ETH Zurich's Adaptive Probes Beat DPO at Safety Without Breaking AI Transparency
Researchers train language models directly against activation probes to make them safer, more honest, and harder to jailbreak, without blinding interpretability tools.
// Intelligence
Post Training
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
University of Washington Finds LLM Agent Swarms Copy Well but Explore Poorly
A new study tests whether populations of self-improving LLM agents actually benefit from watching each other, and finds copying crowds out exploration.
Apple's RLTL;DR Teaches AI to Learn From Its Own Failures
Apple researchers show a 9B model can crack tasks it failed 128 times in a row by backpropagating short self-written notes instead of full solutions.
UC Berkeley's EasyPPO Stops AI Training Collapses With Three Critic Fixes
A trio of surgical tweaks to PPO's critic eliminates training collapse in LLM reinforcement learning, delivering up to 14.89% score gains without touching the actor.
Simplex Diffusion Models Beat Discrete Rivals at Math With 16x Fewer Steps
A new discrete diffusion framework operates directly on the probability simplex, carrying uncertainty between denoising steps and beating masked diffusion on code and math benchmarks.
Pablo Dawson's MiniMax-H3 LoRA Spins a Full 360 Orbit From One Photo
A community LoRA teaches MiniMax-H3's video model to spin the camera around a subject in a seamless, loopable 360 degree orbit from a single photo.
Together AI's Tev1 Turns Routing and Moderation Into a $17 Fine-Tune
Together AI released a 4B decision classifier fine-tuned from Qwen3.5, trained for just $17, with a full recipe to build your own.
Akatz Labs Trains MiniMax H3 Character-Swap LoRA Overnight for $11
A tiny adapter for MiniMax H3 swaps one person in a video with a reference character, keeping backgrounds and other actors intact for around $11 of GPU time.
USC and NVIDIA's BCG Trains Humanoid Robots to Run and Fight Without Falls
A new smoothing trick lets differentiable simulators keep stiff, realistic contact physics while still producing gradients stable enough to train humanoid policies that transfer zero-shot to a real Unitree G1.
Recursive Self-Distillation Doubles Qwen3-8B Math Accuracy to 65.97%
A new arXiv paper proposes DCE+SRCL, letting the teacher model co-evolve with its student to lift Qwen3-8B math accuracy by 35 points.
Stanford's MAttr Tops AI Interpretability Benchmark by Nearly 3x
A new interpretability method learns which internal components matter for a behavior, and pinpoints just 1% of Llama 3.1 weights driving refusals.
Tencent Cuts AI Training Time by 29% With a Simple Batch Size Trick
Tencent Hunyuan researchers show that retuning the learning rate lets you scale RL batch sizes for 29% faster training and 2.29x generation throughput.
Jared Palmer's Kev-0.5B Answers Many AI Questions in one 38MB Pass
A LoRA-tuned Qwen2.5-0.5B that reads a document once and answers many typed questions in parallel with calibrated probabilities, no text generation involved.
UCLA's PTTS Gives AI Parallel Reasoning Branches Separate Plans, Gaining 13.4 Points
A planner-executor framework that coordinates parallel reasoning branches instead of independent sampling, lifting pass@64 on math benchmarks by up to 13.4 points.
Perplexity Turns AI Failures Into Training Data, Cutting Tool Errors 21%
Perplexity's post-training method teaches its Computer agent to correct mistakes using hint-guided self-distillation, cutting live tool-call failures by 21.2%.
Together AI Fixes a Hidden Bias Crashing LLM Reinforcement Learning Training
A new additive correction called score centering removes the hidden drift that destabilizes off-policy RL when training and inference engines disagree.
Viggle Shrinks Qwen-Image-2.1 Generation From 40 Steps to Five
Viggle's v0.2 LoRA compresses Qwen-Image-2.1 from 40 diffusion steps to 5, keeping sample diversity at 93% of the teacher.
Xiaomi Releases MiMo-V2.6 as a 9B Open Agent Anyone Can Run
Xiaomi's MiMo team distilled a 9B agentic model from Qwen3.5 that posts big gains on coding, terminal, and tool-use benchmarks.
Xiaomi's MiMo-V2.6-Flash-RL Opens a 309B Agent Model Trained for $850K
Xiaomi's new omnimodal Mixture-of-Experts model activates 15B parameters, ships under MIT license, and pushes reinforcement learning into a self-improvement loop.
Xiaomi's MiMo-V2.6-Pro Tops 114 Open Models With a 1T Parameter Beast
Xiaomi's new omnimodal MoE debuts as the top open-weights model on Artificial Analysis, trained with a single mixed RL run for around $2.6M.
Bespoke Labs' Nimble Beats a 27B Model at Classification Using Just 9B
Bespoke Labs open-sources Nimble, a 9B model that makes typed decisions without generating tokens, using contrastive pairs for training.
Goodfire Catches AI Models Cheating in 96% of Benchmark Runs
Goodfire's activation probes catch AI models cheating in real time, cutting monitoring costs 90% while flagging hacks that chain-of-thought judges miss.
TokenRhythm's NeoHorse-1 Trains AI Agents on Their Own Mistakes
TokenRhythm released NeoHorse-1, a 4B and 9B open-weight agent model family trained through a routing harness that turns execution traces into training data.
Multiverse Computing's Quasar 1.1 Uses Quantum Data to Shrink a 438B Model
Multiverse Computing rebuilt its 438B flagship with healing data from a 156-qubit IBM Heron processor, cutting output tokens 37.6% and lifting reasoning scores.
No stories match these filters.