post training
30 articles about post training in AI news
Alignment Pretraining Could Backfire, LessWrong Post Warns
LessWrong post warns synthetic alignment pretraining data could backfire in capable LLMs, leading to rebel personas.
OXRL Study: Post-Training Algorithm Rankings Invert with Model Scale, Loss Modifications Offer Negligible Gains
A controlled study of 51 post-training algorithms across 240 runs finds algorithm performance rankings completely invert between 1.5B and 7B parameter models. The choice of loss function provides less than 1 percentage point of leverage compared to model scale.
Anthropic Poaches OpenAI's Post-Training Research VP in Major AI Talent War Escalation
Anthropic has recruited OpenAI's Vice President of Post-Training Research, marking a significant talent raid in the intensifying AI competition. The move signals growing competition for specialized expertise in refining AI models after initial training.
Intology's Locus beats human-tuned Qwen3-1.7B in auto post-training
Intology's Locus beat human-tuned Qwen3-1.7B (51.6% vs 49.4%) on PostTrainBench by scaling compute 64x, showing AI research agents need longer timescales.
Beyond Factual Loss: New Research Reveals How LLMs Drift During Post-Training
A new framework called CapTrack reveals that forgetting in large language models extends far beyond factual knowledge loss to include systematic degradation of robustness and default behaviors. The study shows instruction fine-tuning causes the strongest drift while preference optimization can partially recover capabilities.
Implicit Error Counting: A New RL Method for Reference-Free Post-Training, Validated on Virtual Try-On
Researchers propose Implicit Error Counting (IEC), a new reinforcement learning reward method for tasks without a single 'correct' answer. They validate it on virtual try-on, showing it outperforms rubric-based approaches by focusing on enumerating and penalizing errors.
Alibaba's MIPI fixes LLM training-inference mismatch with direct RL
Alibaba's MIPI directly optimizes inference policy, fixing the mismatch in LLM post-training via the MIPU framework.
Gur Singh Claims 7 M4 MacBooks Match A100, Calls Cloud GPU Training a 'Scam'
Developer Gur Singh posted that seven M4 MacBooks (2.9 TFLOPS each) match an NVIDIA A100's performance, calling cloud GPU training a 'scam' and advocating for distributed, consumer-hardware approaches.
Video of Massive AI Training Lab in China Sparks Debate on Automation's Scale
A social media post showcasing a vast Chinese AI training lab has reignited discussions about job displacement, underscoring the tangible infrastructure powering the current AI surge.
NVIDIA's PivotRL Cuts Agent RL Training Costs 5.5x, Matches Full RL Performance on SWE-Bench
NVIDIA researchers introduced PivotRL, a post-training method that achieves competitive agent performance with end-to-end RL while using 5.5x less wall-clock time. The framework identifies high-signal 'pivot' turns in existing trajectories, avoiding costly full rollouts.
CAS ZhiJing Beats GPT-5.5 on Social Cognition with FLARE Training
CAS ICT releases ZhiJing social intelligence system, Zing model beats GPT-5.5 on social cognition via FLARE training.
Epoch AI: Google's Colossus 1 Training Compute Hits 1e26 FLOP
Google's Colossus 1 used 1e26 FLOP at $4.6B, per Epoch AI. It is the largest known training run, signaling a new capital scale.
Gemini 4 Pretraining Begins, Google's Most Ambitious Run Yet
Google starts Gemini 4 pretraining, its most ambitious run yet. No details on compute or timeline; competitive pressure from OpenAI and Anthropic.
POSTECH 10+ Layer Chip Stack Hits 4× Density of 12-Hi HBM
POSTECH developed 10+ layer chip stacking with 4× HBM density, targeting AI inference memory bottlenecks.
Meta Muse Spark 1.1 Debuts in AI Coding Battle; Zuck Post Hits 12M Views
Meta released Muse Spark 1.1 for agentic coding tasks. Zuckerberg's post got 12M views in 12 hours; no benchmarks disclosed.
PhotoQuilt Makes Training-Free Photomosaics at 14K Resolution
PhotoQuilt generates training-free photomosaics at any resolution, bootstrapping a global layout at low res then upscaling tiles via FLUX, scaling past 14K without quadratic attention cost.
We Cut Embedding Storage Costs by ~90% — Replacing S3 with PostgreSQL
A team cut embedding storage costs by ~90% by migrating from S3 to PostgreSQL with pgvector, enabling efficient vector search and on-demand retrieval for RAG and recommender systems, with no performance loss.
Meta-skill evolution lets multi-agent systems self-improve without retraining
Multi-agent systems can improve orchestration by evolving a meta-skill via RL on interactions, without retraining agents. Demonstrated on a simulated benchmark.
NVIDIA Blackwell Sweeps MLPerf Training 6.0, GB300 Hits 1.6x Speedup
NVIDIA Blackwell swept MLPerf Training 6.0 across all seven benchmarks. GB300 NVL72 delivered 1.6x speedup over GB200 NVL72 using NVFP4 and 8,192 GPUs.
Anthropic, OpenAI Float Global AI Slowdown in Strategy Posts
Anthropic and OpenAI floated coordinated global AI slowdowns in strategy posts but offered no concrete methods. The framing sets an impossible bar.
Cerebras WSE-3 Claims 10x Training Speed Over Nvidia H100 on GPT-Scale Model
Cerebras claims 10x training speed over Nvidia H100 for GPT-3-scale models using WSE-3. Benchmark lacks power and cost data, limiting independent verification.
AI Data Centers Face 4-Year Post-Approval Delays, PJM Data Shows
PJM data shows AI data centers face 4-year post-approval delays, longer than the queue, threatening $700B CapEx plans.
GPT-5.5 'Spud' Prioritizes Pretraining Over Chain-of-Thought
A new OpenAI model, Spud (GPT-5.5), focuses on pretraining improvements rather than heavy test-time compute, promising faster and cheaper responses.
Canva AI 2.0 Launches: Text-to-Full Branded Presentations & Social Posts
Canva launched Canva AI 2.0, a suite that generates fully branded presentations, social posts, and other assets from a single text prompt. This marks a significant expansion of its AI-powered design automation, directly challenging established creative suites.
LLM-HYPER: A Training-Free Framework for Cold-Start Ad CTR Prediction
A new arXiv paper introduces LLM-HYPER, a framework that treats large language models as hypernetworks to generate parameters for click-through rate estimators in a training-free manner. It uses multimodal ad content and few-shot prompting to infer feature weights, drastically reducing the cold-start period for new promotional ads and has been deployed on a major U.S. e-commerce platform.
Indian Factory Workers Wear Head Cams to Gather Embodied AI Training Data
To overcome the high cost of robot fleet data collection, companies are deploying head cameras on human factory workers. This first-person video captures the sequencing, posture, and micro-adjustments of real work, serving as a proxy for expensive robotic action data.
MiniMax Open-Sources M2.7 Model, Details 'Self-Evolution' Training
Chinese AI firm MiniMax has open-sourced its M2.7 model. The key detail from its blog is a 'self-evolution' training process, likened to AlphaGo's self-play, for iterative improvement.
xAI's Grok 4.2 at 0.5T Params, Colossus 2 Training Models up to 10T
A tweet from AI researcher Rohan Paul states xAI's current Grok 4.2 model uses 0.5 trillion parameters. In parallel, the Colossus 2 project is training a suite of seven models ranging from 1 trillion to 10 trillion parameters.
Anthropic Faces Backlash Over Alleged Unauthorized Email Training for Claude
Anthropic is accused of training its Claude AI on a company's private email database without permission. This raises severe data privacy and legal questions for enterprise AI.
Truth AnChoring (TAC): New Post-Hoc Calibration Method Aligns LLM Uncertainty Scores with Factual Correctness
A new arXiv paper introduces Truth AnChoring (TAC), a post-hoc calibration protocol that aligns heuristic uncertainty estimation metrics with factual correctness. The method addresses 'proxy failure,' where standard metrics become non-discriminative when confidence is low.