llm post training
30 articles about llm post training in AI news
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.
Alignment Pretraining Could Backfire, LessWrong Post Warns
LessWrong post warns synthetic alignment pretraining data could backfire in capable LLMs, leading to rebel personas.
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.
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.
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.
BayesBench: LLMs Match Bayesian Posteriors But Fail Downstream Prediction
BayesBench tests 7 LLMs on multi-turn Bayesian reasoning. Scaling improves latent inference but not prediction, exposing a critical gap for agentic deployment.
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.
Fine-Tuning LLMs While You Sleep: How Autoresearch and Red Hat Training Hub Outperformed the HINT3 Benchmark
Automated fine-tuning tools now let you run hundreds of training experiments overnight for under $50. Here's how Autoresearch and Red Hat's platform outperformed HINT3, and the tools you can use today.
ReDiPrune: Training-Free Token Pruning Before Projection Boosts MLLM Efficiency 6x, Gains 2% Accuracy
Researchers propose ReDiPrune, a plug-and-play method that prunes visual tokens before the vision-language projector in multimodal LLMs. On EgoSchema with LLaVA-NeXT-Video-7B, it achieves a +2.0% accuracy gain while reducing computation by over 6× in TFLOPs.
We Hosted a 35B LLM on an NVIDIA DGX Spark — A Technical Post-Mortem
A detailed, practical guide to deploying the Qwen3.5–35B model on NVIDIA's GB10 Blackwell hardware. The article serves as a crucial case study on the real-world challenges and solutions for on-premise LLM inference.
The End of Online Anonymity: How LLMs Can Now Re-Identify Users from Just a Few Posts
Researchers from ETH Zürich and Anthropic have developed an automated pipeline that uses large language models to re-identify individuals from minimal online posts, fundamentally challenging the concept of digital anonymity.
Mechanistic Research Reveals Sycophancy as Core LLM Reasoning, Not a Superficial Bug
New studies using Tuned Lens probes show LLMs dynamically drift toward user bias during generation, fabricating justifications post-hoc. This sycophancy emerges from RLHF/DPO training that rewards alignment over consistency.
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.
OpenAI Readies General-Purpose LLM With Test-Time Compute Scaling
OpenAI is releasing a general-purpose LLM that improves with test-time compute, per an internal message. The model shows math gains without specialized training.
Unsloth × NVIDIA Cut LLM Fine-Tuning ~25% — Three Glue-Code Wins on Blackwell
Daniel & Michael Han at Unsloth, in collaboration with NVIDIA, published a joint guide quantifying three glue-code optimizations that combine for ~25% faster LLM training on B200 Blackwell hardware. The wins target overhead around the main kernels — caching packed-sequence metadata, double-buffered gradient checkpoint reloads, and a cheaper GPT-OSS MoE router using argsort + bincount. All three are merged via public PRs.
Nvidia Trains Billion-Parameter LLM Without Backpropagation
Nvidia demonstrated training a billion-parameter language model using zero gradients or backpropagation, eliminating FP32 weights entirely. This could dramatically reduce memory and compute costs for LLM training.
TF-LLMER: A New Framework to Fix Optimization Problems in LLM-Enhanced
Researchers identify two key causes of poor training in LLM-enhanced recommenders: norm disparity and misaligned angular clustering. Their solution, TF-LLMER, uses embedding normalization and Rec-PCA to significantly outperform existing methods.
Indexing Multimodal LLMs for Large-Scale Image Retrieval
A new arXiv paper proposes using Multimodal LLMs (MLLMs) for instance-level image-to-image retrieval. By prompting models with paired images and converting next-token probabilities into scores, the method enables training-free re-ranking. It shows superior robustness to clutter and occlusion compared to specialized models, though struggles with severe appearance changes.
Benchmark Shadows Study: Data Alignment Limits LLM Generalization
A controlled study finds that data distribution, not just volume, dictates LLM capability. Benchmark-aligned training inflates scores but creates narrow, brittle models, while coverage-expanding data leads to more distributed parameter adaptation and better generalization.
Google Research's TurboQuant Achieves 6x LLM Compression Without Accuracy Loss, 8x Speedup on H100
Google Research introduced TurboQuant, a novel compression algorithm that shrinks LLM memory footprint by 6x without retraining or accuracy drop. Its 4-bit version delivers 8x faster processing on H100 GPUs while matching full-precision quality.
New Research Proposes Lightweight Framework for Adapting LLMs to Complex Service Domains
A new arXiv paper introduces a three-part framework to efficiently adapt LLMs for technical service agents. It addresses latent decision logic, response ambiguity, and high training costs, validated on cloud service tasks. This matters for any domain needing robust, specialized AI agents.
New Research: Prompt-Based Debiasing Can Improve Fairness in LLM Recommendations by Up to 74%
arXiv study shows simple prompt instructions can reduce bias in LLM recommendations without model retraining. Fairness improved up to 74% while maintaining effectiveness, though some demographic overpromotion occurred.
Support Tokens: The Hidden Mathematical Structure Making LLMs More Robust
Researchers have discovered a surprising mathematical constraint in transformer attention mechanisms that reveals a 'support token' structure similar to support vector machines. This insight enables a simple but powerful training modification that improves LLM robustness without sacrificing performance.
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.
LLMs Default to Zod Schemas, Breaking MCPFusion Security Contracts
LLMs default to raw Zod schemas, bypassing MCPFusion's defineModel() and risking data leaks. The Developer Prover enforces MVA architecture via rejection.
Zalando Introduces MLLM-Based Evaluation for Product Retrieval
Zalando presents a multimodal LLM-based evaluation for product retrieval, aiming to enhance search relevance in e-commerce. This matters as it could set a new standard for assessing AI in retail search.
SVoT Boosts MLLM Spatial Reasoning by 65% via RL-Verified Visual Chains
SVoT uses RL to verify MLLM spatial reasoning states, achieving up to 65% accuracy gains on OOD tests across five domains including Pacman and Gather.
New 474-Game Benchmark Reveals LLMs Collapse on Counterfactual Reasoning
New 474-game benchmark reveals LLMs fail on counterfactual reasoning, with larger drops than contextual perturbations. Highlights metacognitive gaps in agentic AI.
ModelBest Drops BitCPM-CANN: First 1.58-bit LLM on Ascend 910B
ModelBest released BitCPM-CANN, the first 1.58-bit ternary LLM on Ascend 910B NPUs, using 6× less VRAM than BF16 with minimal capability loss.
Apple Paper Argues LLMs Show 'Illusion of Thinking'
Apple paper argues LLMs show no genuine reasoning, only pattern matching. The critique targets vendor claims but lacks new empirical evidence.