Borui Kang presents Harness Continual Learning (HCL), where an agent's harness (prompts, memories, tools, skills, routing rules) evolves around a frozen foundation model, unlike updating model parameters. The paper defines "harness-level forgetting" as losing reliable behavior due to harness updates and proposes a guarded evolution mechanism. A Continual Optimizer generates candidate harnesses from feedback, and a Continual Evaluator commits changes only after verifying improvement, retention, and validity. Experiments in textual reasoning, multimodal perception, and open-world interaction show capability accumulation and failure recovery, with over 10% relative gains versus baselines.
- Four execution-facing harness components: Task Interface, Experience Memory, Capability Map, and Adaptive Router.
- Controlled retention sweeps show the stability''-plasticity trade-off can be explicitly adjusted at the harness level.
- The work reframes continual learning away from parameter updates toward externalized, inspectable agent state.
- Four execution-facing harness components: Task Interface, Experience Memory, Capability Map, and Adaptive Router.
- Controlled retention sweeps show the stability''-plasticity trade-off can be explicitly adjusted at the harness level.
- The work reframes continual learning away from parameter updates toward externalized, inspectable agent state.
This research introduces fuzzy-function programming via the Program-as-Weights (PAW) paradigm. The method compiles natural-language specifications into compact, locally executable neural artifacts using a 4B compiler and a 0.6B interpreter. This approach matches the performance of much larger models while using a fraction of the memory and providing faster inference. By shifting the role of foundation models from per-input solvers to tool builders, PAW allows for the creation of small, reusable, and efficient programs for tasks that resist rule-based implementation.
Main points:
- Proposes Program-as-Weights (PAW) to compile natural language into neural programs.
- Uses a 4B compiler and 0.6B interpreter for efficient local execution.
- Achieves high performance with significantly reduced memory requirements.
- Reframes foundation models as tools for building reusable artifacts.
- Releases the FuzzyBench dataset with 10M examples.
Main points:
- Proposes Program-as-Weights (PAW) to compile natural language into neural programs.
- Uses a 4B compiler and 0.6B interpreter for efficient local execution.
- Achieves high performance with significantly reduced memory requirements.
- Reframes foundation models as tools for building reusable artifacts.
- Releases the FuzzyBench dataset with 10M examples.
MIT researchers developed a new approach that rethinks how a classic method, known as Bayesian optimization, can be used to solve problems with hundreds of variables. In tests on realistic engineering-style benchmarks, like power-system optimization, the approach found top solutions 10 to 100 times faster than widely used methods.
Their technique leverages a foundation model trained on tabular data that automatically identifies the variables that matter most for improving performance, repeating the process to hone in on better and better solutions. The researchers’ tabular foundation model does not need to be constantly retrained as it works toward a solution, increasing the efficiency of the optimization process.
The technique also delivers greater speedups for more complicated problems, so it could be especially useful in demanding applications like materials development or drug discovery. The research will be presented at the International Conference on Learning Representations.
Their technique leverages a foundation model trained on tabular data that automatically identifies the variables that matter most for improving performance, repeating the process to hone in on better and better solutions. The researchers’ tabular foundation model does not need to be constantly retrained as it works toward a solution, increasing the efficiency of the optimization process.
The technique also delivers greater speedups for more complicated problems, so it could be especially useful in demanding applications like materials development or drug discovery. The research will be presented at the International Conference on Learning Representations.
By mid-2025 China had become a global leader in open-source large language models (LLMs). According to Chinese state media, by July 2025 China accounted for 1,509 of the world’s ~3,755 publicly released LLMs, far more than any other country. This explosion reflects heavy state and industry investment in domestic AI, open licensing (often Apache- or MIT-style), and a strategic pivot by Chinese tech giants and startups toward publicly shared models. The result is a "revival" of open-source AI, with dozens of Chinese LLMs now available for download or use via Hugging Face, GitHub, or cloud APIs. These range from general-purpose foundation models dozens of billions of parameters in size to specialized chatbots and domain experts, many built on Mixture-of-Experts (MoE) architectures.
This article explores the rise of foundation models in time series forecasting. The authors discuss the increasing success of these approaches in areas such as natural language processing and their potential impact on the field of predictive analytics.
