MagicBeanStalk targets the “zero-advantage trap” in RLVR data synthesis
MagicBeanStalk argues that simply adding more verifiable RLVR problems doesn’t guarantee learning signal, because many datasets collapse into a “solved every rollout or failed every rollout” pattern. The paper introduces a recursive decomposition approach that grows seed problems into calibrated families of tasks where some rollouts succeed, some fail, and at least one correct solution fits the training budget, leading to sizable gains on math benchmarks and finance datasets.
Meta launches Muse Wonder Full for subscribers, aiming to power its Muse agent
The Verge reports that Meta is rolling out two new models – Muse Wonder and Muse Wonder Full – to Muse Code subscribers, with Wonder also reachable via Meta’s model API. Meta says the longer-term plan is to plug the models into its Muse AI agent and further broaden access with specialized models, though it withholds detailed technical specifics in this launch coverage.
CLEAT teaches agents when clarification is worth asking
CLEAT frames interactive intent alignment as clarification that must translate into better decisions, not just more information. The method co-trains a clarification-value controller with the task policy by estimating the task utility of asking versus proceeding, showing improved performance on multiple benchmarks and underspecification settings by adapting what to ask as the agent learns.
Microsoft’s Maia 300 chip promises higher-efficiency AI inference than GPUs
Microsoft announces its third-generation in-house Maia 300 as a primary inference engine for AI workloads across Azure products and services. The Verge highlights a planned 2027 production rollout and claims up to 40% more tokens per megawatt than leading GPU-based systems, with positioning focused on both efficiency and better support for reasoning and agentic workloads.
Faithful hardware knowledge graphs from messy datasheets, without one-shot context limits
This paper tackles the brittleness of feeding long, table-heavy datasheets directly into an LLM by building a hierarchical hardware knowledge graph through chunk-based extraction, semantic routing, and iterative refinement. It reports 98.8% property faithfulness versus a 76.1% one-shot baseline, aiming to reduce overlooked constraints that can cascade into incorrect component selection or integration decisions.