AREX-2
An LLM agent framework with self-improvement capability that iteratively refines solutions at test time through two domain-agnostic mechanisms—reflection and long-horizon execution—while supporting cross-domain capability transfer.
Timeline (last 90 days)
BAAI's AREX-2 is a 27B-parameter model, released under the Apache 2.0 license and not gated, and the architecture class declared in its config.json is Qwen3_5ForConditionalGeneration.
AREX-2 is trained to solve complex problems in rounds, with a workflow of proposing a plan, measuring results, and reflecting to correct.
AREX-2's config.json declares mtp_num_hidden_layers as 1, but its safetensors weight index contains no tensors prefixed with mtp (0), whereas the index of the base model Qwen3.8 contains 15 mtp tensors.
In AREX-2's tokenizer_config.json, the pre-tokenizer and decoder settings differ from those of Qwen3.8-27B, which may cause garbled tool calls after the agent runs for several hours.
There is currently no NVFP4 build of AREX-2; available quantized versions include BAAI's original BF16 (54.7 GB), prithivMLmods' FP8 (35.9 GB), cyankiwi's AWQ INT4 (20.2 GB), and numsu's INT4 W4A16 (18.6 GB).
AREX-2 requires mamba_ssm_dtype to be set to float32, while the base command forces bfloat16, a setting that works on Qwen3.8 but has not yet been verified on AREX-2.
AREX-2 achieved the following scores on deep research tasks: BrowseComp 84.0, HLE 52.6, GAIA 92.2, and DeepSearchQA 93.8
Despite training data coming from machine learning and algorithmic programming, AREX-2 can transfer to deep research tasks, supporting the assumption that capabilities are domain-agnostic
The core objective of AREX-2 is to advance the self-improvement capability of LLM agents to iteratively optimize their own solutions at test time
Evaluation metrics for machine learning tasks and test cases for algorithmic programming tasks provide objective, automated feedback, making large-scale trajectory synthesis possible and avoiding reliance on manual annotation
7 more timeline events
All Facts (17)
BAAI's AREX-2 is a 27B-parameter model, released under the Apache 2.0 license and not gated, and the architecture class declared in its config.json is Qwen3_5ForConditionalGeneration.
50%UnverifiedAREX-2 is trained to solve complex problems in rounds, with a workflow of proposing a plan, measuring results, and reflecting to correct.
50%UnverifiedAREX-2's config.json declares mtp_num_hidden_layers as 1, but its safetensors weight index contains no tensors prefixed with mtp (0), whereas the index of the base model Qwen3.8 contains 15 mtp tensors.
50%UnverifiedIn AREX-2's tokenizer_config.json, the pre-tokenizer and decoder settings differ from those of Qwen3.8-27B, which may cause garbled tool calls after the agent runs for several hours.
50%UnverifiedThere is currently no NVFP4 build of AREX-2; available quantized versions include BAAI's original BF16 (54.7 GB), prithivMLmods' FP8 (35.9 GB), cyankiwi's AWQ INT4 (20.2 GB), and numsu's INT4 W4A16 (18.6 GB).
50%UnverifiedAREX-2 requires mamba_ssm_dtype to be set to float32, while the base command forces bfloat16, a setting that works on Qwen3.8 but has not yet been verified on AREX-2.
50%UnverifiedAREX-2 achieved the following scores on deep research tasks: BrowseComp 84.0, HLE 52.6, GAIA 92.2, and DeepSearchQA 93.8
50%UnverifiedDespite training data coming from machine learning and algorithmic programming, AREX-2 can transfer to deep research tasks, supporting the assumption that capabilities are domain-agnostic
50%UnverifiedThe core objective of AREX-2 is to advance the self-improvement capability of LLM agents to iteratively optimize their own solutions at test time
50%UnverifiedEvaluation metrics for machine learning tasks and test cases for algorithmic programming tasks provide objective, automated feedback, making large-scale trajectory synthesis possible and avoiding reliance on manual annotation
50%UnverifiedThe core conclusion of the AREX-2 paper is that long-horizon reflection data is an effective path toward self-improving agents
50%UnverifiedAs the iteration budget increases, the AREX-2 agent continues to improve rather than hitting a ceiling prematurely
50%UnverifiedAREX-2 decomposes self-improvement capability into two complementary mechanisms: reflection and long-horizon execution
50%UnverifiedAREX-2 is based on a core assumption that reflection and long-horizon execution capabilities are domain-agnostic and can be transferred from scenarios suitable for supervised training to different task domains
50%UnverifiedAREX-2 synthesizes long-horizon improvement trajectories from two domains—machine learning tasks and algorithmic programming tasks—as training data
50%UnverifiedAREX-2 is built on Qwen3.8-27B
50%UnverifiedAREX-2 scored 81.8 on MLE-bench Lite and 70.7 on Frontier-CS
50%