arXiv subjects

Song Cao

Publications and source records attributed to Song Cao.

10 recordsLinked to original sources

arXiv2026

Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA

Macaron-V1 is an open agent-model family for experiential intelligence: learning from experience in real environments and continuing to learn after deployment. It is organized around two system goals. Adaptation is pursued through recursive improvement of versioned model-harness pairs, where experience from one configuration is evaluated under an external contract and used to construct its successor. Collaboration is pursued via the Mixture-of-LoRA (MoL) architecture that freezes a base model, composes specialist LoRA adapters, and selects one LoRA per user turn. The flagship Macaron-V1-Venti (748B) combines a 744B GLM-5.2 base with four LoRAs for chat, agent, coding, and GenUI; the Qwen3.6-35B-based Macaron-V1-Tall (50B) uses the same design for local deployment. This report presents Macaron-V1 as a co-designed system spanning architecture, algorithms, and infrastructure. The MoL architecture supports continual learning through extensible LoRA specialists. The algorithm combines Model-Harness Co-design and recursive self-improvement loop, including the UI4A component-native GenUI harness, a stateful action substrate, versioned Harness Context Protocol contract, and the agentic RL framework MindForge. The supporting infrastructure includes the post-training platform MinT, the long-context RL method LongStraw, and stability techniques for sparse MoE and DSA base models. We evaluate Macaron-V1 on Personal Intelligence, GenUI, and general capability benchmarks against frontier baselines. Our results validate the current system, while compounding gains from continual learning and collective intelligence remain open questions.

cs.LG↗

arXiv2026

On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters

Parameter-efficient fine-tuning (PEFT) is usually treated as a cheaper alternative to full fine-tuning. We study a broader role: small trainable adapters as persistent local state on top of strong shared foundation models. In this framing, the base model provides shared competence while adapters carry instance-specific behavior such as preferences, skills, tool habits, and memory-like updates. We organize the problem around three scaling axes: Scale Up, where stronger shared priors make small local updates more useful; Scale Down, where we study how small adapters can be while remaining reliable; and Scale Out, where many persistent adapted instances coexist. MinT provides one infrastructure example for managing adapter identity, revision, provenance, evaluation, and serving residency. Together, the results suggest that PEFT can be a compact substrate for persistent personal models rather than only a budget substitute for full fine-tuning.

cs.LG↗

arXiv2026

Macaron-A2UI: A Model for Generative UI in Personal Agents

As personal agents evolve to handle complex, user-centric tasks, static plain-text chat is rapidly becoming a bottleneck. Generative UI emerges as the necessary new interface layer, dynamically synthesizing the right controls, options, and state from the interaction context in real time. We present Macaron-A2UI, a model for Generative UI in personal agents. Our goal is to move beyond text-only interaction by enabling agents to generate natural language together with lightweight, executable UI actions for information collection, preference refinement, confirmation, and multi-goal organization. We build a large-scale Generative UI corpus from heterogeneous dialogue sources, introduce A2UI-Bench for controlled evaluation, and train 30B, 235B and 754B models with parameter-efficient LoRA-based supervised fine-tuning followed by reward-driven reinforcement learning. The best Macaron-A2UI model reaches 75.6 overall on A2UI-Bench without explicit schema hints, surpassing the strongest full-schema frontier baseline. We release the models, benchmark, and evaluation protocol to support future work on Generative UI for personal agents.

cs.HC↗

arXiv2026

MinT: Managed Infrastructure for Training and Serving Millions of LLMs

We present MindLab Toolkit (MinT), a managed infrastructure system for Low-Rank Adaptation (LoRA) post-training and online serving. MinT targets a setting where many trained policies are produced over a small number of expensive base-model deployments. Instead of materializing each policy as a merged full checkpoint, MinT keeps the base model resident and moves exported LoRA adapter revisions through rollout, update, export, evaluation, serving, and rollback, hiding distributed training, serving, scheduling, and data movement behind a service interface. MinT scales this path along three axes. Scale Up extends LoRA RL to frontier-scale dense and MoE architectures, including MLA and DSA attention paths, with training and serving validated beyond 1T total parameters. Scale Down moves only the exported LoRA adapter, which can be under 1% of base-model size in rank-1 settings; adapter-only handoff reduces the measured step by 18.3x on a 4B dense model and 2.85x on a 30B MoE, while concurrent multi-policy GRPO shortens wall time by 1.77x and 1.45x without raising peak memory. Scale Out separates durable policy addressability from CPU/GPU working sets: a tensor-parallel deployment supports 10^6-scale addressable catalogs (measured single-engine sweeps through 100K) and thousand-adapter active waves at cluster scale, with cold loading treated as scheduled service work and packed MoE LoRA tensors improving live engine loading by 8.5-8.7x. MinT thus manages million-scale LoRA policy catalogs while training and serving selected adapter revisions over shared 1T-class base models.

cs.LG↗

arXiv2026

Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution

Multimodal Large Language Model (MLLM) agents facilitate Graphical User Interface (GUI) automation but struggle with long-horizon, cross-application tasks due to limited context windows. While memory systems provide a viable solution, existing paradigms struggle to adapt to dynamic GUI environments, suffering from a granularity mismatch between high-level intent and low-level execution, and context pollution where the static accumulation of outdated experiences drives agents into hallucination. To address these bottlenecks, we propose the Darwinian Memory System (DMS), a self-evolving architecture that constructs memory as a dynamic ecosystem governed by the law of survival of the fittest. DMS decomposes complex trajectories into independent, reusable units for compositional flexibility, and implements Utility-driven Natural Selection to track survival value, actively pruning suboptimal paths and inhibiting high-risk plans. This evolutionary pressure compels the agent to derive superior strategies. Extensive experiments on real-world multi-app benchmarks validate that DMS boosts general-purpose MLLMs without training costs or architectural overhead, achieving average gains of 18.0% in success rate and 33.9% in execution stability, while reducing task latency, establishing it as an effective self-evolving memory system for GUI tasks.

cs.AI↗

arXiv2025

Optimal Allocations under Strongly Pigou-Dalton Criteria: Hidden Layer Structure & Efficient Combinatorial Approach

We investigate optimal social welfare allocations of $m$ items to $n$ agents with binary additive or submodular valuations. For binary additive valuations, we prove that the set of optimal allocations coincides with the set of so-called \emph{stable allocations}, as long as the employed criterion for evaluating social welfare is strongly Pigou-Dalton (SPD) and symmetric. Many common criteria are SPD and symmetric, such as Nash social welfare, leximax, leximin, Gini index, entropy, and envy sum. We also design efficient algorithms for finding a stable allocation, including an $O(m^2n)$ time algorithm for the case of indivisible items, and an $O(m^2n^5)$ time one for the case of divisible items. The first is faster than the existing algorithms or has a simpler analysis. The latter is the first combinatorial algorithm for that problem. It utilizes a hidden layer partition of items and agents admitted by all stable allocations, and cleverly reduces the case of divisible items to the case of indivisible items.In addition, we show that the profiles of different optimal allocations have a small Chebyshev distance, which is 0 for the case of divisible items under binary additive valuations, and is at most 1 for the case of indivisible items under binary submodular valuations.

cs.GT↗

arXiv2025

D-Artemis: A Deliberative Cognitive Framework for Mobile GUI Multi-Agents

Graphical User Interface (GUI) agents aim to automate a wide spectrum of human tasks by emulating user interaction. Despite rapid advancements, current approaches are hindered by several critical challenges: data bottleneck in end-to-end training, high cost of delayed error detection, and risk of contradictory guidance. Inspired by the human cognitive loop of Thinking, Alignment, and Reflection, we present D-Artemis -- a novel deliberative framework in this paper. D-Artemis leverages a fine-grained, app-specific tip retrieval mechanism to inform its decision-making process. It also employs a proactive Pre-execution Alignment stage, where Thought-Action Consistency (TAC) Check module and Action Correction Agent (ACA) work in concert to mitigate the risk of execution failures. A post-execution Status Reflection Agent (SRA) completes the cognitive loop, enabling strategic learning from experience. Crucially, D-Artemis enhances the capabilities of general-purpose Multimodal large language models (MLLMs) for GUI tasks without the need for training on complex trajectory datasets, demonstrating strong generalization. D-Artemis establishes new state-of-the-art (SOTA) results across both major benchmarks, achieving a 75.8% success rate on AndroidWorld and 96.8% on ScreenSpot-V2. Extensive ablation studies further demonstrate the significant contribution of each component to the framework.

cs.AI↗

arXiv2025

DPGP: A Hybrid 2D-3D Dual Path Potential Ghost Probe Zone Prediction Framework for Safe Autonomous Driving

Modern robots must coexist with humans in dense urban environments. A key challenge is the ghost probe problem, where pedestrians or objects unexpectedly rush into traffic paths. This issue affects both autonomous vehicles and human drivers. Existing works propose vehicle-to-everything (V2X) strategies and non-line-of-sight (NLOS) imaging for ghost probe zone detection. However, most require high computational power or specialized hardware, limiting real-world feasibility. Additionally, many methods do not explicitly address this issue. To tackle this, we propose DPGP, a hybrid 2D-3D fusion framework for ghost probe zone prediction using only a monocular camera during training and inference. With unsupervised depth prediction, we observe ghost probe zones align with depth discontinuities, but different depth representations offer varying robustness. To exploit this, we fuse multiple feature embeddings to improve prediction. To validate our approach, we created a 12K-image dataset annotated with ghost probe zones, carefully sourced and cross-checked for accuracy. Experimental results show our framework outperforms existing methods while remaining cost-effective. To our knowledge, this is the first work extending ghost probe zone prediction beyond vehicles, addressing diverse non-vehicle objects. We will open-source our code and dataset for community benefit.

cs.RO↗

arXiv2025

Discrete Effort Distribution via Regret-enabled Greedy Algorithm

This paper addresses resource allocation problem with a separable objective function under a single linear constraint, formulated as maximizing $\sum_{j=1}^{n}R_j(x_j)$ subject to $\sum_{j=1}^{n}x_j=k$ and $x_j\in\{0,\dots,m\}$. While classical dynamic programming approach solves this problem in $O(n^2m^2)$ time, we propose a regret-enabled greedy algorithm that achieves $O(n\log n)$ time when $m=O(1)$. The algorithm significantly outperforms traditional dynamic programming for small $m$. Our algorithm actually solves the problem for all $k~(0\leq k\leq nm)$ in the mentioned time.

cs.DS↗

arXiv2024

Scheduling two types of jobs with minimum makespan

We consider scheduling two types of jobs (A-job and B-job) to $p$ machines and minimizing their makespan. A group of same type of jobs processed consecutively by a machine is called a batch. For machine $v$, processing $x$ A-jobs in a batch takes $k^A_vx^2$ time units for a given speed $k^A_v$, and processing $x$ B-jobs in a batch takes $k^B_vx^2$ time units for a given speed $k^B_v$. We give an $O(n^2p\log(n))$ algorithm based on dynamic programming and binary search for solving this problem, where $n$ denotes the maximal number of A-jobs and B-jobs to be distributed to the machines. Our algorithm also fits the easier linear case where each batch of length $x$ of $A$-jobs takes $k^A_v x$ time units and each batch of length $x$ of $B$-jobs takes $k^B_vx$ time units. The running time is the same as the above case.

cs.DS↗