Search posts, papers, and topics
Jevlike is an open‑source starter model that scores a list of text options in a single forward pass. It provides a minimal architecture (option queries, shared dot‑product scorer), synthetic data generation, training/evaluation CLI, and examples on Doom and chess. The repo supports a byte‑level encoder or a frozen Hugging‑Face encoder (e.g., Qwen2.5‑0.5B), runs on CPU/MPS/CUDA, and reports benchm…
The paper presents Infinite-Parameter LLMs, where a compact hypernetwork creates feed‑forward weights from live user data and updates a Bayesian latent code online, keeping the stored model size constant while effectively having infinite parameters. This design aims to improve over standard in‑context learning and retrieval by persisting knowledge in weights and freeing context space.
On-policy distillation (OPD) can lead to excessively long student responses, a phenomenon called length inflation. This paper identifies "termination-token mismatch" between base students and post-trained teachers as a key source, where models place stopping probability on different EOS tokens. Treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigate…
FAMOS is a feed-forward model that predicts 3D articulation, including movable-part segmentation and joint parameters, from sparse, unordered point clouds. It leverages multiple observations through a Multi-state Articulation Transformer and a novel observed articulation span objective, outperforming existing baselines.
When2Think introduces a post‑training framework that lets a large reasoning model decide per‑instance how much reasoning depth to allocate, using difficulty‑aware reward shaping (IDAC) and verifier rewards. It cuts token usage by ~28% while boosting Pass@3 by 10% on AIME24 and reaches 40% Pass@3 on AIME25, outperforming compression and routing baselines.
VC-Attention introduces a training‑free low‑bit attention pipeline for diffusion transformers. It smooths value tensors via lightweight online clustering (V‑Smooth) and quantizes only the residual after subtracting block means, restoring the mean from the softmax row sum. It also replaces the FP32 softmax exponential with a fused FP8 cast (ExpCast‑FP8) that maps log‑scores directly to E4M3 probab…
The paper decomposes transformer representation updates into parallel and perpendicular components, showing that parallel updates beyond the residual path are significant. Suppressing parallel updates during pretraining improves validation loss and downstream metrics.
RelateAnything is a 53 M‑parameter model that predicts arbitrary textual relations between any supplied image regions, using a text‑embedding bank instead of a fixed classifier. Trained on a new 4.3 M‑relation dataset, it outperforms prior open‑vocab methods by 2.3‑3.5× mean recall while running in ~20 ms per frame.
The authors cast transformer block removal as a constrained binary optimization problem equivalent to an Ising glass, using a Hessian‑derived energy as a proxy for downstream quality. Solving the resulting QUBO with classical or quantum‑inspired solvers yields up to 23 MMLU points improvement over prior block‑removal baselines at 50 % depth compression.
The IBM research blog explains two new theoretical results that prove shallow constant‑depth quantum circuits can outperform decoder‑only transformers on a functional task (iterated index) and diffusion language models on a sampling task (parity‑sampling). The proofs give asymptotic separations but are not yet practical.
The paper presents TREVIS, a method that encodes decision trees into a continuous latent space using a Tree Transformer VAE, allowing gradient-based optimization of both accuracy and structural sparsity. Experiments claim TREVIS matches the predictive performance of near-optimal algorithms while producing sparser trees.
