MintAct is a family of vision-language models (2B/4B/8B) that unifies UI grounding, multi-step navigation across mobile, desktop, and web, and…
The paper introduces RRSI, a method for automatically refining LLM agent harnesses while mitigating overfitting during recursive self-improvement. It constrains…
The paper presents D-RAC, a technique that normalizes heterogeneous enterprise documents into PDF before converting them into retrieval-optimized Markdown using…
The paper addresses uneven performance across task categories in software-engineering agents by training category-specific expert models via reinforcement learning with…
The paper introduces Harness-Zero, a method for transferring behaviors learned with an optimized external agent harness into a model’s own…
The paper introduces Jev-Mem, an agentic memory architecture that splits control into a lightweight ‘System-One’ controller (fast query routing, retrieval-budget…
The paper analyzes gradient estimation noise in sparse on-policy distillation and proposes an ‘information-efficiency ratio’ (IER) metric, based on signal-to-noise…
The paper introduces Complex KDA (CKDA), an enhanced Kimi Delta Attention variant for linear RNNs that extends parameter ranges (gates…
The paper presents the Functionalizer, a pre-tokenization framework that decomposes orthographic variations (capitalization, diacritics, etc.) into reversible opcode/operand pairs encoded…
The paper introduces EdgeGen, a framework that automatically generates synthetic edge-case tasks by extracting compliance rules from agent specifications and…
The paper identifies that FP8 quantization noise in RL pipelines distorts importance ratios, causing training instability. It proposes ‘Calibrated Clipping,’…
Researchers introduce Taste-Bench, a benchmark that measures an LLM agent’s “taste” — its ability to make good long-horizon decisions at…
