DenoiseRL — Reinforcement Learning for Self-Correction from Failures

DenoiseRL is an RL framework that learns from a weak model's own erroneous reasoning traces, without relying on strong external teacher models or carefully curated datasets. It treats flawed reasoning processes as opportunities for recovery, enabling more scalable learning that is independent of external resources.

On competitive math and general reasoning benchmarks, DenoiseRL consistently outperformed strong on-policy RL baselines. The results also confirm a trend in which self-correction behavior strengthens as training difficulty increases, making it a noteworthy scalable approach for enhancing LLM reasoning capabilities.

Source**: [DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes](https://huggingface.co/papers/2605.28421) (Hugging Face Daily Papers)

NVIDIA LocateAnything — High-Accuracy Grounding at 10x Qwen3-VL Speed

NVIDIA has released LocateAnything, a fast, high-quality Vision-Language Grounding model built on Parallel Box Decoding that achieves roughly 10x speedup compared to Qwen3-VL.

The 3B-parameter model is available on Hugging Face, along with a demo. Source code is provided in the NVlabs Eagle repository.

Source**: [NVIDIA LocateAnything](https://huggingface.co/nvidia/LocateAnything-3B) (Reddit LocalLLaMA)

OpenBMB Releases UltraData-SFT-2605 — High-Quality SFT Data with Over 15 Million Samples

OpenBMB has released UltraData-SFT-2605, the core SFT dataset used for post-training MiniCPM5-1B-SFT. It covers seven domains including math, code, knowledge, and instruction following, and contains a total of more than 15 million samples, including both Deep Thinking and Non-thinking samples.

Its standout feature is a six-stage high-quality SFT data curation pipeline. Rigorous quality control covers everything from query filtering and answer quality verification to per-sample validation and benchmark decontamination. Released under the Apache 2.0 license, it can be used for training small models and researching post-training methods.

Source**: [openbmb/UltraData-SFT-2605](https://huggingface.co/datasets/openbmb/UltraData-SFT-2605) (Hugging Face Trending Datasets)

B³D-RWKV — A Triplet-Block Design Enabling Bidirectional Diffusion on RWKV

Diffusion models require bidirectional attention, but efficient linear models like RWKV are constrained to causal (unidirectional) computation. B³D-RWKV's Triplet-Block layout resolves this fundamental conflict.

It enables bidirectional discrete diffusion to run in parallel while preserving RWKV's O(L) inference efficiency. At 7.2B parameters, it achieves roughly 1.6x higher decoding throughput over existing baselines while maintaining competitive accuracy.

Source**: [Triplet-Block Diffusion RWKV](https://huggingface.co/papers/2605.25969) (Hugging Face Daily Papers)

Fast-dDrive — Block-Diffusion VLA for Autonomous Driving

In end-to-end approaches to autonomous driving, the key challenge is achieving both precise trajectory planning and efficient inference. Autoregressive models face memory bandwidth constraints on edge hardware, while diffusion models make KV cache reuse difficult.

Fast-dDrive is a Block-Diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across units. It achieves SOTA ADE@3s and ADE@5s on the WOD-E2E test set. On nuScenes, it reduces mean L2 error to 0.32m, a 22% improvement. Integration with SGLang delivers 12x the throughput of AR baselines.

Source**: [Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving](https://huggingface.co/papers/2605.23163) (Hugging Face Daily Papers)

Optimizing LLM Post-Training Data Engineering with SAEs

A new method leverages the internal representations of Sparse Autoencoders (SAEs) to predict data diversity, difficulty, and quality in LLM post-training. This enables more systematic guidance for data selection in RL post-training.

Source**: [Guiding LLM Post-training Data Engineering with Model Internals from SAE](https://huggingface.co/papers/2605.27354) (Hugging Face Daily Papers)

Practical KV Cache Quantization Benchmarks — q5/q6 Emerge as the Optimal Choice

A detailed benchmark of 38 quantization pairs using Qwen 3.6 27B has been published. Key findings: q5_0 and q5_1 KV caches are underrated, while q8_0/q4_* pairs are not as strong as their community reputation suggests.

The practical recommendation is a tiered approach: `q8_0/q6_0` or `q8_0/q5_1` for high quality, `q5_0/q5_0` or `q5_0/q4_1` when VRAM is tight, and `q4_0/q4_0` only as a last resort.

Source**: [KV cache quant benchmarks](https://www.reddit.com/r/LocalLLaMA/comments/1tp9d1w/kv_cache_quant_benchmarks_q5_q6_are_underrated/) (Reddit LocalLLaMA)

---