World Wire

Researchers from Stanford, EPFL, and UNC introduce Weak-for-Strong Harnessing, W4S , a new Reinforcement Learning RL framework that trains a small meta-agent to design and refine code workflows that call a stronger executor model. The meta-agent does not fine tune the strong model, it learns to orchestrate it. W4S formalizes workflow design as a multi turn Markov decision process, and trains the meta-agent with a method called Reinforcement Learning for Agentic Workflow Optimization, RLAO . The research team reports consistent gains across 11 benchmarks with a 7B meta-agent trained for about 1 GPU hour. https://ift.tt/2W6zucw W4S operates in turns. The state contains task instructions, the current workflow program, and feedback from prior executions. An action has 2 components, an analysis of what to change, and new Python workflow code that implements those changes. The environment executes the code on validation items, returns accuracy and failure cases, and provides a new sta...

Microsoft Research proposes BitNet Distillation , a pipeline that converts existing full precision LLMs into 1.58 bit BitNet students for specific tasks, while keeping accuracy close to the FP16 teacher and improving CPU efficiency. The method combines SubLN based architectural refinement , continued pre training , and dual signal distillation from logits and multi head attention relations. Reported results show up to 10× memory savings and about 2.65× faster CPU inference , with task metrics comparable to FP16 across multiple sizes. What BitNet Distillation changes? The community already showed that BitNet b1.58 can match full precision quality when trained from scratch, but converting a pretrained FP16 model directly to 1.58 bit often loses accuracy, and the gap grows as model size increases. BitNet Distillation targets this conversion problem for practical downstream deployment. It is designed to preserve accuracy while delivering CPU friendly ternary weights with INT8 activa...

Kong has open-sourced Volcano , a TypeScript SDK that composes multi-step agent workflows across multiple LLM providers with native Model Context Protocol (MCP) tool use. The release coincides with broader MCP capabilities in Kong AI Gateway and Konnect , positioning Volcano as the developer SDK in an MCP-governed control plane. Why Volcano SDK?   because 9 lines of code are faster to write and easier to manage than 100+. Without Volcano SDK? You’d need 100+ lines handling tool schemas, context management, provider switching, error handling, and HTTP clients.  With Volcano SDK:  9 lines . Copy Code Copied Use a different Browser import { agent, llmOpenAI, llmAnthropic, mcp } from "volcano-ai"; // Setup: two LLMs, two MCP servers const planner = llmOpenAI({ model: "gpt-5-mini", apiKey: process.env.OPENAI_API_KEY! }); const executor = llmAnthropic({ model: "claude-4.5-sonnet", apiKey: process.env.ANTHROPIC_API_KEY! }); const database = ...

Are your LLM code benchmarks actually rejecting wrong-complexity solutions and interactive-protocol violations, or are they passing under-specified unit tests? A team of researchers from UCSD, NYU, University of Washington, Princeton University, Canyon Crest Academy, OpenAI, UC Berkeley, MIT, University of Waterloo, and Sentient Labs introduce AutoCode , a new AI framework that lets LLMs create and verify competitive programming problems, mirroring the workflow of human problem setters. AutoCode reframes evaluation for code-reasoning models by treating problem setting (not only problem solving) as the target task. The system trains LLMs to produce competition-grade statements , test data , and verdict logic that match official online judges at high rates. On a 7,538-problem benchmark built from prior datasets, AutoCode achieves 91.1% consistency with official judgments ( FPR 3.7%, FNR 14.1% ). On a separate, more difficult 720 recent Codeforces problems (including interactive task...

Reinforcement Learning RL post-training is now a major lever for reasoning-centric LLMs, but unlike pre-training, it hasn’t had predictive scaling rules. Teams pour tens of thousands of GPU-hours into runs without a principled way to estimate whether a recipe will keep improving with more compute. A new research from Meta, UT Austin, UCL, Berkeley, Harvard, and Periodic Labs provides a compute-performance framework —validated over >400,000 GPU-hours —that models RL progress with a sigmoidal curve and supplies a tested recipe, ScaleRL , that follows those predicted curves up to 100,000 GPU-hours . Fit a sigmoid, not a power law Pre-training often fits power laws (loss vs compute). RL fine-tuning targets bounded metrics (e.g., pass rate/mean reward). The research team show sigmoidal fits to pass rate vs training compute are empirically more robust and stable than power-law fits, especially when you want to extrapolate from smaller runs to larger budgets. They exclude the very...