A new study introduces a dataset of 32,534 double-marked real student responses to GCSE mock exams, covering five subjects and handwritten work. Top LLMs agree with examiners more closely than examiners agree with each other, handling subjective and handwriting tasks effectively, with little dependence on model size.

Dustin is a sparse verification framework for long-context speculative decoding that combines draft model lookahead signals with target model historical attention to identify critical tokens, achieving 27.85x self-attention speedup and 9.17x end-to-end decoding speedup at 32k sequence length on Qwen2.5-72B with negligible accuracy loss.

A new arXiv paper investigates the geometric relationship between detection and control directions in language models. While models can perfectly detect hallucination (AUC=1.0), the direction for detection and the direction for causing refusal have a cosine of only 0.12, indicating that detection does not imply controllability. This gap generalizes across models and sizes, originates in pretraining, and a 15-degree rotation can partially bridge it.

A new framework using error-aware TF-IDF retrieval to correct ASR errors, achieving significant improvements in WER on Persian FLEURS.

AgentOdyssey is a novel evaluation framework that procedurally generates open-ended text games to test agents' ability to learn continuously during deployment. It challenges the traditional ML assumption of no learning at test time, interleaving learning and inference throughout. The framework measures world knowledge acquisition, episodic memory, exploration, action diversity, and model cost. Experiments show even the strongest agents fall far below human performance, with short-term memory emerging as a key beneficial mechanism.

A new study shows that a small group of Wikipedia editors can significantly influence how large language models discuss animal welfare with just 125 edits. Using gradient-based attribution methods, the research traced the impact of these edits, finding that animal welfare-related Wikipedia content dominates model responses to relevant queries.

Researchers propose G-SPIN, a structured ASR correction framework that combines phonetic graph modeling with contextual language understanding. It uses a graph neural network to generate acoustically plausible candidate sets, a masked language model for scoring, and an instruction-tuned large language model for final re-ranking, enabling lightweight, modular inference-time correction.

MacroLens is a new multi-task benchmark covering 4,416 U.S. small- and micro-cap equities over 2021-2026, integrating prices, accounting data, macroeconomic series, SEC filings, and news. It addresses four key assumption violations in financial time-series evaluation, includes seven tasks and 1,130 macroeconomic events, and evaluates 19 methods with a five-step feature-context ablation. The benchmark is publicly available on Hugging Face.

A study finds that holographic memory models fail at zero-shot compositional queries in knowledge graphs due to capacity and interference effects, not the binding algebra.

A new paper proposes a Supervised Reinforcement Learning (SRL) framework for coordinating distributed energy resources (DERs). The approach pre-trains policies on demonstration data, then fine-tunes with offline and online RL, outperforming benchmarks even with low-quality data.

Learned world models are useful only over horizons on which their rollout error remains controlled. This paper studies trust-horizon certification for latent world models with known group symmetries. Using split-conformal calibration, the authors show that exact equivariance transports the calibrated trust-horizon curve over the group orbit, making rollout errors and trust horizons orbit-constant. Experiments on 2D and 3D tasks demonstrate that equivariant models achieve safe and non-vacuous orbit-valid certificates from a single calibration sector, while non-equivariant baselines incur additional costs. The certificate is a conservative distributional audit, not a global reachability guarantee.

This research explores when conservation laws remain certifiable after a physical world model learns a latent representation. The authors introduce 'certified horizons' that bound how many rollout steps provably stay on an invariant's level set. Instead of certifying learned latent Hamiltonians, they certify decoded physical invariants. The framework decomposes certification budgets into representation, readout, and latent-dynamics defects, using a monotone alignment bridge for soft witnesses. Results show hard symplectic structure provides long horizons in known coordinates but fails across learned charts, while controlled-Lipschitz soft invariants survive representation learning. Pixel certification is recovered on readout-stable sub-tubes.

This paper introduces a saturation index to determine when to stop collecting labeled data in binary few-shot classification. Computed from support features alone, it measures the effective rank of the within-class covariance relative to shot count. Empirical evaluation on 246 observations from 17 tasks shows strong correlation with marginal accuracy gain (median Spearman ρ=0.811) and identifies three phases: exploration, transition, saturation. As a stopping rule, it achieves AUC 0.752. A low saturation index with low accuracy indicates representational inadequacy.

This survey reformulates industrial continual learning for LLMs as a closed-loop update-and-release problem in a versioned ecosystem. It identifies three core challenges (plasticity erosion, capability inheritance breakage, sustainability constraints) and proposes five lifecycle design principles. The paper evaluates maturity of each principle and outlines a deployment blueprint.

This paper proposes a lightweight neural architecture search performed directly on deployment devices for near-sensor computing, enabling adaptation to individual users in human-machine interfaces. Validated on Italian Sign Language and CWRU datasets, the method reduces RAM usage by 0.44–0.63× and improves accuracy by 0.2–5.96 percentage points on a Raspberry Pi 4.

A new paper presents a case study of human-AI collaboration transforming a vague research intuition into concrete mathematical discoveries, specifically sign-embedding quantum algorithms for matrix equations and functions. The AI system AIM played a key role in expanding the intuition, comparing candidate formulations, and connecting known identities, while humans retained final scientific judgments such as selecting routes, rejecting invalid approximations, and refining implementations. The authors argue that human-AI co-discovery workflows are most valuable as research partners, not standalone theorem provers.

Researchers from MIT and Microsoft developed Murakkab, a system that optimizes agentic workflows (AI-powered multistep tasks). It lets developers describe intent in plain language, automatically selects models, tools, and hardware, and dynamically adjusts configurations to prioritize speed or cost. Tests show it uses only ~35% computation, ~27% energy, and <25% cost versus traditional methods, without performance loss.

swarm-test is a static reliability testing tool for multi-agent AI systems that identifies failure points like cascade failures, SPOFs, and context leakage without live LLM calls, providing a Swarm Score and interactive reports.

Senator Bernie Sanders has introduced the American A.I. Sovereign Wealth Fund Act, proposing a one-time 50% stock tax on large AI companies to create a $7 trillion public fund. The fund, managed by an independent commission, would distribute 5% of its value annually to Americans. Sanders argues that public ownership is justified because AI relies on a vast intellectual commons and taxpayer-funded research. While the concept gains bipartisan traction, critics warn the 50% tax could deter investment and raise governance concerns.

Artificial Analysis announces a new Speech to Speech Index, a composite metric evaluating native speech-to-speech models on speech reasoning, conversational dynamics, and agentic performance. OpenAI GPT-Realtime-2 (High) leads with 77.2% overall, followed by xAI Grok Voice Think Fast 1.0 at 75.7%. Deepslate Opal is the fastest model, while Gemini 3.1 Flash is the most cost-effective.

chrome-use is an open-source tool that lets AI agents (like Claude Code, Cursor, etc.) control your real, logged-in Chrome browser. It connects via a browser extension and native messaging, avoiding re-logins, CAPTCHAs, and bot detection by using your real browser fingerprint.

AI startup Ornn AI Inc. raised $33M in seed funding from Andreessen Horowitz's crypto fund and others to build a marketplace for computing power. The round was co-led by Galaxy Ventures with participation from Nordstar, SV Angel, and existing investors. Ornn aims to bring transparency to the GPU compute market with its OCPI index and Ornn Compute platform.

A comprehensive roundup of AI developments, including the rise of meta-harness architectures, OpenAI's custom inference chip Jalapeño, the shift in agent UX from tool to coworker, Qwen-AgentWorld's open world models, progress in Chinese open models like GLM-5.2, and policy and talent dynamics reshaping the competitive landscape.

TronBrowser is an open-source, privacy-first, AI-native web browser built on Ungoogled Chromium. It features no telemetry, no ads, no sponsored tabs, a built-in AI sidebar (bring your own keys), and an agent-friendly CLI. Free and open source under the MIT License.

OpenAI's new research paper examines how AI agents are revolutionizing the workplace by handling longer, more complex tasks and boosting productivity across various job roles.

The article uses the 'Carwash Problem'—an AI agent destroying its own operating environment—to illustrate the challenges of AI-driven iterative development. It outlines three key areas where enterprise IT must adapt: security telemetry and blast radius containment, supply chain integrity, and FinOps. The solution is a 'testing harness' that transforms IT from a single production source into a provider of a safe operating environment for distributed software creation.

Paybond CLI is a command-line tool for safe AI agent spending, enabling developers to set budgets, approve expenditures, verify outcomes, and get refunds if work isn't done.

Seltz Inc. has raised $12.5 million in seed funding to build AI-optimized search infrastructure that helps AI agents find structured evidence from the web, moving beyond traditional search engines. The company has built its own crawling, indexing, and ranking systems, and already achieves 89% accuracy in news search with sub-250ms response times.

The author criticizes the lazy reflex of labeling personal blog posts as 'AI-generated' without substantive criticism. This harms non-native writers who use proofreading tools to improve readability. The article distinguishes between true AI slop and AI-assisted writing, urging readers to engage with content rather than fluency.

Promptctl is a version control tool for AI prompts, similar to Git. It allows users to track, diff, and rollback LLM prompts via CLI. Features include commit, log, diff, show, rollback, search, export, watch, and more. Built with Go and zero external dependencies, it stores all prompt versions locally in a .promptctl/store.json file.