Career moves
No published career moves.
Papers
Paper2026-10-07 · arXiv (Cornell University)
Long-WAM: Scaling the Context of World-Action Models
- What
- Researchers have developed Long-WAM, a model that efficiently utilizes past visual information of robots under real-time constraints.
- Key result
- When pre-training the foundational model that processes visual information using an autoregressive (AR) method, the utilization of past memory is maximized, and robot task success rates are improved.
- Why it matters
- It demonstrated that real-time operation is possible in various hardware environments, which can contribute to future complex long-horizon robot control tasks.AI summary
Paper2026-10-06 · arXiv (Cornell University)
Humanize: Judgement Engineering for Agentic Coding
- What
- Introduced Humanize, a multi-AI collaborative workflow that applies clear, mechanical decisions between the planning, implementation, and review stages.
- Key result
- Production AIs and review AIs from different providers worked in alternating turns, achieving high performance in competitive problem-solving and performance evaluation.
- Why it matters
- Demonstrates that independent reviews help catch errors in AI-assisted programming processes.AI summary
Paper2026-09-29 · arXiv (Cornell University)
SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video
- What
- The researchers proposed SoL-Refiner, a refiner model that transforms low-resolution videos into high-resolution ones through a single denoising step.
- Key result
- At 2K resolution, it achieved higher average scores on VBench and UniPercept than comparable external refiners. At 3840x2176 resolution, it improved upon both metrics compared to the three-step LTX-2.3 refiner. It was reported that under 2K latency settings, refining latency became 8.91 times faster compared to the same baseline model.
- Why it matters
- High-resolution video generation significantly increases computational load and cost burdens. This approach shows that increasing quality in a single step instead of multiple steps can be a direction to reduce this burden. However, since the results presented in the abstract were measured under conditions set by the authors, they cannot be generalized as overall effects in actual services.AI summary
Paper2026-09-29 · arXiv (Cornell University)
LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization
- What
- The paper proposes LeapQuant, a training-free method that compresses the recurrent states, which large language models repeatedly read and write when processing long contexts, into low precision.
- Key result
- According to the abstract, 8-bit compression maintains accuracy comparable to the FP32 baseline while achieving an average speedup of 2.05 to 3.70 times at the GPU kernel level and 1.47 times in overall inference.
- Why it matters
- This research can contribute to improving operational efficiency for AI services handling long contexts by reducing memory and computational burdens during inference. However, since the abstract only presents results from specific models and GPUs, it is difficult to conclude that the same effects will appear across all actual commercial services.AI summary
Paper2026-09-28 · arXiv (Cornell University)
Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge
- What
- Created a complete execution pipeline to run the 33-billion parameter video generation AI model MiniMax-H3 faster and with less memory on both cloud servers and small devices.
- Key result
- According to the abstract, up to a 30x end-to-end speedup and about 20% memory reduction were achieved, generating a 5-second, 1344x768 resolution video with audio 3.5x faster than real-time on 8 GB200 server nodes.
- Why it matters
- This is significant as it demonstrates a way to lower the generation latency and memory limits that are the biggest obstacles when applying video generation AI to real services, though the commercialization effect and scale of cost reduction are difficult to judge from the abstract alone.AI summary
Paper2026-09-28 · arXiv (Cornell University)
d-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models
- What
- The authors propose d-OPD, a distillation training method that converts conventional autoregressive language models (which generate one token at a time) into block diffusion language models that generate multiple tokens in parallel.
- Key result
- Experiments with Qwen3 models (0.6B to 8B) showed that d-OPD improved the average score across six benchmarks by up to 4.0 points compared to the existing OPDLM method, while reducing training time by 1.35 to 1.58 times.
- Why it matters
- The abstract suggests that generating multiple tokens simultaneously in blocks has the potential to increase text generation speed, and this study demonstrates that the efficiency of the training process for creating such models can be improved.AI summary
