Evolve Quick Reviews
Public Paper Reviews
Structured preliminary reviews of recent open-access papers, generated and continuously updated by the Evolve analysis engine. Each review condenses what the paper claims, how solid the evidence is, and the questions a rigorous peer reviewer would ask — English analysis with abstracts in your language. These are independent AI analyses, not peer review; the canonical record always lives at the publisher.
Ownership in AI-Assisted Everyday Tasks
Megan Wei, Melanie Subbiah, Audrey Lee, Annya Dahmani et al.
Exploratory survey (n=64): AI-task ownership hinges on process, voice, comprehension; insightful but biased sample and untested percentages limit claims.
Transcript-Supervised Post-Training of Generative Speech Enhancement on Real Recordings via Reinforce Adjoint Matching
Julius Richter, Christoph Boeddeker, Yoshiki Masuyama, Kohei Saijo et al.
RAM reward post-training of FlowSE on real CHiME-4 recordings with transcript rewards lifts WAcc by 5.08 points with metrics preserved; credible, baseline and noise-ceiling caveats.
Machine Unlearning for Gibbs Supervised Learning Algorithms
Yaiza Bermudez, Samir M. Perlaza, Iñaki Esnaola
Theory paper: exact unlearning for Gibbs algorithms via risk maximization on the delete-set; clean identity but no algorithm or experiments.
Large Language Models for Programming: Actually Fixing or Reimplementing Incorrect Code?
Alexandru Stefan Stoica, Traian Rebedea, Marian Cristian Mihaescu
Finds LLMs rewrite buggy code instead of minimally fixing it and solve more from scratch; plausible, but 567 pairs from 7 users, no significance tests.
Limits of Confidence in Diffusion
Russ Webb, Amitis Shidani, Alice Bizeul, Dan Busbridge
Theory: parallel discrete-diffusion steps can't be exact for dependent tokens; theorems are sound but empirical support is a single synthetic task.
Language-model groups overstate consensus when replaying human deliberation on a reasoning task
Tengfei Shao
Belief-anchored LLM groups overstate consensus vs the human Wason groups they replay (+34–44 pp); credible preregistered evidence with a post-hoc headline.
Controlling Backchannels in Streamable Full-duplex Models
Maike Züfle, Peter Polák, Sefik Emre Eskimez, Jan Niehues et al.
Proposes a tunable backchannel head for full-duplex models; probing, benchmark, and human results support it, but the human study covers only the 7B model.
FreqCondNorm: Towards Cross-domain Predictive Maintenance through a Frequency-Conditioned Transformer Foundation Model
Zaynab Raounak, Camille LHermine, Zhiguo Zeng
FiLM-style frequency-conditioned normalization for PHM transformers; promising diagnosis results, but the core layer is never ablated against plain LayerNorm
agentic-ger: terminology recovery in long-form speech using global context
Yanqiao Zhu, Wupeng Wang, Zhifu Gao, Xiangang Li et al.
LLM-agent GER using full-transcript context cuts long-form ASR term errors across 32 configs (up to 36.8% B-CER); good ablation, no significance tests.
Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation
Peiying Zhu, Sidi Chang
A preregistered claim-safe protocol (refuse/decompose/refresh) for closed-loop AI evaluation; rigorous, honest simulation study, but single synthetic system.
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li et al.
Rigorous benchmark of the RLCD model Jev as zero-shot alignment-failure detector (median AUROC 0.886); strong methods, but scope limited to a single vendor.
How Do Agent Harnesses Create Value? Planning Information and Release Control in Stateful LLM Agents
Yukun Zhang, Kemu Xu, Yishen Chen
Claims prewritten plans lift agent success 7.17 pp over word-matched controls and cheap verifiers curb false passes; headline effect is statistically fragile.
Diagnose, Recover, Certify: Task Readiness under Hidden Dynamics Changes
Nguyen Viet Tuan Kiet, Huynh Thi Thanh Binh
EG-MPT couples dormant-fault diagnosis, recovery, and deploy-or-abstain certificates; broad experiments impress, but guarantees are model-conditional.
The Organization of Inference: Information, Resource Constraints, and AI Production
Yukun Zhang, Kemu Xu, Yishen Chen
Claims planning's value hinges on token budget and issue visibility; credible preregistered SWE-bench evidence, but narrow tasks and drifting API backends.
Detecting Glaucoma Across Multi-ethnic Myopic and Non-Myopic Populations Using an Uncertainty-Aware Vision Transformer: A Multicentre Model Development and Validation Study
Raghavan Lavanya, Yangqin Feng, Ten Cheer Quek, Quan V. Hoang et al.
Credible ViT glaucoma screening model with strong external validation (16 datasets, AUROC 86–100%); headline metrics aided by excluding uncertain cases.
Two Emojis of Difference: What Multilingual Affective Generation Benchmarks Actually Measure
Fardeen Sadab, Adib Sakhawat
Rigorous self-audit: a multilingual emoji benchmark's leaderboard collapses once annotators are a random factor; persuasive, but rests on only 3 raters.
A Qualitative Model for Reasoning about Path and Support
Abhishek Jaiswal, Zoe Falomir
Hybrid qualitative + center-of-mass solver for Camelot Jr. puzzle; clear formalization, but no experiments back the claim it solves arbitrary levels.
Structured Four-Stage Legal Translation: From Natural-Language Traffic Rules to PROLOG
May Myo Zin, Wachara Fungwacharakorn, Ken Satoh, Katsumi Nitta
Claims a four-stage structured prompt lifts GPT-4.1 traffic-rule-to-Prolog accuracy to 75% vs 60/55%; suggestive, but 20 rules, one model, no significance.
YODAS v3: Over 1 Million Hours of High-Bandwidth, Stereophonic, Multilingual Speech
William Chen, Shinnosuke Takamichi, Sayaka Shiota, Satoru Fukayama et al.
Claims largest open 1.1M-hour 48kHz stereo corpus, 147 languages; evidence supports scale and fidelity, but language tags and transcript quality go unverified.
AgentPProf: Semantic Profiler for Long Horizon AI Agents
Yusheng Zheng, Chaokun Chang, Yu Mao, Tianyuan Wu et al.
Claims semantic profiling of agent trajectories recovers task structure (0.764 B3 F1) and lifts fault localization; plausible gains but evidence rests on case studies and a proprietary segmenter LLM.
Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution
José Luis Pino
Proposes out-of-band kernel halts for rogue AI agents; sound syscall-boundary insight, but closed-source core and unverifiable incident data limit trust.
Precise Convergence Speed of Clipped SGD
David A. R. Robin
Tightens clipped-SGD convergence theory on (L0,L1)-smooth functions: stepsize 1/9β→1/β, stronger criterion, explicit constants; sound but incremental.
Decoupled Learning and Selection in Slate Recommendation for Privacy and Stability Under Noisy Scores
Sam Urmian, Qinyi Liu, Mohammad Khalil
Claims a privacy-scope contract + logged margin certificate for slates; proofs sound though elementary; stability gains solid, utility value unproven.
SWE-Prometheus: Measuring Engineering Governance Improvements in Real-World Repositories
Jiajun Wu, Leixin Sun, Zihan Tan, Yitao Liu et al.
Benchmark for LLM agents improving repo governance; rigorous controls (no-op, template, gate ablation), but single-rollout LLM-judged scoring caps precision.