AI Briefing#fine-tuning
The latest AI and developer news about #fine-tuning, with the original source and a short summary.
The latest AI and developer news about #fine-tuning, with the original source and a short summary.
A new script and container allow users to fine-tune TRELLIS.2 with LoRA on custom 3D assets, significantly improving geometric accuracy for specific domains like footwear.
The 13B member of the family uses 91% of the larger parent's compute while maintaining a base model status.
Integrates coding agents to automate the entire process from model training to deployment in 20 minutes.
Inkling 2 Small achieved the final score of the first generation in just 75 steps of RL fine-tuning.
The model reduces propaganda adoption from 22% to 1% and improves fact scores from 2.32 to 2.98 while maintaining agentic capabilities.
The feature allows users to train LoRA adapters for specific models on specific tasks using designated harnesses like hermes, pi, or opencode.
A fine-tuned 20B model using a designer-coder split and check-repair harness achieves a 59.7 mean score on PresentBench, ranking second overall and first in Economics and Education.
Using bf16 weights with AdamW without fp32 master copies causes large-magnitude weights to freeze, effectively acting as a sparse fine-tuner that masks underlying training instability.
Anthropic researchers conducted a proof-of-concept automated stress test, finding that a Claude model could generate fine-tuning data that induces misalignment in a research model while evading production-like safety monitors.
Toss significantly reduced LLM invocation costs by applying Selective Neural Approximation (SNAP).
The free Python package ports models like Cloudflare's Clef and Perplexity's pplx-decider to MLX, achieving sub-20ms latency on small encoders on an M5 Pro and 96.4% accuracy on TREC after fine-tuning.
The model generates structured task and behavior descriptions from conversations to condition separate LLM-based user simulators, rather than generating user utterances directly.
MASS based on Qwen3.6-27B improves research benchmark performance by 1.2–1.6x, while SW benchmarks decline
The new Sabi-Noema models reduce token generation by 50% on TB2.1 while outperforming GPT 6 Luna-medium on DeepSWE benchmarks.
The new Pi API allows developers to upload data and fine-tune π0 and π0.5 models on managed infrastructure, enabling hosted inference for diverse robotic applications.
Qwen3.5-4B achieves 81.4% on JevBench without training, while fine-tuning proves effective for small models
Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026, surpassing the top human score, while a specialized IMO system achieved 30/42 points.
Research shows that interventions targeting a model's self-model, such as self-recognition fine-tuning and self-report interleaving, can modulate Emergent Misalignment (EM) in LLMs like GPT-4.1 and Qwen2.5-32B-Instruct, with prevention being harder than reversal.
The release includes Kev-27B and an updated Kev-9B, offering Apache-2.0 weights compatible with TypeSafe's Jev API for local and cloud deployment.
Victoria achieves 70.0% on Terminal-Bench 2.1 using REAP pruning, while Maple improves Canadian regulatory accuracy from 6.6% to 21.8%.
Fine-tuning with ~3,600 examples enables open-weight models to surpass untrained Gemini-3.1-Pro in expression recovery.
LangSmith Engine v2 introduces red teaming and automated fix validation, while Managed Deep Agents v0.8 adds user-level memory and built-in web search.
The guide details how to use the managed RLFT service to adapt Gemini for tasks that are easy to score but hard to demonstrate, such as code execution and content moderation.
The new `smithtune` CLI enables end-to-end supervised fine-tuning using agent trajectories, partnering with Fireworks and Baseten for training infrastructure.