibration tooling are likely to drive the next wave of unit-cost reductions and latency wins.

Agent capabilities, safety, and security risks

Capability metrics have risen in tested settings: workplace agents complete more tasks with fewer harmful actions, agents generate lab-executable protocols, and domain-grounded adapters unlock expert-quality simulations. Dual-use risk is real in some areas; standardized tests report meaningful autonomous penetration success, and bio-capability benchmarks tighten the feedback loop from text to lab. Creativity remains a weak spot for unconstrained models, which cluster around plausible but conservative ideas, though structured environments with metrics and iteration can push agents into more productive search. Taken together, capability and safety appear to be improving in tandem on evaluated tasks in these benchmarks, but governance and red-line testing need to keep pace.

Inequality, labor market exposure and governance

Institutional context conditions AI’s benefits: financial markets with strong governance are associated with liquidity and stability gains, while weaker markets risk information asymmetry. Within countries, exposure to AI-intensive work is uneven, India’s caste-based gaps signal who may be left out of wage gains absent targeted training and access. Management syntheses stress culture and leadership alongside governance, echoing findings that policies for machine contributors in open source remain fragmented. Editorially, this looks less like a generic “AI inequality” story and more like a tractable governance and access problem that policy can shape.

Oversight architecture reduces failure in AI-assisted research established

A preregistered RCT finds human gates and deterministic computation reduce critical failures from 72% to 16%.

Implication: Funders and universities should standardize gated, non-executing AI workflows for data analysis.

Offline RL via weight multipliers appears to improve marketplace efficiency without degrading service suggestive

A production switchback test shows a store-level policy that reweights an existing optimizer increases batching and lowers courier time while holding delivery quality steady.

Implication: Platforms should prefer control-layer learning over end-to-end replacement to capture gains safely.

Provider-side KV caches cut prefill compute by 9–50x with token-exact reuse descriptive

Measurements show hosted key-value caches reproduce prefill exactly and indicate order-of-magnitude cost savings for repeated reads.

Implication: Move repeated-document workloads to provider-hosted caching and adjust pricing to storage/egress economics.

Multi-agent orchestration is not a free lunch descriptive

Automatically generated multi-agent systems often underperform strong single-agent chain-of-thought and cost up to 10x more, except when tasks favor parallelization and expert design.

Implication: Default to single-agent baselines, add agents only with measured parallelizable bottlenecks.

Governance quality is associated with AI’s market-level benefits suggestive

Cross-market analyses associate GenAI adoption with better liquidity and stability in well-governed settings and potential imbalances in weak ones.

Implication: Regulators should tie AI-enabled trading permissions to disclosure, audit, and market-quality safeguards.

Provider-hosted precomputed KV cache and reuse: Can I Buy Your KV Cache?

Describes a practical serving primitive that can replicate prefill token-exactly while cutting compute by up to 50x, which could change cost models for retrieval- and document-heavy agents.

Multiplier interface for offline multi-agent RL: Multi-Agent RL from Delayed Marketplace Feedback

Presents a control-layer approach where learning adjusts objective weights of an existing optimizer, enabling offline training under delayed, coupled rewards.

Human‑in‑the‑Loop Economic Research (HLER) workflow RCT: (Human) Attention Is (Still) All You Need

Isolates the causal effect of workflow constraints, offering a blueprint for reproducible, lower-failure AI-assisted research.