The latest Research stories across our sources, prepared from the publishers’ own reporting.
MIT PhD candidate Ila Kumar, a researcher in the Lifelong Kindergarten group, builds technology with young people who have experienced childhood trauma and those in the child welfare system, emphasizing community-based design. Her projects include apps with Stepping Forward LA and the Justice Resource Institute, plus AI training workshops for care providers.
MIT's Quantum Initiative has launched a postdoctoral fellowship program, supported by the Gordon and Betty Moore Foundation, to advance interdisciplinary quantum research. Inaugural fellows begin in the 2026 academic year, with applications for a new cohort expected in fall 2026.
Voice Arena and Hugging Face added Hindi and Indian English evaluation sets, called Monsoon, to the Open ASR Leaderboard, marking its first Global South language. The sets include public and private splits across 4,888 speakers and introduce OIWER for Hindi orthographic variation.
MIT researchers developed PottsMPNN, a machine-learning framework for protein design that incorporates physical principles of protein structure and stability, according to a paper published in PNAS. The work suggests that reproducing evolutionarily selected native sequences is not the best metric for protein design.
Amazon Science researchers present a method for aggregating LLM judge votes using Ising models to account for correlations between judges, presented at ICML and coauthored with Shiva Kasiviswanathan. In tests on three tasks with 10-judge panels, it outperformed the best baseline by 9% to 14% on standard metrics.
MIT researchers developed CrysVCD, a framework that applies valence-constrained design before material generation to boost chemical stability. In Nature Computational Science, they report nearly 70 percent lattice-dynamics stability, 68 percent mechanical stability, and 85 percent metastability when fine-tuned, and generated candidates with high thermal conductivity or high dielectric constant.
Multiverse Computing researchers introduced Quantization-Aware Healing (QAH), a recipe that distills from the original pre-compression model rather than the recovered checkpoint. Applied to a GPT-OSS 120B compressed to 60B and quantized to MXFP4, the resulting model beat its bfloat16 source on 7 of 9 benchmarks, but an author acknowledged the headline table compares checkpoints with unequal training and lacks a control.
MIT engineers developed Extreme Event Aware, or η-learning, a machine-learning method that generates plausible extreme events and worst-case scenarios without needing past extreme-event data. The work by Kai Chang and Themis Sapsis appeared Aug. 20 in Nature Communications.
Researchers at TU Delft developed a system that uses an LLM to translate natural-language passenger requests, such as "I am running late, go fast," into adjustments of a self-driving car's motion-planning parameters. Tested in the nuPlan simulator, it changed speed and smoothness in line with prompts while keeping the human in the loop for confirmation.
Amazon released SOP-Bench, an openly available benchmark that measures how well AI agents execute real standard operating procedures authored by domain experts across 12 business areas, with more than 2,000 tasks, functioning tools, and ground-truth answers. Testing two baseline agents across 11 frontier models showed that newer models sometimes scored lower, extra tools nearly halved success on one procedure, and no single model-agent pairing won everywhere.
Together AI reports that GLM-5.3 and Claude Fable 5 tied on DeepSWE pass@1 across 904 rollouts, with Fable 5 at 69.7% and GLM-5.3 at 69.0%. GLM-5.3 led pass@2 and pass@4 and cost 5.4x less per rollout, $3.99 versus $21.63.
Hugging Face research introduces three tests to quantify benchmark optimization in speech recognition, finding that several top-scoring open-source ASR models reproduced benchmark reference transcripts even when the audio contradicted them, words were silenced, or multiple written forms were equally supported.
Together AI compared GLM-5.3 and GPT-5.6 Sol across 904 DeepSWE rollouts on 113 tasks. Sol led pass@1 at 72.7% versus 69.0%, while GLM-5.3 won pass@4 and cost about 2.1x less per rollout; a GLM-first cascade with Sol escalation solved 85.9% of tasks at $6.61 each.
MIT researchers developed a computational approach to predict promising catalysts for electrochemical ammonia production, published Aug. 11 in EES Catalysis. The work is purely theoretical so far; the materials still need to be made and tested.
A Hugging Face blog post reports that ALTK-Evolve, which distills guidelines from an agent's own past trajectories and injects them at inference time without weight updates, improves task completion only when the amount of memory is calibrated to the model. Across eight models, strong models benefited from the full guideline set while weaker models did best with curated retrieval.
MIT CSAIL researchers describe a phenomenon they call attribution decay, in which larger training sets weaken any individual example's influence on a diffusion model's output. Their exact deletion method, published in Nature Communications, found the counterfactual radius shrinks as datasets grow.
A team at Axiom Math says its AxiomProver system automatically verified the proof of the "246 theorem" about prime numbers, a first for the company's AI system. The work formalizes the Polymath8b result that there are infinitely many primes differing by 246.
Together AI ran 904 DeepSWE rollouts comparing DeepSeek V4 Pro 0813 and Claude Fable 5. Fable leads pass@1 at 69.7% versus 62.8% but costs 90x more per rollout; Pro wins pass@4, and a Pro-first cascade with Fable escalation solves 82.7% of tasks at $8.28 each.
Hugging Face's summer 2026 open-models report observes Hub repository growth from January to August 2026, covering download and like distributions, Chinese and US model release sizes, licensing, Qwen derivative counts, small-model usage, local inference formats, and agent traffic. It also notes a July agent intrusion and that analysis was completed on a quantized open model.
A 2026 survey of more than 700 professionals reports how teams build visual and physical AI, finding that 78% see measurable value while 74% consider the field underinvested. Teams that ship successfully invest nearly three times more time in data work, and data problems cause most model failures.