Papers1055

Beyond Imitation: Reinforcement Learning for Active Latent Planning

The paper shows how to make AI think faster and smarter by planning in a hidden space instead of writing long step-by-step sentences.

#latent reasoning#chain-of-thought#variational autoencoder

Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening

Intermediate

Xiaotong Ji, Rasul Tutunov et al.Jan 29arXiv

The paper shows a fast, training-free way to boost an LLM’s step-by-step reasoning by smartly reusing the model’s own probabilities.

#power distribution sampling#distribution sharpening#low-temperature sampling

KromHC: Manifold-Constrained Hyper-Connections with Kronecker-Product Residual Matrices

Intermediate

Wuyang Zhou, Yuxuan Gu et al.Jan 29arXiv

Hyper-Connections (HC) make the usual single shortcut in neural networks wider by creating several parallel streams and letting the model mix them, but this can become unstable when stacked deep.

#Hyper-Connections#Manifold-Constrained Hyper-Connections#Doubly Stochastic Matrix

Shaping capabilities with token-level data filtering

Intermediate

Neil Rathi, Alec RadfordJan 29arXiv

The paper shows a simple way to teach AI models what not to learn by removing only the exact words (tokens) related to unwanted topics during pretraining.

#token-level data filtering#capability shaping#sparse autoencoders

ASTRA: Automated Synthesis of agentic Trajectories and Reinforcement Arenas

Intermediate

Xiaoyu Tian, Haotian Wang et al.Jan 29arXiv

ASTRA is a fully automated way to train tool-using AI agents by making both their practice stories (trajectories) and their practice worlds (environments) without humans in the loop.

#tool-augmented agents#multi-turn decision making#verifiable environments

MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning

Intermediate

Yaorui Shi, Shugui Liu et al.Jan 29arXiv

MemOCR is a new way for AI to remember long histories by turning important notes into a picture with big, bold parts for key facts and tiny parts for details.

#MemOCR#visual memory#adaptive information density

Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization

Intermediate

Jiecong Wang, Hao Peng et al.Jan 29arXiv

This paper introduces PLaT, a way for AI to think silently in a hidden space (the brain) and only speak when needed (the mouth).

#latent chain-of-thought#planning in latent space#planner-decoder architecture

Self-Improving Pretraining: using post-trained models to pretrain better models

Intermediate

Ellen Xiaoqing Tan, Shehzaad Dhuliawala et al.Jan 29arXiv

This paper teaches language models to be safer, more factual, and higher quality during pretraining, not just after, by using reinforcement learning with a stronger model as a helper.

#self-improving pretraining#reinforcement learning#online DPO

Qwen3-ASR Technical Report

Intermediate

Xian Shi, Xiong Wang et al.Jan 29arXiv

Qwen3‑ASR is a family of speech models that hear, understand, and write down speech in 52 languages and dialects, plus they can tell you when each word was spoken.

#ASR#forced alignment#timestamps

Grounding and Enhancing Informativeness and Utility in Dataset Distillation

Intermediate

Shaobo Wang, Yantai Yang et al.Jan 29arXiv

This paper tackles dataset distillation by giving a clear, math-backed way to keep only the most useful bits of data, so models can learn well from far fewer images.

#dataset distillation#data condensation#Shapley value

Scaling Embeddings Outperforms Scaling Experts in Language Models

Intermediate

Hong Liu, Jiaqi Zhang et al.Jan 29arXiv

The paper shows that growing the embedding part of a language model (especially with n-grams) can beat adding more MoE experts once you pass a certain sparsity 'sweet spot.'

#N-gram Embedding#Mixture-of-Experts (MoE)#Embedding Scaling

Do Reasoning Models Enhance Embedding Models?

Intermediate

Wun Yu Chan, Shaojin Chen et al.Jan 29arXiv

The paper asks a simple question: if a language model becomes better at step-by-step reasoning (using RLVR), do its text embeddings also get better? The short answer is no.

#text embeddings#RLVR#contrastive learning

35 36 37 38 39