A university team has shown a training method that lets a neural network absorb new skills for months without degrading the abilities it already had. The approach reserves a small fraction of weights per task and rehearses compressed memories of earlier data, sidestepping the catastrophic forgetting that has forced most systems to be retrained from scratch. In a year-long trial across forty sequential tasks, accuracy on the earliest task fell by under two percent. Independent labs are now attempting to replicate the result at larger scale before anyone declares the problem solved.
A published architecture change replaces dense attention over very long inputs with a learned routing layer that sends each query to a small set of relevant memory blocks. Benchmarks show comparable quality on document question answering while using roughly a tenth of the computation at million-token lengths. The authors caution that gains shrink on tasks requiring genuinely global reasoning, where every token matters. Several open-weight projects have already begun porting the technique, and early community reproductions broadly match the reported numbers on smaller models.
A laboratory has trained multi-fingered robotic hands to perform delicate tasks such as threading cable ties and opening pill bottles, using only ordinary video of people doing the same jobs. The system infers contact points and force profiles without motion-capture gloves or teleoperation data, which have long been the expensive bottleneck in robot learning. Success rates remain below human levels on unfamiliar objects, but the training pipeline costs a fraction of previous approaches. The team plans to release the dataset so other groups can test generalisation.
Computational biologists report a generative model that designed novel enzymes which functioned as intended when synthesised in the wet lab, without the usual rounds of trial and error. Three of eight candidate designs catalysed their target reaction at rates comparable to natural counterparts. The result matters because failed designs, not compute, dominate the cost of protein engineering. Reviewers note the targets were relatively well-understood reaction classes, and independent replication on harder chemistry will determine whether the approach generalises beyond the demonstration set.
A machine-learning weather system has outperformed a leading physics-based operational forecast across several standard measures at ten- to fourteen-day horizons, running in minutes on a single accelerator rather than hours on a supercomputer. Meteorologists welcomed the speed but stressed that the model inherits biases from its training reanalysis and has not been tested through a full cycle of unusual seasons. National forecasting agencies are running it in shadow mode alongside conventional systems before considering any operational role in public forecasts.
Engineers have described a training schedule that prunes connections continuously as a model learns, arriving at a network roughly a third the size of a comparable dense model with equivalent benchmark scores. Unlike post-training compression, the sparsity pattern emerges during learning, so the network adapts around its own removals. The technique reduces inference cost meaningfully for deployment on modest hardware. Practitioners note that current accelerators do not exploit unstructured sparsity efficiently, so realised speedups depend heavily on the hardware and kernel used.
A multilingual speech model can now reach usable transcription accuracy for languages with as little as ten hours of recorded audio, down from the hundreds of hours previously required. The gains come from a shared phonetic representation learned across a thousand languages, which transfers to new ones with minimal fine-tuning. Community linguists have begun applying it to endangered languages where recordings are scarce. Researchers stress that transcription quality still depends on dialect coverage, and that speaker communities should control how their recordings are used.
A newly published evaluation suite of freshly written problems, never posted online, found that leading systems score far below their results on established public benchmarks. The authors argue the gap reflects contamination of training data with test items rather than genuine reasoning ability. Model developers dispute the size of the effect, pointing to differences in prompt formatting and grading rubrics. Regardless of the exact figure, the study has strengthened calls for held-out, regularly rotated evaluations maintained by parties independent of the labs being tested.
A joint academic and industrial team has identified a solid electrolyte candidate for lithium batteries after an automated search proposed thousands of compositions and an autonomous lab synthesised the most promising fifty. The winning material remained stable across hundreds of charge cycles in early cells. Battery researchers describe it as an encouraging demonstration of closed-loop discovery rather than a finished product, since manufacturability at scale is untested. The group published the full candidate list, including failures, which reviewers praised as unusually transparent practice.
A generative geometry model can produce a navigable three-dimensional scene from one ordinary photograph, filling in surfaces the camera never saw with plausible detail. Architects and game developers have shown early interest as a rough drafting tool. The researchers are careful to describe the hidden regions as invention rather than reconstruction, warning against use in forensic or measurement contexts where the fabricated geometry could be mistaken for evidence. Code has been released under a licence that explicitly excludes surveillance applications.
Researchers have shown that training a small model on the intermediate reasoning steps of a much larger one, rather than only its final answers, transfers a surprising share of its problem-solving ability. A model small enough to run on a laptop recovered most of the accuracy of a system many times its size on mathematics and coding tasks. The method depends on the quality of the generated reasoning traces, and the authors found that filtering out incorrect chains before training was essential to the result.
An interpretability group has traced a specific pathway inside a language model that retrieves factual associations, and demonstrated they could edit individual facts by adjusting a small number of weights. Changing where a landmark is located, for instance, propagated consistently through the model without disturbing unrelated knowledge. The work advances the case that these systems can be understood mechanically rather than treated as opaque. The authors caution that fact editing at scale risks introducing subtle inconsistencies that current evaluation methods would not catch.