Model Releases

  • Carbon-X Genomics LLM — A Qwen3-14B-based LLM fine-tuned on genomic data for variant-effect prediction and conversational biology queries. Built with standard transformers and text-generation-inference tooling, it’s aimed at filling the bio-NLP gap for researchers who want to chat with their genome data instead of writing pipelines.

  • Fibo-2.0 Diffusion Model — A diffusion model released with ComfyUI support and arXiv backing, packaged in a single-file format to make sharing and versioning less of a headache in ComfyUI workflows. If you’re generating images and tired of juggling model files, this might save you a few clicks.

Open Source Releases

  • jaxonomy 3.5.0 — A JAX-native simulation engine for hybrid dynamical systems, now at 3.5.0 with block-diagram modeling, ODE/DAE solvers, event handling, acausal physics, and full autodiff support. Think of it as JAX-powered Simulink for people who want differentiable physics simulations without leaving the NumPy-ish ecosystem.

  • claude-code v2.1.311 — Adds --hold-idle-file and --max-hold-min flags to keep self-hosted sessions alive longer, plus a small UX tweak where Claude explains its reasoning in a more readable output style. Useful if you’re running long-lived agent sessions and keep losing context to timeouts.

  • jorbit 1.7.2 — A JAX-based library for solar system orbit fitting and integration, targeting astrophysics and engineering folks who need differentiable celestial mechanics. If you’re modeling spacecraft trajectories or exoplanet dynamics and want gradients through your orbital integrator, this is your new best friend.

  • rishe 0.6.0 — A Rust-native Persian NLP engine that does tokenization, lemmatization, spell-checking, ezafe detection, informal-to-formal normalization, and embedding generation — all at “blazingly fast” speeds. For teams working with Persian text who are tired of Python-based pipelines choking on large datasets.

  • llm-failover-chain 0.6.1 — A multi-provider LLM failover library with rate-limit-aware routing, automatic retries, and seamless provider switching. If your app depends on LLM APIs and you’ve ever had a provider hiccup take down your service, this is the duct tape you didn’t know you needed.

Research Worth Reading

  • Prospect Theory in Large Language Models — A meta-analysis showing that LLMs exhibit loss aversion, diminishing sensitivity, and nonlinear probability weighting across decision-making prompts — basically, they’re biased like humans, just in different ways. The empirical curves suggest these biases are consistent enough to be modeled, not just quirks.

  • End-to-End Differentiable Diffusion Conditioning for Physics-Informed Optimization — Introduces a differentiable diffusion conditioning framework that bakes physics constraints into offline model-based optimization, letting you do gradient-driven design of physical systems from fixed datasets alone. If you’re doing engineering design and hate having to hand-craft objective functions, this might help.

  • When and How Does Muon Precondition? — Theoretically and empirically breaks down Muon’s preconditioning under structured gradient noise, showing that non-isotropic noise unlocks faster convergence than standard optimizers in large-scale LLM training. Comes with practical noise scheduling guidelines — read this if you’re tuning optimizers and want to squeeze out extra speed.

  • Optimal Correction Sets for Provable Redress of Contestable Causal Discovery — Formalizes repairs in causal discovery as defeasible commitments using argumentation theory, and proposes optimal correction sets that guarantee alignment with a subset of conditional independence tests. If you’re building causal inference pipelines and dealing with noisy or conflicting test results, this gives you a principled way to fix things.

  • AOS-Bench: Can LLM Agents Manage Dynamic Workloads? — A benchmark testing whether LLM agents can handle concurrent, dynamic tasks like builds, data processing, and bug fixes on a single machine. The results expose real trade-offs between parallelism and context interference — useful if you’re evaluating agentic workflows for CI/CD or devops automation.

  • Scoped Semantic Invariance as a Foundation for LLM Reliability — Proves that full semantic invariance in LLMs is undecidable, then proposes scoped invariance as a practical workaround. Demonstrates improved consistency tracking on real tasks — worth a look if you’re trying to make LLM outputs more predictable across reworded prompts.

Today’s Synthesis

If you’re building ML systems that need to reason about real-world constraints — whether in biology, physics, or orbital mechanics — the common thread this week is that differentiable, constraint-aware modeling is getting practical. jaxonomy 3.5.0 and jorbit 1.7.2 both offer full autodiff through complex dynamical systems, meaning you can now gradient-descent your way through spacecraft trajectories or hybrid physical systems without leaving JAX. Pair that with End-to-End Differentiable Diffusion Conditioning for Physics-Informed Optimization , which bakes physics constraints directly into offline model-based optimization, and you’ve got a toolkit for designing physical systems from fixed datasets using gradient-driven methods. The practical takeaway: if you’re tired of hand-crafting objective functions for engineering design problems, or wrestling with black-box simulators that won’t give you gradients, these tools let you treat physics as differentiable code — no symbolic derivations required.