Research papers, technical reports, benchmarks and datasets.
Voice Arena and Hugging Face have added Monsoon en-IN and Monsoon hi-IN to the Open ASR Leaderboard, making Hindi the first Indian language represented on the multilingual leaderboard.
Why it matters: The Open ASR Leaderboard has added its first Hindi and Indian English benchmarks, showing how speaker attributes are being incorporated into public evaluations.
Cleaning data and aligning the reward function for RLVR takes expertise and effort upfront, but the result is a model that's state-of-the-art on a complex task. Guest post by researchers at UIUC and Bridgewater, in collaboration with our team. https://thinkingmachines.ai/news/putting-task-expertise-into-rl
Google Research has introduced the experimental Planetary Prediction Engine (PPE) under Google Earth AI. From a natural-language query, it can autonomously discover geospatial data, engineer features, train and evaluate models, and produce a report—compressing weeks of manual data engineering into minutes.
Why it matters: The article reports quantified improvements for PPE in public health, food security, and epidemic forecasting, giving a view of the practical limits of geospatial modeling automation.
Google has released GlucoFM, a self-supervised foundation model with a dual-stream design that separates slow glucose trends from short-term fluctuations while preserving timing and missingness information. Across 14 evaluations spanning four cohorts and seven clinical tasks, it improved average PR-AUC by 5.8 percentage points over the best GluFormer variant and achieved the lowest MAE for predicting post-meal glucose response. It was pretrained on 109,066 hours of unlabeled CGM data and performed best at cross-dataset transfer and few-shot adaptation.
Google has released AgentHands, a research prototype described in a paper at CHI 2026, that maps LLM reasoning output to gestures synchronized with speech in XR. This lets an agent point things out and demonstrate objects or steps in the user's physical space. The system comprises four modules: object registration, a gesture-event library, gesture-embedding inference, and synchronized execution in a headset. Gestures are categorized by dimensions including left or right hand, spatial anchoring, and VFX. In a controlled study with N=12, AgentHands significantly improved spatial reference over a voice-only baseline.
Multiverse Computing has published a paper proposing Quantization-Aware Healing (QAH). After compressing GPT-OSS 120B to 60B parameters and quantizing it to MXFP4, the method distills directly from the original model before compression rather than from a recovered bfloat16 checkpoint.
Why it matters: The paper presents a repair recipe that distills from the original model after compression and quantization, with stability comparisons against QAT.
METR research finds that AI's acceleration of science is uneven: cybersecurity-vulnerability reports sped up considerably from 2025 to 2026, mathematical research accelerated only slightly, and AI research itself showed no measurable acceleration. A university collaboration proposed SPADE, a self-play framework in which LLMs alternately generate executable training environments and solve them. Qwen3-30B-A3B scored 58.3 across the game-environment suite, an 8.1-point gain over its base model.
Google has released Biomarker Discovery Framework, a multi-agent system that organizes candidate biomarker prioritization into an iterative, human-supervised research loop combining hypothesis generation, parallel statistical analysis, model training, adversarial validation, and literature reasoning. Across 9,279 participant observations in three cohorts, it identified 41 candidate digital biomarkers for mental health and 25 for metabolic health. One example: variability in sleep duration correlated with PHQ-8 severity (ρ = 0.252).
Jim Fan announces T-Rex, an open-source tactile robot learning methodology developed through a collaboration between NVIDIA and Berkeley. Its mixture-of-transformer architecture pairs a slow visuomotor expert with a fast tactile expert running at 4 touch ticks per vision tick; a 50-hour (~5,500 episodes) synchronized robot play dataset collected on tactile hand hardware with 22 degrees of freedom is available today on HuggingFace. Extending EgoScale, its training recipe combines human egocentric video pretraining with tactile robot play mid-training, which Fan says helps bridge contact-free pretraining and contact-rich manipulation.
Why it matters: T-Rex offers a transferable approach to contact-rich robot learning by combining human video pretraining, synchronized tactile robot data, and separate visual and tactile processing rates.
Google Research has proposed the Mobility-Embedded POIs (ME-POIs) framework, combining anonymized mobility patterns with text descriptions to generate embeddings that encode both a place's identity and its changing function.
Hugging Face has published research measuring benchmark optimization in speech recognition through three tests: consensus disagreement, masked-entity retrieval, and spelling switches. The study evaluates 11 open-source ASR models.
Why it matters: The study presents three reproducible probes and quantifies how often ASR models reproduce incorrect benchmark transcripts, with direct implications for evaluation design.
Microsoft Research has released Skala 1.1, a deep-learning DFT functional trained on 2.5 times as much data as the previous public version. It ranks first in 32 of the 55 GMTKN55 benchmark categories, with a weighted mean error of 2.8 kcal/mol. The release is available in CP2K and is being integrated into Psi4, FHI-aims, ORCA, and VASP. Microsoft has also introduced a living benchmark that tracks computational performance across versions.
The IBM research team has published follow-up work on ALTK-Evolve, evaluating eight models on 585 multi-step tasks in AppWorld. It finds that agent memory is not an on/off switch but a dosage that needs to be calibrated for each model.
Why it matters: A comparison across eight models shows that the right amount of agent memory varies with model capability and offers two transferable ways to control costs: retrieval and prompt caching.
Import AI 469 introduces DiG-bench (Discovery in Games), a new benchmark of 70 games with hidden rules that tests AI's ability to infer environmental rules through exploration. Opus 5 and Fable 5 with Claude Code perform best, but complete only 0.2 of the tasks at Tier 7, far below the human success rate of 100%.
Google Research has proposed PhotoScan, a deep-learning framework that estimates body-fat percentage, the android-to-gynoid fat ratio (A/G), and the visceral-to-subcutaneous fat area ratio (V/S) from ordinary 2D phone photos. It was pretrained on more than 35,000 UK Biobank records, fine-tuned on a cohort of 677 adults, and validated on an independent cohort of 132 people.
Hugging Face ran the ICML 2026 Open Reproduction Challenge from July 15 to August 2. A community of 1,221 used coding agents such as Claude Code, Codex, and Cursor to reproduce papers, publishing 6,816 Trackio logbooks covering 2,226 papers—about 34% of all papers at the conference.
Why it matters: Hugging Face has published reproduction results and adjudication data for more than 2,200 ICML papers, offering a view of the limits of agents in scientific reproduction.
Microsoft Research has introduced MindTopo, a benchmark for topological reasoning in multimodal LLMs. It covers connectivity, separation, ordering, containment, and knot relations, across static reasoning and interactive planning. Current models are substantially better at static recognition than interactive planning, but both remain far below human performance; failures occur more often in planning than perception. Image and video generation helps only when topology is preserved within a single frame and remains unreliable across multi-step operations.
Google Research has proposed a knowledge-profiling framework that distinguishes factual errors caused by encoding failures from those caused by retrieval failures, and released WikiProfile, a benchmark containing 2,150 Wikipedia facts.
Why it matters: By separating encoding failures from retrieval failures, knowledge profiling offers a new way to diagnose the sources of factual errors in frontier models.
Google Research has released AMIE (Video), which uses Gemini and Project Astra for real-time video clinical interviews. It can perceive nonverbal cues, guide virtual physical examinations, and reason about diagnoses in real time.
Why it matters: AMIE extends from text to live video consultations, showing how an asynchronous multi-agent architecture can balance conversational latency with clinical reasoning.
Microsoft Research has released CARE-X, a research model and unified chest X-ray vision-language model that can generate reports, make structured diagnostic predictions, and localize lesions. It uses reinforcement learning (DAPO) to reward clinical correctness across multiple tasks.
Hugging Face's ALTK-Evolve and ACE are both agent-memory approaches that require no weight updates or human annotations. ALTK-Evolve treats memory delivery as a tunable control, retrieving a small number of well-supported principles for each task rather than injecting the full playbook at every step.
A new Hugging Face paper proposes two systems improvements that sharply reduce the training cost of distilling LLMs: cache the teacher model's top-100 logits offline so teacher and student do not need to reside in GPU memory simultaneously, and use a fused, chunked KL loss to avoid generating a full vocabulary-by-sequence-length matrix.
Researchers from the University of Toronto, Vector Institute, University of Cambridge, and ServiceNow built a prototype computer worm that uses GPUs on compromised machines to run open-weights LLMs, sustaining inference while spreading attacks. It achieved about 80% vulnerability-detection success, 53% exploit success, 88% self-replication success, and about 37% end-to-end attack success.
Google Research has proposed Chain-of-Evidence (CoE), a framework for verifiability implemented in the autonomous-research prototype Science One Framework and the automated auditing tool CoE Audit.
Why it matters: The article reports a verifiability framework and audit metrics, helping readers compare autonomous research systems on citation and code consistency.
Microsoft Research has introduced Echoverse, a set of 12 training worlds for computer-use agents: 10 full-domain environments and two capability environments. Each consists of an application, tasks, and a database-based verifier.
Why it matters: Microsoft Research has released 12 computer-use-agent training environments with verifiers and reports comparisons of transfer performance between deep and shallow environments.
Berkeley AI Research and IBM Research have extended the evolutionary kernel-search framework K-Search to MLX and added a structured CUDA-to-MLX translation layer that automatically adapts existing CUDA-kernel knowledge to Apple Silicon.
Why it matters: The article details the CUDA-to-MLX translation layer and reports two kernel benchmarks, showing how kernel knowledge can be transferred across platforms.
Epoch and METR have released MirrorCode, a benchmark testing AI systems on coding tasks that take humans a long time to complete. Of 25 target programs, 17 achieved a perfect-score run at least once, while eight never reached the 100% threshold.
Berkeley AI Research has proposed ABBEL, a framework that separates an LLM's recursive summaries into natural-language “belief states” and scores their content with supervised signals to reduce performance loss from context compression in long interactions. On the collaborative-programming benchmark CollabBench, reconstruction-based belief scoring cut the performance gap with a full-context model by about 50%, reduced training from 100 steps to 50, and kept peak context length well below the full-context setup.
Google Research has published SymptomAI, a study in which five Gemini Flash 2.0-based conversational agents interviewed 13,917 participants and evaluated differential diagnoses (DDx).
Why it matters: Google's randomized study compares five questioning strategies across 13,917 people and, for the first time, compares diagnoses with Fitbit physiological signals, offering evidence on real-world conversational triage.
In a paper published in Nature, Google Research proposes using reinforcement learning to let autonomous agents learn from quantum error-correction detection events, continuously adjusting thousands of control parameters during computation to counter drift.
Why it matters: Google trains a reinforcement-learning agent on quantum error-correction data to continuously calibrate control parameters without interrupting computation.
DharmaOCR scored 0.925 on a Portuguese OCR benchmark, ahead of Mistral OCR4 at 0.798 and Unlimited-OCR at 0.7587. Two-stage training with Portuguese supervised fine-tuning followed by DPO concentrates all model parameters on Brazilian Portuguese, giving it a domain-specific advantage in extraction quality and stability.
In a paper at ICLR 2026, Google Research explains that diffusion-model creativity comes from score smoothing during neural-network training. Regularization makes the learned score function smooth, so denoising interpolates between training data points and generates new samples.
Hume has released Real World VoiceEQ, a benchmark for human quality in voice interaction. It covers more than 40 proprietary and open-source voice models, over 15 evaluation dimensions, and more than 60 metrics across ASR, TTS, speech-to-speech, and voice understanding.
Why it matters: Hume built a voice-AI benchmark from more than a million human ratings, showing why current metrics can overestimate real conversational performance.
Google Research and Google DeepMind have introduced SensorFM, a large sensor foundation model that learns directly from unlabeled wearable data. Its pretraining corpus contains more than one trillion minutes of multimodal sensor signals from five million consenting participants across more than 100 countries, all 50 U.S. states, and over 20 Fitbit and Pixel Watch devices.
Why it matters: The article reports pretraining-data scale, scaling experiments, and downstream-task comparisons, helping readers assess how generalizable wearable foundation models are.
In the first large-scale real-world study of its kind, published in Nature Cities, Google Research modified the Google Maps algorithm to direct about 2% of trips to alternative routes across 10 U.S. cities for six months. Median travel speeds on target roads rose by about 2%, median fuel-consumption rates fell by 0.5% to 1.0%, and each city could cut thousands of tonnes of CO2e emissions annually.
Hugging Face has released ScarfBench, an open-source benchmark for evaluating AI agents on enterprise Java framework migrations. It covers the Spring, Jakarta EE, and Quarkus ecosystems, with 34 applications, 102 framework implementations, and 204 migration tasks.
A Hugging Face blog introduces DiScoFormer (Density and Score Transformer), a single model that estimates both the density and score of a data distribution in one forward pass, without retraining for a new distribution.
NVIDIA has released ENPIRE, a framework in which coding agents drive physical robots through autonomous experiments and improvements. It achieved 99% success on tasks including PushT, peg sorting, and cutting zip ties with a cutter; tests also tried inserting a GPU into a motherboard.
Google Research has introduced an architecture that attaches a multi-token-prediction (MTP) head to the frozen Gemini Nano v3 model, speeding up on-device generation without changing the backbone weights.
Why it matters: Google attaches an MTP head to frozen Gemini Nano v3 and reports concrete methods and gains under on-device memory and energy constraints.
Google Research has proposed linear elastic caching, modeling page eviction as a ski-rental problem and using lightweight machine learning to adjust cache size dynamically. After months of testing in Spanner production, memory use fell 15.5%, cache misses rose just 5.5%, total cost of ownership fell about 5%, and the impact on actual I/O costs was only 0.5%.
