A 4B Model Gains 5.93 Points as AuK Runs 4.5× Faster

Featured

01 Can Agent Routers Turn Traffic Into Training?

Agent systems generate daily traces of capability needs, model choices, and task outcomes. Teams usually use this telemetry for monitoring, not training. NeoHorse-1 turns the router into a data intake pipeline.

It records the requested capability, assigned model tier, and final interaction for every request. Schema validation, six-dimensional semantic evaluation, and subscenario labels convert those records into training examples. The examples retain reasoning, tool calls, and runtime context.

Router signals organize a three-stage fine-tuning curriculum and online distillation. Evaluation feedback sets the next round’s data mix, closing the evaluation-selection-update loop. The reported 11-benchmark macro average rose from 58.94 to 64.87 for the 4B model. The 9B model improved from 65.60 to 69.04.

These results show measurable gains and a smaller overall gap between the two model sizes. For agent teams, NeoHorse-1 offers infrastructure for continuously converting production interactions into training assets. It is not merely another post-training recipe.

One experiment does not prove recursive self-improvement. Stable gains across multiple cycles, resistance to amplified data bias, and transfer to real tasks still need confirmation.

Key takeaways:

• Promote routing logs from operational data to training data by recording capability needs, model tiers, and task outcomes.
• The 4B model’s gains suggest capability-gap-driven training allocation may narrow differences between model sizes.
• Treat this as testable data-flywheel infrastructure, not evidence that AI systems can already improve themselves autonomously.

Source: NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

02 One Interface for Speech Generation and Editing

AuK provides one interface based on natural-language instructions and audio context. It covers speech generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing.

About 3.03 billion instruction-audio examples provide 1.95 million hours of effective supervision. That scale makes AuK closer to an extensible speech foundation than a point solution.

AuK-Flash completes inference in four steps and delivers a 4.5× end-to-end speedup under identical conditions. A unified workflow therefore does not necessarily require high latency.

The paper reports leading results in zero-shot generation, instruction-controlled generation, and general editing. It also stays competitive on signal-restoration tasks. Open code and weights let developers test new interaction patterns, but they do not make training cheap to reproduce. Quality across every supported scenario still requires full-paper review and hands-on testing.

Key takeaways:

• Natural-language instructions could unify fragmented speech generation and editing features behind one product interface.
• Four-step inference and a 4.5× speedup bring a unified model closer to deployment.
• Evaluate weight availability, training reproducibility, and production reliability as three separate questions.

Source: AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing

03 Change the Environment Before the Agent

A long-horizon agent may fail because its environment does not expose progress soon enough. More demonstrations may not solve that problem.

This work moves the training intervention from supervised fine-tuning to the environment. Its feedback-enriched environment provides richer observations during single-task exploration and later rounds of multi-stage training. Those observations gradually replace direct action guidance.

The method produced consistent gains across Qwen3 model sizes and several RL algorithms on SciWorld and BFCL. It also reduced fluctuations in training entropy and encouraged active exploration on difficult tasks.

The environmental cues became encoded in policy parameters instead of serving only as temporary inference aids. Feedback consistency within each sample group was an important condition for stable optimization. When rewards are sparse, inspect how well the environment exposes state changes before buying more SFT data.

Key takeaways:

• When agent training fails, first check whether the environment exposes task progress promptly.
• Enrich observations instead of providing direct action hints indefinitely.
• Keep feedback consistent within sample groups to avoid destabilizing training.

Source: Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks

04 Miles Makes RL Training Components Verifiable

Miles organizes the RL training loop around accuracy, efficiency, reliability, and scalability. Verifiability, clean design, and customization guide every component.

The system separates inference generation, model training, and weight synchronization into replaceable units. Multiple training backends and synchronization methods give teams clearer boundaries for adapting deployment topologies and locating failures.

One architecture supports full-parameter training, LoRA, online distillation, and supervised fine-tuning. Teams no longer need separate infrastructure for every training method.

The reported case ran fully asynchronous agent training for GLM-5.2 744B-A40B on 64 NVIDIA GB300 GPUs. That proves Miles can operate at this scale. It does not establish general performance or cost advantages.

Teams building post-training platforms can treat Miles as an inspectable engineering baseline. Claims of production readiness still require long-term stability testing under their own workloads.

Key takeaways:

• Evaluate post-training platforms for correctness, fault tolerance, and scalability, not only algorithmic metrics.
• Verifiable, customizable component boundaries support system changes and failure diagnosis.
• A large-scale run proves operational capacity but cannot replace cross-workload performance and cost testing.

Source: Miles v0.1: Production-Level Post-Training