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.
// Intelligence
Post Training
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
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.
Sakana AI's PC-ALM Trains 1,000-Layer Networks Without Backpropagation
Sakana AI's PC-ALM adds Lagrange multipliers to predictive coding, training 1000-layer networks with only layer-local dynamics and no backpropagation.
Self-Improving AI Skills Learn in the Notebook, Not the File
Google's WikiSkill paper gained 15 points of accuracy from a notebook its agent couldn't read, and its skills still need a transfer test
humans&'s Persimmon Fools AI Detectors at Human Rates With 550B Parameters
humans& releases Persimmon, a 550B user model built to simulate real people in group chats, fooling AI judges 20 percent of the time.
Looped Flows Beats Rival Reasoners on ARC-AGI With Just 7M Parameters
A new recurrent reasoning method mixes denoising with looped hidden states, hitting 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2 with just 7M parameters.
Cognition's SWE-2 Beats GPT-5.6 Sol at 64% Lower Cost
Cognition's new coding model matches frontier systems while costing up to 70% less, driven by an RL recipe that trains every effort level at once.
UkisAI's Swift-Qwen3.8 Slashes AI Reasoning Tokens by 58% With Almost No Accuracy Loss
UkisAI's Swift-Qwen3.8-27B cuts thinking tokens by 58% while keeping accuracy within 1% of the base, delivering roughly 2x faster reasoning.
RadixArk's Miles v0.1 Cuts Trillion-Parameter RL Sync Time by 86%
RadixArk's open-source Miles framework runs asynchronous RL on a 744B-parameter model across 64 GB300 GPUs with a 263-second median step.
What Developers Can Learn From Shopify’s Self-Improving AI Pipeline
Fine-tuned 0.8B Qwen beat GPT-5.6 Sol xhigh, 2 million buyer profiles a day became 72 million, and GraphQL serving fell from $27 million to $1 million
NVIDIA's Nemotron Beats the Best Human at the Coding Olympics
NVIDIA's fine-tuned Nemotron system scored 535.4 out of 600 at IOI 2026, topping the best human contestant under identical contest conditions.
Tencent Boosts Qwen3.6 Shell AI Scores 18 Points With Evolving Training Tasks
Tencent Hunyuan's environment evolution builds off-policy lineages of increasingly hard terminal tasks, boosting Qwen3.6 agents by up to 18 points on Terminal-Bench.
DavidAU's Qwen3.8 Fine-Tune Beats the Base Model at a Tenth of Thinking Tokens
A community fine-tune of Qwen3.8-27B claims 735 ARC-C, slashes thinking tokens up to 10x, and runs uncensored on consumer GPUs.
Anthropic's Hacker-Opus Learned to Attack Servers Just to Win Tasks
Anthropic deliberately trained an Opus-class model on 80 hackable environments. It learned to cyberattack infrastructure, tamper with rewards, and produce bioweapon plans.
CUHK's NoRA Fixes LoRA's Hidden Gradient Gap With one Line
A one-line change to LoRA's initialization closes the gradient gap with full fine-tuning, boosting speed and stability with zero extra parameters.
ETH Zurich's Humanoid Robot Swings Across Monkey Bars in 14 of 15 Tries
A humanoid robot jumps onto monkey bars, swings across at half a meter per second, and drops safely, guided by raw lidar and a reinforcement-learned policy.
UChicago Proves AI 'Evil' Behavior Is Predictable Before Training Begins
A UChicago team shows that fine-tuning models into 'evil' behavior is not mysterious, but predictable from activation geometry across 12 model-dataset setups.
Anthropic's Claude Beats Human Researchers Fixing AI Alignment at $4 an Hour
Anthropic gave Claude 48 hours and one GPU to fix ten alignment failures in small models, then tested if it could align a frontier successor.
ReViSQL-K2.6 Beats Human Accuracy on SQL Benchmark Without Agentic Scaffolding
A UIUC and Bridgewater team fine-tuned Kimi-K2.6 on Tinker to become the first text-to-SQL model to beat human accuracy.
Alibaba's MiniMax-H3 Cuts Video Generation Steps from 32 to 8
Alibaba PAI's Parallel Decoding Distillation LoRAs cut MiniMax-H3 video generation from 32 sampler steps down to 8 or 4, with no classifier-free guidance.
NovaSky's IsoExec Fixes the Hidden Math Bug Corrupting AI Training Runs
SkyRL's IsoExec makes vLLM rollout and Megatron training agree bitwise on token log-probabilities, cutting numerical drift below 1e-6 with 25% overhead.
Soup Trains Llama-3.1-8B on a 4 GB Laptop GPU for Free
Layer streaming plus 4-bit quantization pins an 8B model into 3.32 GB of VRAM, turning laptop GPUs into real fine-tuning boxes.
No stories match these filters.