AI / ML

Kernel Forge is an agentic system that leverages large language models to generate and optimize low-level CUDA kernels, targeting compute-intensive operations like matrix multiplication and convolution. Unlike previous tools that produce isolated code snippets or rely on random tensor inputs, this framework aims to streamline the optimization workflow with reduced human intervention. The research addresses current limitations in existing LLM-based optimizers, which often lack robust debugging capabilities and integration support for broader model architectures.

• Automates CUDA kernel optimization for ML runtime bottlenecks using LLM agents.
• Targets common compute kernels: matmul, convolution, and normalization.
• Aims to reduce reliance on expert engineers for hand-written GPU code.
• Addresses gaps in existing tools regarding debugging and model integration.
• Moves beyond isolated kernel generation toward more practical deployment workflows.

The article details how Zig manages incremental compilation by tracking file dependencies and caching build outputs. It explains the hashing mechanisms used to detect changes and avoid redundant work. The discussion covers how the build system integrates with these internals to optimize compilation times for large projects.

• Zig uses content hashing to detect source changes and invalidate caches accurately.
• Incremental builds skip recompilation when dependencies and inputs remain unchanged.
• The build system integrates directly with compiler internals for efficient dependency tracking.
• Optimized caching reduces build times significantly for large-scale Zig projects.

RoCo-ACE introduces a rollout-conditioned online distillation method to inject new knowledge into pretrained multimodal large language models while minimizing behavioral drift. It employs a likelihood contrast mechanism that reallocates distillation weight specifically to rollout tokens supported by the reference data. This approach avoids the coarse supervision of uniform reference-conditioned distillation by directly reinforcing reference-aligned outputs.

• Targets drift in updated MLLMs by focusing supervision on reference-supported tokens.
• Uses likelihood contrast to weight distillation, improving over uniform reference methods.
• Online distillation leverages model-generated rollouts rather than static datasets.
• New method balances factual injection with retention of prior non-updated behaviors.

TabRank introduces a framework for training reasoning-based re-rankers specifically for tabular retrieval tasks. It leverages Large Reasoning Models with explicit chain-of-thought capabilities to improve ranking quality over conventional sparse or dense retrievers. The approach aims to refine candidate lists generated by first-stage systems using enhanced semantic understanding and reasoning.

• Focuses on table retrieval, a key structured information task.
• Uses Large Reasoning Models for explicit chain-of-thought ranking.
• Outperforms conventional sparse and dense retrieval models.
• Distills reasoning capabilities into specialized re-rankers.

Temporal-Distance JEPA improves latent model predictive control by mining directed temporal costs from reward-free demonstration logs. Unlike prior methods that rely on latent Euclidean distance as a proxy for goal progress, this approach explicitly optimizes multi-step ranking. It retains the LeWM encoder-predictor backbone while replacing heuristic geometry with mined temporal progression metrics.

• Replaces latent Euclidean distance with mined temporal costs for planning.
• Enables better multi-step ranking of imagined futures from offline logs.
• Retains LeWM backbone while optimizing for goal progress directly.
• Improves reward-free planning by aligning representation with temporal flow.