The Hub is quiet on new weights. Two blog posts lead the day.
On the Hub
IBM ALTK-Evolve consistency guidelines: will it pass the second time
IBM Research posted a consistency guidelines write up on September 15. It covers ALTK-Evolve. The post introduces a Consistency Analyzer. The analyzer resamples an agent's own recorded trajectory. It finds flip prone decision points. It needs one trace. It needs no ground truth. Default k is five completions per decision. 1
The paper is arXiv 2609.08832. It was submitted September 8. It targets AppWorld with ReAct and GPT-4.1. The blog claims Mean@5 77.4 percent. Pass^5 is 53.0 percent. The consistency gap is 24.4 percentage points. After consistency guidelines the gap falls to 12.0 percentage points. Same task Pass^5 rises 16.0 percentage points. Similar task rises 13.0 percentage points. Average accuracy is unchanged. 2
The repo is github.com/AgentToolkit/altk-evolve. It is Apache-2.0. It has 109 stars. No Hub checkpoint exists for ALTK-Evolve. 3
Reef: inference as the start of learning
Human-Agent-Society published a Community Article on September 15. It is about Reef infrastructure for continual self improving agents. The post frames live serving traces as the start of a learning loop. It claims ownership of experience, whole agent, and updates. The loop can run asynchronously. 4
The GitHub repo is github.com/Human-Agent-Society/reef. It is Apache-2.0. It has 2,115 stars. It was created August 31. It was pushed September 16. The blog shows an OpenAI format endpoint client.post("/v1/chat/completions"). A response header x-reef-agent- marks stored inference records. The stack keeps a structured experience stream. It handles off policy staleness, session merging and dedup. 5
No Hub models are linked. Reef is infra, not a foundation model.
