A focused feed of models, agents, research and open-source releases for people building with AI.
Microsoft's LLMLingua-2 BERT compressor crosses 800K+ downloads, trending on Hugging Face. Reframes prompt compression as token classification instead of entropy scoring, guaranteeing extractive faithfulness.
Rikiya Takehi released an edited Qwen3.5-0.8B on Hugging Face, modified by GPT-6 Astra without any training. Astra changed 0.016% of weights (131,072 values) across two MLP down-projection tensors in layers 15 and 19.
New framework reformulates RAG grounding as case + claim + evidence support verification. Builds training data with counterfactual wrong-state and topic-related negatives automatically. Verifier hits 97.4 AUROC on MIMIC-CXR , collapses to 55 under swapped evidence.
Value Forcing trains flow-matching critics by enforcing Bellman consistency at every noise level, not just at the endpoint. A new flow Bellman equation adds a Gaussian correction term so the equation holds from pure noise to clean returns.
Anthropic automated red-teaming by having Claude agents search for training data that evades monitors but induces misalignment. Successful datasets hid misaligned behavior in roleplay and creative-writing frames like beekeeping Q&A and wrestling kayfabe.
JEV-9B is an open-weights 9B model combining calibrated System 1 decisions and System 2 reasoning in one engine, via Hugging Face Mean KL divergence of about 0.019 nats from closed TypeSafe Jev 1.13's output distributions on 25,376 held-out questions Trained with Blocks of Experts: 40.2M-parameter
New preprint decomposes task vectors into Task Recognition, Task Learning, and residual components. Decomposition uses candidate-label unembeddings as the basis, keeping it linear and cheap. Bayesian optimization over two coefficients per layer beats the original task vector.
BYOD converts frozen AR models into masked diffusion LMs using LoRA only, in 3-4 GPU hours. Works on Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B with one unified recipe. Adapters are 4.1-5.8% of parameters; training uses at most 102.4M tokens over 25k steps.
Round-to-nearest fp32 to bf16 conversion silently discards most RL updates; only 0.4 to 0.5% of weights change per step. The served rollout policy lags the fp32 optimizer by 50 to 150 steps, invisible to existing mismatch fixes.
Anthropic introduced self-assessment , where Claude grades the safety of each conversation right after responding. Flags 78% of harmful conversations and 94% of high-priority ones at 1% false-positive rate on Fable 5.
Kyutai shows Neural Thickets gains come from logit diversity, not weight-space experts. Logit PSV adds one random bias vector over the vocabulary, model stays frozen. Beats full-weight PSV on 13 of 15 reasoning benchmarks across four tested LLMs.
CMU-led team unifies pretraining, SFT, and RL into one maximum-likelihood objective called unified training. RL rollouts reframed as likelihood fitting to model-generated targets via MaxRL extension . Single trainer mixes static text and RL generations; staged training becomes a special case.
New paper shows flow robot policies produce near-Gaussian per-step action distributions, with under 2% Wasserstein error. Gaussian Action Readout replaces the flow policy's final layer with a Gaussian head trained on 0.02% of states.
A new survey reframes foundation model adaptation for medical imaging as a risk-control problem, mapping five mechanisms to clinical deployment trade-offs.
Google Research proposes a biologically inspired sleep phase that lets LLMs consolidate new facts into weights without forgetting, scaling context to 10M tokens.
Reflection AI reveals the raw infrastructure behind Beam's reinforcement learning run and commits to OpenEnv as its future environment standard.
A new training recipe cuts hallucinations in multimodal reasoning models by trimming redundant thinking tokens and contrasting against deliberately induced errors.
A new literature review unifies the fragmented field of on-policy distillation, giving the student-rollout training loop a shared f-divergence framework and three-axis taxonomy.
A new training objective called MCLR reproduces classifier-free guidance effects at training time, revealing CFG as implicit contrastive alignment.
A new framework lets language model agents master dynamic network control by reflecting on their own experience and self-finetuning, no reward engineering required.
A self-supervised method treats multiple X-ray projections of the same patient as paired data, beating supervised and vision-language baselines on three datasets.
A new self-training framework called OPST lets language models improve themselves using only topic names, with no human-written questions, answers, or reasoning traces.
FMER swaps diffusion for flow matching in online RL, delivering tractable entropy, mirror descent updates, and up to 7x faster training.
Two overlooked floating-point bugs in softmax and cross-entropy have been quietly corrupting gradients. A new paper fixes them in a few lines of JAX.
No stories match these filters.
