meta-llama/Llama-3.2-3B

In the model registry

Timeline

1. Oct 16, 2026 · Research paper · 1 source
   Reusing Calibration Gains across Model Representations
   Scale calibration learns an update for one model representation.
2. Oct 15, 2026 · Research paper · 1 source
   WaveLLM: Wavelet-Domain LLM Entropy Modeling for Lossless Image Compression
   Motivated by this insight, we propose WaveLLM, a wavelet-domain LLM-based entropy model for lossless image compression.
3. Oct 14, 2026 · Opinion / analysis · 1 source
   Moving a running model's layers between GPUs mid-reply (8 ms, no re-prefill)
   In the video, a cloud VM with T4s: mid-reply, all layers of Llama-3.2-3B move from one GPU to the other and back.
4. Oct 14, 2026 · Research paper · 1 source
   INTACT: Inducing Latent Reasoning on a Frozen Backbone
   We use a small external encoder to induce latent reasoning in a frozen backbone model.
5. Oct 13, 2026 · Research paper · 1 source
   Two For The Price Of One: Curvature-Aware Pruning Preserves Influence Functions Too
   Post-training pruning enables efficient deployment, but conventional methods do not explicitly preserve internal-gradient alignment between sparse and dense models.
6. Oct 13, 2026 · Research paper · 1 source
   inLoRA-MINT: Internal Representations for Auditing Training Data in Domain-adapted LLMs
   This work investigates the application of the Membership Inference Test (MINT) to large language models (LLMs) fine-tuned with LoRA, aiming to determine whether specific samples were used during training and address concerns about data privacy, leakage, and transparency.
7. Oct 11, 2026 · Research paper · 1 source
   Who Sets the Pace of Judge Hacking? Judge Capability versus the Policy-Judge Gap
   When a policy is trained with GRPO against a frozen, reference-free LLM judge, the judge accepts more of its answers, but in a post-hoc analysis we detect no gain in held-out greedy accuracy: in all 12 (policy, judge) cells, the seed-mean gain never exceeded 0.04.

Often appears with