As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world. Our code is publicly available at https://github.com/A-EVO-Lab/a-evolve.
LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.
AI 速览：该研究探讨将语言模型用作研究世界模型（RWM），以预测候选实验干预的结果，并利用真实实验经验提升预测与实验选择效率。
AI research agents automate the cycle of proposing, implementing, and evaluating experiments, opening a path toward recursive self-improvement. Yet their ability to propose experiments outpaces their capacity to execute them in real environments, making outcome prediction a key capability for sustained self-improvement under limited experimental budgets. We investigate language models as Research World Models (RWMs), which predict the outcomes of candidate interventions across research environments. Our evaluation draws on over 2,600 experimental records from nine research environments spanning pretraining, post-training, and inference, representing more than 171,000 H100 GPU-hours of experimentation. Research knowledge acquired from real experimental experience improves RWM predictions of unseen interventions within the same environment (Spearman +0.27), and can be reused across environments. For example, using only pretraining experience from OLMo3, Marin, and Nanochat, an RWM reduces selection regret in the Qwen3 environment by 78% compared with zero-experience setting. These benefits extend to multi-round Autoresearch under a fixed selection budget: RWMs with in-env and cross-env research knowledge increase the best gain achieved by 15.8% and 11.6%, respectively. Ablations across 13 language models used as RWMs show that adding research knowledge can improve intervention ranking more than changing models or increasing reasoning effort alone. These findings support language models as RWMs and motivate accumulating experimental data for future RWM training.
Diffusion language models (DLMs) have attracted wide attention for their fast, parallel decoding, but they are believed to be far less efficient to train than autoregressive language models (ARMs): a widely cited study reports that a DLM needs roughly 16x the training compute of an ARM to match its likelihood, a figure that has made pretraining DLMs from scratch look prohibitive at scale. We find that the methodology behind this figure has fallen out of step with current DLM practice: it scores the model with its evidence lower bound (ELBO) and trains a full-sequence, decoder-only model. We therefore revisit the scaling laws of DLMs, adopting the exact likelihood under the model's own decoding order, block diffusion training, an encoder-decoder architecture, and a sweep over the block sizes used in practice. Under this new protocol, we study across ten model variants, about 1,700 models in total, and find that pretraining DLMs is far cheaper than previously thought. Concretely, scoring with the exact likelihood reduces the compute multiple to 12.8x; block diffusion training and an encoder-decoder architecture further reduce it to 7.8x at block size 32, and smaller block sizes narrow it to 4.6x at block size 8 and 3.0x at block size 4. We complement these scaling analyses with an empirical rule for the optimal decoder depth of encoder-decoder block diffusion, a held-out compute budget beyond the fitting range at which the gap between variants persists, and downstream benchmark results where likelihood improvements carry over to accuracy. Overall, the 16x figure substantially overstates the pretraining cost gap between DLMs and ARMs, and pretraining DLMs from scratch is far more viable than previously assumed.
Large language models (LLMs) are increasingly used to automate AI research, including language-model (LM) development, offering a potential path toward recursive self-improvement. However, turning experimental feedback into a learned ability to propose effective improvements across models faces two challenges. First, training such an ability with real feedback is expensive: reinforcement learning (RL) requires evaluating tens of thousands of proposals, each potentially requiring a full model-training run. Even at ten GPU-hours per run, reward collection alone would consume hundreds of thousands of GPU-hours, motivating a cheaper surrogate. Second, surrogate feedback can become unreliable as policy optimization favors overestimated proposals when the proposal distribution shifts. To address these challenges, we propose LM2LM, which trains LLMs with RL to generate effective LM improvement proposals. To provide scalable feedback, we use a research world model (RWM), a frozen LLM that predicts proposal gains from real experimental records provided in context, as the reward model for RL. To maintain feedback reliability, we periodically execute current-policy proposals ranked highly by the RWM and add their measured outcomes to its context. Across eight autoregressive and diffusion LM development environments, real-execution evaluations on held-out tasks show that LM2LM enables Qwen3.8-27B to outperform GPT-6-sol, even when GPT-6-sol is given 300 real experimental records in context, while requiring 33% lower data-generation cost. LM2LM also generalizes to a model family unseen during training.
Jiaheng Xiao, Jiangtao Han, Jiachen Wang, Honghai Li, Lijun Sun, Kay Axhausen, Chang Liu, Zhengbing He
Understanding mobility dynamics and the effects of urban transportation management policies requires examining how individual travel decisions, traffic conditions, and policy interventions influence one another over time. We present AgenticMobility, an open-source experimental platform for studying how mobility AI agents and other AI agents interact, adapt through repeated experience, and collectively shape mobility outcomes. Persistent agents retain identities and relevant states across rounds, with explicit interfaces defining information access, admissible actions, and decision schedules; linked records trace decisions, execution, and feedback. In the current implementation, large language model (LLM) agents act as travelers choosing routes and departure times and as a manager periodically adjusting link charges within MATSim. A preliminary demonstration spans 100 simulated days in each of five configurations, with one run per configuration and a shared simulation seed. In the joint traveler--manager run, all 1,179,300 finalized target decisions satisfied action constraints; execution checks verified implementation of accepted choices and preservation of personal memories across days. Trajectories show stable route choices alongside departure-time adjustments and differences between controlled-trip and population-wide outcomes. The platform supports experiments on individual adaptation, collective outcomes, and management interventions by varying information access, retained experience, and management feedback. Source code is available at https://github.com/Agentic-City-Research/Agentic-Mobility
Auto-harness systems improve LLM agents by adapting prompts, skills, tools, and supporting infrastructure from execution feedback. Sustaining these gains on open-ended task streams requires handling growing histories, heterogeneous tasks, and shifting distributions. We introduce Adaptive Auto-Harness, a unified framework that separates persistent evolution workspaces from deployable harnesses and supports interchangeable evolution procedures. We compare configurations adapted from five prior approaches on recorded streams of prediction markets, security challenges, and event forecasts. With the same Sonnet solver, our full system reaches 80.9\% market all-task accuracy and 50.2\% security Pass@1, versus 22.2\% and 37.2\% without evolution; market accuracy also exceeds a market-following control's 72.0\%. Our analyses show that artifact accumulation does not guarantee sustained gains, and knowledge that transfers across tasks can still lose value as the task distribution changes. Continued evolution can therefore help preserve earlier gains through revision of accumulated knowledge. Code: https://github.com/A-EVO-Lab/a-evolve.
Zihan Chang, Shuibing He Zhejiang University, Bo Zhou, Sheng Xiao Zhejiang University, Siling Yang, Rui Wang Zhejiang University, Zhe Pan
In response to the increasing demand for largescale machine learning training jobs, many organizations have deployed GPU clusters across geographically distributed regions. However, existing ILP- or genetic-based cross-cluster training approaches largely overlook the topology of decentralized clusters, lacking both topologyaware task scheduling mechanisms and automated model parallelization strategies. As a result, naively applying these optimization-based methods in cross-cluster settings leads to prohibitive scheduling overhead, due to the drastically enlarged search space induced by complex inter-cluster topologies. To address these challenges, we propose SpiderFlow, a topologyaware scheduling system specifically designed for decentralized GPU clusters. We formulate cross-cluster task scheduling as a graph optimization problem and introduce SpinSearch, a low-overhead topology-aware scheduling algorithm. In addition, for automated model parallelization, we propose TPA, a two-level scheduling framework that combines heuristic methods at the inter-cluster level with ILP-based optimization within clusters, effectively reducing the search space while maintaining high training throughput with substantially lower scheduling overhead. We evaluate SpiderFlow on a physical platform comprising 8 decentralized clusters, as well as on a simulation platform with up to 64 decentralized clusters. Experimental results demonstrate that SpiderFlow reduces job completion time (JCT) by 1.2-1.3×, improves throughput by 1.12-1.25×, and reduces scheduling overhead by 20-90× on average compared to state-of-the-art scheduling systems.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表