Research
Solving Long-Form Coherence in Small Open-Weight LLMs
An analysis of the POLARIS paper and its approach to preventing quality degradation and structural collapse in long-form creative writing for small models.
Research
Research
Papers that actually matter
49 articles in this section.
Research
AURA: Solving the KV Cache Problem for Continuous Embodied AI
AURA introduces action-gated memory to prevent VRAM bloat in robots, allowing long-term policies to run indefinitely without crashing or hallucinating.
Research
Reducing LLM Long-Context Latency with Adaptive Runtime Termination
Explore how Adaptive Runtime Termination (ART) reduces memory bandwidth bottlenecks to improve token throughput during long-context LLM inference.
Research
BitsMoE: Reducing VRAM Requirements for Mixture-of-Experts Models
BitsMoE uses spectral energy to guide non-uniform bit allocation, potentially allowing massive MoE models to fit on consumer GPUs.
Research
EAGLE 3.1: Fixing Attention Drift in Speculative Decoding
EAGLE 3.1 addresses attention drift to provide more consistent and predictable throughput for LLM inference via speculative decoding.
Research
Together AI’s OSCAR: 2-Bit KV Cache Quantization for Long Context
Together AI’s OSCAR system uses attention-aware rotation to compress KV caches to 2-bit, significantly expanding context windows on consumer GPUs.
Research
ByteDance Research: QA-Centric Training Improves LMM Document Analysis
A ByteDance study suggests that training multimodal models via question-answering outperforms transcription-heavy methods for analyzing long, complex documents.
Research
Recurrent Depth in Transformers: Balancing Compute and Memory Efficiency
An analysis of recurrent depth and Sparse MoE as a way to trade memory efficiency for gradient stability in transformer architectures.
Research
Multi-Pass Prompt Verification: Addressing Qualitative Loss in Quantized LLMs
A new study explores using multi-pass verification to recover accuracy lost in 2-bit and 3-bit quantized models, though critics argue it’s a workaround.
Research
OpenAI Model Disproves Discrete Geometry Conjecture via Counterexample Search
An OpenAI model has disproven a geometry conjecture, highlighting the shift from human intuition to high-speed automated counterexample searching in mathematics.