When an LLM agent doesn't generate a single answer but keeps correcting it at inference time, the value no longer lies only in the model's parameters. It also lies in the ability to use compute to iterate. AREX-2 is a project that brings this intuition to long-horizon reflective tasks: the researchers built synthetic improvement trajectories from machine learning and algorithmic programming problems, two domains where feedback is verifiable and rewards sustained iteration. The agent, built on Qwen3.8-27B, is trained on these trajectories and learns to reflect on the current solution to produce a better one.

The published results give substance to the method: 81.8 on MLE-bench Lite, 70.7 on Frontier-CS, and, in transfer to deep research tasks, 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA. These are not isolated scores: what matters is that the agent keeps improving as the budget of rounds grows. It is here that the story stops being a leaderboard and becomes a matter of architecture and cost.

The hidden cost of reflective rounds

In a static deployment, the dominant cost is loading the model and serving tokens in a single pass. With a reflective agent, the cost shifts to the number of iterations: each additional round can call the model, re-read the context, and rework the solution. For those managing local infrastructure this is not neutral. A 27-billion-parameter model is relatively small compared to high-end generalist models and, with established quantization techniques, can fit into lower hardware budgets; but the real object to manage becomes the orchestration of rounds. Pipelines are needed that limit consumption, maintain context memory, and stop iteration when returns diminish. AREX-2 suggests that if synthetic training data can teach efficient reflection, the gain comes not only from a larger model but from a smarter use of available compute.

There is a second-order implication for those developing or buying self-hosted systems. The research benchmarks show transfer from domains with verifiable feedback to open research tasks: BrowseComp, HLE, and GAIA are not lab tasks but proxies for real analytical work. If a compact model, trained on synthetic trajectories, reaches those levels with a growing round budget, the barrier to creating internal autonomous assistants lowers. Organizations that do not want to send sensitive data to cloud services could evaluate local reflective loops, provided they accept higher orchestration and monitoring costs. It is not a magic solution: every round consumes resources and introduces latency, and the benchmark says nothing about the overall cost of a production system. But the signal is clear: test-time self-improvement is becoming a design lever, not a side effect. For those assessing on-premise deployment, the issue is not only which model to choose but how to orchestrate reflection cycles without inflating TCO; AI-RADAR collects analytical frameworks at /llm-onpremise for evaluating these trade-offs.