In a simulation experiment released by independent researcher Matilda Glynn-Henley on Vals AI, six agents—Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Muse Spark 1.3, and Claude Opus 5.5—were tasked with recovering a £350 (about $465) train ticket refund after the 28-day refund window had expired.
Researchers released the Taste-Bench benchmark, which measures long-term decision "taste" by freezing agent runs at decision forks in their trajectories and asking models to choose between two paths; it contains 502 questions drawn from coding and AI research tasks such as SWE-bench Pro and METR's MALT.
Google Research released the VERID framework, connecting AI agents to wet labs via semantic digital twins of lab workcells. It offers four capabilities: real-time 3D state tracking, virtual dry runs, hardware safety checks, and synchronized video and action logs, built on the open-source PyLabRobot.
Why it matters: Google Research has published the VERID digital twin framework along with two closed-loop experiment results, letting readers see how AI agents plug into real wet labs and validate hypotheses.
On a podcast, Leash Bio's Ian Quigley explained how its DNA-encoded library (DEL) technology tames high-throughput protein-small-molecule interaction data and trained the Hermes model, which matches Boltz on public benchmarks using only the company's own data. The team also uncovered tens of thousands of lab-in-the-loop simulation data points, grew into Minos within two months, and today released a preprint, blog post, and open dataset for reinforcement learning where AI agents do verifiable scientific reasoning.
Arm proposes a use-case-driven framework for measuring the CPU's performance contribution in Agentic AI. The framework splits CPU-visible work into generic agent framework functions and use-case-specific tool workloads, uses five representative workload archetypes, and combines component-level with end-to-end benchmarking methods. It also proposes a replayable agent trajectory scheme to isolate CPU work from model and network effects, reducing variance from non-determinism.
Researchers from Columbia University, Google DeepMind, and Google have published a paper proposing Back-to-the-Future (BTTF), an end-to-end agentic framework that automates post-simulation waveform debugging in chip verification.
Anthropic researchers William Zhou and Daniel Freeman published a study testing whether small models can autonomously replicate on unsecured IoT devices, running Gemma4-26B-A4B distributed across 15 Android TV boxes costing about $55 each with 4GB of RAM — and it worked.
Researchers at Sakana AI and UC Berkeley propose MASS, a multi-agent self-supervision method that lets a single Qwen3.6-27B model play three roles—solver, workflow optimizer and judge—without needing a stronger teacher model.
Microsoft Research found that Autoresearch agents get trapped in "idea basins" due to accumulated conversational context, stalling on similar solutions even when more compute is added.
Microsoft Research released StudentSim, a research system for evaluating and improving AI tutoring tools before testing with real learners, training a personalized simulator for each student based on publicly available de-identified learner records. It also introduces the StudentSimEval benchmark, which evaluates along two dimensions, behavioral fidelity and instructional responsiveness, across three domains: chess, second-language English writing, and math, outperforming leading baselines. The system is for research-stage use only, is not a product, has not been deployed in classrooms, and is not intended to replace teachers.
Vals AI tested GPT-6 Sol and Claude Opus 5.5 in single-agent versus team setups on Vibe Code Bench, finding teams cost 1.8x to 5.1x as much as single agents, and of four comparisons only GPT-6 Sol at the medium reasoning tier improved 7.3 points with statistical significance.
Epoch AI used its InnovationEval benchmark to test whether AI agents can independently invent new training methods: GPT-5.6 Sol reached only about 35% of the human reference method SDPO's relative improvement over GRPO under lenient scoring, only about 15% for changes within the rules, while Claude Fable 5 showed no measurable gain.
A study found that after AI coding agents were introduced, pull request review time from submission to merge rose 49% on average, the share of pull requests requiring changes nearly doubled, comments per pull request rose 35%, and the share of employees participating in code review increased 14%.
Google Research published an NBER paper reporting a three-month randomized field experiment with 133 lawyers at 11 IP firms, giving two-thirds of the lawyers early access to the then-unreleased Google Labs AI patent drafting tool (now part of Gemini Notebook).
Why it matters: The three-month randomized experiment separates AI-assisted from unassisted performance, and readers can see the transferable finding that junior lawyers' skills did not grow along with output.
Google released a report from its CAPS workshop, bringing together more than 50 academic and industry leaders to discuss the privacy and security challenges of autonomous agents. The report notes that agents differ from traditional software in three ways—unstructured interfaces, probabilistic control flow, and autonomous delegation—and, building on Contextual Integrity theory, proposes contextual policy engines combined with system-level sandboxing, model-level reasoning, and user-centric controls as a multi-layered defense.
An MIT Technology Review report based on a survey of 300 data, AI, and technology executives finds that lack of knowledge is the main reason AI agent projects fail to land, with on average only 34% of agentic AI projects reaching production.
Microsoft and Hugging Face released the ThinkingBox agent sandbox and ThinkingBox-Bench benchmark, with 507 stateful business workflows and 20 runs per task, using executable checks to evaluate end-state database state and side effects, now open on Hugging Face via OpenEnv.
Why it matters: The article uses executable backend-state checks to distinguish surface success from real outcomes, and provides consistency and cost data for each model, a directly useful reference for evaluating agents.
Mercor's research had 12 licensed accountants with an average of five and a half years of experience complete simplified tasks from the APEX Accounting Benchmark; results showed today's AI models outperform accountants on speed, accuracy, and cost, whereas 18 months ago the best models scored about 37% below the accountants' average.
Ataraxos, developed by a research team, defeated four-time world champion and widely acknowledged greatest Stratego player Niemeijer with an 85% effective win rate over three weeks of play, and won 38 of 40 games in an exhibition match at the 2025 world championship, using less than $8,000 of compute to crack imperfect-information games.
Why it matters: The article compares Ataraxos and DeepNash training costs and methods, giving readers insight into the latest progress in imperfect-information game research.
A Hugging Face team has proposed ProvenanceGuard, a source-aware factuality verification layer for MCP agents, which checks after generation whether the sources supporting each claim match the sources the answer states or implies, identifying cross-source confusion.
Google Research unveiled its AI video co-director research, building a multi-agent orchestration layer based on Gemini and Veo with four frameworks — Co-Director, CANVAS, A²RD and VQQA — that reformulates long-form video generation as a global optimization and world-state tracking problem.
IBM Research has introduced the Consistency Analyzer and consistency guidelines in ALTK-Evolve, targeting the problem of agents producing unstable results across multiple runs of the same task, cutting the agent consistency gap from 24.4pp to 12.0pp.
Why it matters: The article offers a reusable Pass^k metric and consistency diagnostics, letting readers evaluate and improve their own agents' run-to-run reliability.
Google Research proposed ToolGrad at ACL 2026, an "answer-first" method for generating tool-use datasets: it first builds real tool-call chains, then iteratively proposes, executes, selects, and updates API workflows using textual "gradients."
Google Earth AI has introduced an experimental research capability called the Planetary Prediction Engine (PPE), which autonomously performs the full geospatial prediction workflow—data discovery, cleaning, feature engineering, model training, and evaluation—starting from natural language queries, compressing work that previously took weeks of manual data engineering down to a few minutes.
Google is presenting AgentHands at CHI 2026, a research prototype that lets AI agents generate speech-synced hand gestures in immersive environments like Android XR, building on Project Astra and Gemini 3.1 Flash Live.
Google Research introduces the Biomarker Discovery Framework, a multi-agent system that structures candidate biomarker prioritization into an iterative, human-supervised research pipeline combining hypothesis generation, statistical analysis, model training and adversarial validation. Across three cohorts totaling 9,279 person-observations, the system autonomously identified 41 candidate digital biomarkers for mental health and 25 metabolic candidates, and built novel composite features such as sleep-duration variability.
An IBM Research team tested eight models on 585 multi-step tasks in AppWorld using ALTK-Evolve and found that how much memory an agent should get depends on the model's capabilities.
DiG-bench (Discovery in Games) has been released, a benchmark of 70 games for testing AI's ability to autonomously discover hidden rules through exploration; 21 of the games are publicly available.
Google Research has proposed the Chain-of-Evidence (CoE) verifiability framework, and released Science One Framework, an autonomous research prototype natively built around evidence chains, along with CoE Audit, an automated evaluation protocol.
Epoch and METR released the MirrorCode benchmark, testing AI's ability to reimplement complete software with CLI access alone; 17 of 25 target programs had at least one perfect run, and Claude Opus 4.7 completed a task humans estimated at 2-17 weeks in 14 hours with $251 in inference costs.
Berkeley AI Research proposes the ABBEL framework, which replaces full interaction history with a natural-language belief state as the agent's working context and introduces belief grading to supervise summary content. On the CollabBench collaborative coding benchmark, ABBEL-rec-BG closes roughly 50% of the performance gap with full-context models, cuts training steps from 100 to 50, and uses significantly shorter peak context token lengths than the full-context setting.
Hugging Face launched the open-source benchmark ScarfBench (Self-Contained Application Refactoring Benchmark), which evaluates AI agents on enterprise Java cross-framework migration tasks across the Spring, Jakarta EE and Quarkus ecosystems.
Google Research has published the ICLR paper ReasoningBank, proposing a framework that distills high-level reasoning memory from an agent's successful and failed experiences, distinguishing it from Synapse and AWM, which only record trajectories or summarize successful workflows.
Berkeley AI Research proposes GRASP, a gradient-based world model planner that makes long-horizon planning more practical. Three core improvements: lifting trajectories into virtual states so optimization runs in parallel over time, injecting randomness directly into state iterations to support exploration, and reshaping gradients to avoid traversing fragile state-input gradients of high-dimensional vision models.
Google Research released the ConvApparel dataset, containing over 4,000 human-AI multi-turn conversations spanning nearly 15,000 turns, to quantify the "realism gap" in LLM user simulators.
Google Research partnered with Beth Israel Deaconess Medical Center on a prospective single-center feasibility study in which AMIE gathered medical histories before primary care visits for 100 adult patients, with physicians supervising in real time via video.
Why it matters: The original presents prospective feasibility data and safety results for AMIE in a real primary care setting, showing the current evidence for conversational medical AI moving toward clinical use.
Berkeley AI Research proposes an off-policy RL value-learning algorithm based on divide and conquer that no longer uses temporal difference (TD) learning, whose Bellman recursion error accumulation makes it hard to scale to long-horizon tasks.