Atomic fact-checking triples clinician trust in LLM oncology recommendations

Lisa C. Adams, Linus Marx, Erik Thiele Orberg, Keno Bressem, Sebastian Ziegelmayer, Denise Bernhardt, Markus Graf, Marcus R. Makowski, Stephanie E. Combs, Florian Matthes, Jan C. Peeken

RCT, high evidence established

A randomized trial with 356 oncology clinicians (7,476 trust ratings) finds a claim-by-claim, “atomic” verification interface increases expressed trust from 26.9% to 66.5% versus standard explainability (Cohen’s d = 0.94). For high-stakes deployments, presentation that certifies verifiability appears pivotal for safe adoption and should be evaluated for effects on actual decisions and outcomes.

Targeted AI feedback on joint attention sharply improves pair-programming debugging performance

Anahita Golrang, Kshitij Sharma

multi-study experiments, high evidence established

Using dual eye-tracking and pupillometry, high-performing dyads show higher joint mental effort and gaze alignment; reactive feedback on deviations improves collaboration, with combined (effort + gaze) feedback yielding the biggest gains. Time-series modeling indicates effort leads attention, implying process-level nudges can causally enhance team debugging beyond output suggestions.

GenAI coding assistants raise developer productivity but do not improve learning outcomes

Sebastian Maier, Moritz Gunzenhäuser, Jonas Schweisthal, Manuel Schneider, Stefan Feuerriegel

meta-analysis, medium evidence suggestive

A meta-analysis of 23 studies estimates a moderate productivity boost (Hedges’ g = 0.33) from coding assistants but no statistically significant effect on learning (g = 0.14, CI includes zero), with sizable heterogeneity across settings. Expect real, context-dependent output gains, not automatic skill development.

Water resource tax reform increases grain yields by reducing industrial pollution and spurring green innovation Yashu Qin, Shiyao Yuan, Yue Wang, Luwei Wang (quasi_experimental, medium evidence)

A staggered difference-in-differences (DID) across Chinese prefectures is associated with higher grain yields via reduced industrial wastewater and more green patenting, suggesting environmental pricing can complement productivity policy.

SignSGD provably outperforms SGD under sparse-noise, l1-stationarity geometry Hongyi Tao, Dingzhi Yu, Lijun Zhang (theoretical, medium evidence)

Theory provides conditions under which sign-based optimizers dominate standard SGD when noise is sparse and objectives satisfy l1-stationarity, guiding optimizer choice for high-dimensional, sparse regimes.

Model-only protocol predicts AI can reduce population-level idea diversity and quantifies excess crowding Nafis Saami Azad, Raiyan Abdul Baten (theoretical, medium evidence)

A modeling toolkit offers early-warning metrics (Δ, ρ) to anticipate diversity crowding from LLM use, enabling platforms and policymakers to monitor creative homogenization risk ex ante.

Trajectory-fidelity metric exposes hidden workflow shortcuts in payment agents that outcome metrics miss Donghao Huang, Joon Kiat Chua, Zhaoxia Wang (descriptive, medium evidence)

Agentic Success Rate (ASR) evaluates entire action trajectories in autonomous agents, revealing skipped checkpoints and unsafe shortcuts that task success rates overlook; deploy trajectory-level audits in sensitive agent deployments.

Decision theory shows optimal cascades are pairwise thresholds; routers often beat cheap-then-expensive cascades Dylan Bouchard (theoretical with empirical validation, medium evidence)

A formal cost–quality analysis indicates optimal cascades are pairwise threshold policies and that pre-generation routing can outperform naive “cheap-then-escalate” designs in practice.

Semantic-envelope repair yields provable certificates against platform gaming of scalar safety metrics Florian A. D. Burnat, Brittany I. Davidson (theoretical, medium evidence)

By formalizing how announced metrics get gamed, the paper provides classwise semantic-envelope repairs that can yield certifiable reductions in harm, useful for regulators designing auditable metrics.

Enforcing static fairness can worsen long-run disparities; investment policies can reduce long-term inequity at low cost Shahin Jabbari, Chen Wang (theoretical, medium evidence)

A sequential selection model indicates one-shot fairness constraints can increase long-run disparities, while targeted investments can eliminate them with modest efficiency loss.

High-quality AI drafts halve audio-description completion time for novices, but low-quality drafts give marginal gains Lana Do, Shasta Ihorn, Charity M. Pitcher-Cooper, Sanjay Mirani, Gio Jung, Hyunjoo Shim, Zhenzhen Qin, Kien T. Nguyen, Vassilis Athitsos, Ilmi Yoon (quasi_experimental, medium evidence)

In accessibility workflows, AI drafts are associated with large time and cognitive-load reductions for novices only when draft quality exceeds a domain threshold, underscoring the value of quality control.

WIOA retraining rarely shifts participants into less automation-exposed jobs; wage recovery drives measured 'success' Julian Jacobs, Jordan Canedy (administrative-panel analysis, medium evidence)

Analysis of 23 million records finds retraining seldom moves workers into less-automatable roles, with employer-led apprenticeships performing best on wage recovery.

A tunable governance + learner system adaptively allocates tasks and can reduce fatigue while preserving performance Vicente Pelechanoa, Antoni Mestre, Manoli Albert, Miriam Gil (quasi_experimental, medium evidence)

HAAS combines rules with a contextual-bandit to shift human–AI task allocation; moderate governance appears to improve outcomes and buffer fatigue as the learner gains experience.

An RL Feasibility Index identifies jobs RL can learn that prior exposure metrics miss Philip Moreira Tomei, Bouke Klein Teeselink (descriptive, medium evidence)

A reinforcement learning (RL)–based learnability score across 17,951 O*NET tasks flags occupations with higher RL feasibility than prior exposure indices suggested, informing workforce planning.

Firm-level AI development raises skill premiums via substitution and capital deepening Hui Liang, Xuxia Zhang, Jingbo Fan (quasi_experimental/panel, medium evidence)

Using AI-related patents, firm-level AI activity is associated with higher skill premiums, driven by substitution of lower-skilled labor and capital deepening.

AI boosts TFP for resource-constrained firms but shows diminishing returns for tech-advanced firms Xu Chu, Qingyu Han (panel correlational, medium evidence)

Panel evidence from Chinese firms indicates uneven total factor productivity (TFP) gains: larger for firms with constrained hardware or weaker human capital, minimal where capabilities are already saturated.

Human–AI decision consistency boosts confidence, trust and performance even if AI paradigm alone doesn't Yingying Wang, Qin Ni, Haoxin Xu, Jiaqi Yin, Tingjiang Wei (RCT, medium evidence)

In an experiment with pre-service teachers, human–AI agreement increases confidence, which elevates trust and improves performance, spotlighting alignment over interface style.

Screen at the algorithmic margin and serve highest-risk people directly when uncertainty is irreducible Santiago Cortes-Gomez, Mateo Dulce Rubio, Carlos Patino, Bryan Wilder (theoretical, medium evidence)

A two-stage allocation model indicates screening should target cases near the algorithmic decision boundary, and its value rises with aleatoric (irreducible) uncertainty; useful for social and humanitarian targeting.

Automated audit finds hundreds of thousands of hallucinated citations concentrated in AI-active fields and early-career teams Zhenyue Zhao, Yihe Wang, Toby Stuart, Mathijs De Vaan, Paul Ginsparg, Yian Yin (quasi_experimental/audit, medium evidence)

An audit of 111M references across 2.5M papers finds a sharp rise in non-existent citations since the LLM era, concentrated in AI-active areas and early-career authorship, raising integrity and detection concerns.

Strict scoring with endogenous benefits makes truthful probabilistic reporting impossible unless approval is binary Lauri Lovén, Sasu Tarkoma (theoretical, n/a evidence)

A mechanism-design result indicates smooth approval functions distort incentives and break truthful scoring; a binary threshold restores truth-telling and can match welfare under Brier scoring in some settings.

Where agents substitute for humans, competitive wages are bounded by compute rental rates Siqi Zhu (theoretical, n/a evidence)

Modeling AI agents as compute capital implies human wages in substitutable tasks face an upper bound tied to compute rental prices, shifting attention to compute markets and factor shares.

AI-assisted task-splitting makes more granular, more complete plans but developers prefer hybrid workflows Luka Pavlič, Reinhard Bernsteiner, Stephan Schlögl, Christian Ploder (RCT (controlled experiment), medium evidence)

In controlled tasks with GitLab Duo, AI generates more granular and complete plans but injects irrelevant items, reinforcing the case for human-in-the-loop planning.

Cost-aware inline router matches best-model correctness while cutting inference cost ~84% Sharad Agarwal, Pooria Namyar, Alec Wolman, Rahul Ambavat, Ankur Gupta, Qizheng Zhang (descriptive/engineering, medium evidence)

Switchcraft routes tool calls to the cheapest model that maintains correctness, matching best-model accuracy at a fraction of cost, a practical route for agent deployments.

CEO–TMT divides weaken green innovation, and AI adoption magnifies that harm Zhiyu Chen, Jianzu Wu, Chongchong Lyu (panel correlational, medium evidence)

Firm panels indicate that executive-team faultlines correlate with reduced eco-attention and green innovation, with AI adoption associated with a strengthening of this negative pathway, underscoring the role of organizational cohesion.

LSTM/Transformer models outperform classical benchmarks in China A‑share return prediction with improved tail-risk control Haoyu Wang, Dejun Xie, Yuqing Duan, Wenze Xiong, D. Z. Chen (correlational/empirical, medium evidence)

Deep sequential models beat linear and tree benchmarks on backtests in A‑shares while managing downside risk, though standard backtest caveats apply.

Trace-prior RL preserves realistic pricing behavior under partial observability where revenue optimization alone fails Peiying Zhu, Sidi Chang (descriptive/simulator, medium evidence)

In a two-hotel simulator, pure revenue optimization yields unrealistic price paths; learning a prior over realistic traces and penalizing divergence restores market-plausible behavior without sacrificing revenue.

Minimal dynamical model shows human–LLM feedback can drive degenerative low-diversity equilibria Xuening Wu, Yanlan Kang, Qianya Xu, Kexuan Xie, Jiaqi Mi, Honggang Wang, Yubin Liu, Zeping Chen (theoretical/simulation, low evidence)

A compact dynamical system suggests heavy human–LLM coupling can push ecosystems toward low-diversity equilibria, highlighting a potential long-run risk that warrants empirical follow-up.

Schema-1, a 140M-parameter Data Language Model, outperforms GBDTs and AutoML on tabular tasks without preprocessing Eda Erol, Giuliano Pezzoli, Ozer Cem Kelahmet (descriptive/benchmark, medium evidence)

A “Data Language Model” trained on 2.3M tabular datasets reportedly outperforms gradient-boosted decision trees (GBDTs) and AutoML on prediction and imputation from raw tables, promising for messy enterprise data but requiring replication on proprietary stacks.

Government-guided funds accelerate firms' digital–intelligent transformation by easing financing and sharing policy know-how Fangzheng Zhu, Yuexiang Lu (quasi_experimental/DID, medium evidence)

Difference-in-differences results suggest public funds catalyze digital-intelligent upgrades by alleviating financing constraints and diffusing policy knowledge.

AI adoption follows an inverted-U for firm innovation; too much AI ties to homogenization Xu Fan, Benye Wang (panel correlational, medium evidence)

Innovation rises with AI adoption up to a turning point (≈ 2.948) and then correlates with lower novelty and patent-text homogenization, implying absorptive-capacity limits.

Productivity vs learning and skills

The evidence suggests a familiar split: coding assistants and workflow aids increase output, but they do not, by themselves, build durable skills. A large meta-analysis finds moderate productivity gains alongside null effects on learning, and retraining data suggest workers rarely exit highly automatable roles without employer-led programs. Firm panels reinforce the skew: capital- or capability-constrained firms appear to benefit most from AI, while advanced firms see diminishing returns. Editorially, this points to a two-track strategy—optimize tools for output while separately investing in apprenticeships, coaching, and absorptive capacity to convert short-run gains into human capital.

Measurement, metrics, and governance

Standard success metrics can mask brittle behavior. Trajectory-fidelity audits and price-trace diagnostics uncover shortcutting and market-implausible paths that headline metrics bless. Theory indicates that once you announce a scalar metric, platforms can optimize to it unless you repair the metric classwise or move to binary approvals that restore stronger incentives for truthful reporting. Engineering fixes and theoretical certificates are complementary; the former brings realism in production, the latter brings assurance under adversarial conditions. The direction of travel is clear: measure the process, not just the outcome, and design metrics that are harder to game.

Human–AI collaboration, trust, and workflow design

Presentation and process-level nudges often outperform abstract explainability. In clinics, breaking advice into verifiable claims sharply increases trust; in pair programming, adaptive feedback that tracks joint mental effort and gaze alignment improves debugging. Governance can be a lever rather than a brake: tunable task allocation reduced fatigue without harming performance, and high-quality AI drafts meaningfully helped novices, but only past a quality threshold. Scaling these gains will likely require investment in UX, real-time signals, and domain-specific quality control.

Innovation, diversity, and long-run risks

Several models and empirical panels point to homogenization risks as AI scales: inverted-U innovation at firms, diversity crowding in creative outputs, and even degenerative equilibria in coupled human–LLM systems. Organizational context matters—leadership faultlines correlate with weaker green innovation, and AI adoption can amplify those weaknesses. Editorially, short-term efficiency and long-run diversity are in tension; monitoring novelty and investing in exploratory R&D are sensible policy hedges.

Verifiable claims interface unlocks clinical trust established

An RCT shows atomic, claim-level fact-checking triples clinician trust in oncology recommendations relative to standard explainability.

Implication: Regulated deployments should prioritize interfaces that certify verifiability to accelerate safe uptake.

Coding assistants boost output, not learning suggestive

A meta-analysis estimates a moderate productivity gain and no significant learning effect from GenAI support in programming.

Implication: Pair assistants with pedagogy (explanations, reflection prompts) if the goal is skill growth, not just throughput.

Outcome metrics hide unsafe or unrealistic behavior descriptive

Trajectory-level audits and trace diagnostics reveal shortcutting and implausible action sequences that task success rates miss.

Implication: Make trajectory fidelity a go/no-go criterion in agent deployments, alongside accuracy and cost.

Routers can beat naive model cascades framework

Decision-theoretic analysis indicates pairwise-threshold cascades are optimal and that pre-generation routing can outperform cheap-then-escalate pipelines.

Implication: Treat routing as a first-class optimization problem to reduce cost without sacrificing quality.

Hallucinated citations surged where LLMs are prevalent suggestive

A large-scale audit associates the post-LLM period with a sharp rise in fabricated references, concentrated in AI-active fields and early-career teams.

Implication: Journals and institutions should adopt automated reference verification and author attestations.

Systematic meta-analysis of GenAI programming effects

A meta-analysis of the effect of generative AI on productivity and learning in programming

Provides pooled estimates across 23 studies and documents heterogeneity by context, setting a benchmark for realistic effect sizes.

Dual eye-tracking + pupillometry with reactive/proactive AI feedback

Cognitive Alignment Drives Attention: Modeling and Supporting Socially Shared Regulation in Pair Programming

Operationalizes joint mental effort and attention in teams and shows causal improvements from feedback, a template for process-aware tooling.

RL Feasibility Index mapped to occupational tasks

What Jobs Can AI Learn? Measuring Exposure by Reinforcement Learning

Scales task-level learnability across 17,951 O*NET tasks, bridging ML feasibility with labor taxonomies for targeted policy.