Skip to content

Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

🕒 Published (v1): 2025-12-04 16:57 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Nex-N1 introduces a three-component ecosystem—NexAU (agent runtime), NexA4A (agent-framework generator), and NexGAP (trajectory pipeline)—that automates construction of diverse, complex interactive training environments for agentic LLMs. By treating agent environments as declarative, generative specifications rather than hand-coded scaffolds, the system scales to 200+ synthesized frameworks spanning 1–34-node agent graphs and 7 tool-call formats. Models trained on the resulting data achieve SOTA or near-SOTA performance on SWE-bench, τ²-bench, GAIA 2, and BFCL v4.

Problem

Training LLMs as reliable agents requires large-scale, high-fidelity interactive environments that provide behavioral diversity and realistic feedback—but existing approaches either rely on a small number of manually engineered frameworks (narrow behavioral coverage) or purely synthetic/static data (simulation-reality gap, hallucinated tool use, no error recovery). There is no scalable infrastructure for automatically producing the quantity and diversity of interaction trajectories needed for robust policy learning.

Method

NexAU (Nex Agent Universe) is a lightweight, declarative agent runtime built around a recursive ReAct loop. Agents are defined via YAML configs specifying system prompt, LLM config, tools/MCPs, and sub-agents; sub-agents are unified with tools so a parent agent treats a child execution context identically to a function call. Reasoning state is isolated per sub-agent, preventing context overflow in deep hierarchies. The runtime supports MCP integration for live external APIs, dynamically injected Skills (prompt/example/script bundles), and a thread-safe GlobalStorage for cross-step state.

NexA4A (Agent for Agent) is a generative meta-system that takes a natural-language description of a desired agent framework and automatically produces a complete YAML configuration. A MetaAgent selects a workflow pattern, a FrameworkBuilder constructs multi-agent topology (1–3 layers), and an AgentBuilder populates per-node system prompts, meta-skills, tools, and MCP bindings—without writing executable code.

NexGAP (General Agent-data Pipeline) orchestrates end-to-end trajectory synthesis: (1) real MCP tools from curated public repositories serve as capability seeds; (2) a Problem Type Tree (hierarchical, bilingual taxonomy with inverse-frequency sampling) drives Information Fusion Query Synthesis, conditioning each query on persona, problem type, framework context, and difficulty; (3) NexAU executes the generated agents and records raw traces; (4) a Quality Assessment Agent filters trajectories for truncation, repetition, hallucination, and reward hacking (e.g., fabricated test results) by processing long traces in batches and outputting structured JSON issue reports; (5) trajectories are normalized into seven tool-call formats (OpenAI FC, several XML variants).

Key Contributions

  • NexAU: modular recursive agent runtime abstracting heterogeneous framework internals via declarative YAML, enabling large-scale reproducible trajectory generation.
  • NexA4A: automated generative system that synthesizes entire multi-agent frameworks from natural language, replacing manual environment engineering.
  • NexGAP: end-to-end agentic data pipeline producing diverse, grounded trajectories with structured quality filtering; yields 200+ environments, 1–34-node graphs, 7 tool-call format variants.
  • Nex-N1 model family: trained on the above data; outperforms open-source models of comparable size on SWE-bench, τ²-bench, GAIA 2, BFCL v4, and human evaluations.
  • Full open-source release: model weights, inference code, and a training-data subset.

Results

  • τ²-bench: DeepSeek-V3.1-Nex-N1 scores 80.2 vs. base DeepSeek-V3.1 42.8; Qwen3-32B-Nex-N1 72.1 vs. base 41.5 (Claude-Sonnet-4.5 leads at 88.1).
  • GAIA 2: DeepSeek-V3.1-Nex-N1 29.5 vs. base 21.9; Qwen3-32B-Nex-N1 16.7 vs. base 10.8.
  • SWE-bench (verified): DeepSeek-V3.1-Nex-N1 70.6 vs. base 66.0; Qwen3-32B-Nex-N1 50.5 vs. base 12.9.
  • BFCL v4: DeepSeek-V3.1-Nex-N1 65.3 vs. base 56.3; Qwen3-32B-Nex-N1 60.5 vs. base 45.4.
  • Human eval – agentic coding (43 samples, 13 scenarios): Nex-N1 wins/ties in 64.5% of scenarios vs. Claude-Sonnet-4.5, 92.9% vs. Minimax-M2.
  • Human eval – web development (45 samples): Nex-N1 wins/ties vs. all models except Claude-Sonnet-4.5 (44.5% win rate there).
  • Deep Research Benchmark: Nex-N1-powered deep research agent scores 47.0%.

Limitations

  • Visual feedback used for code-quality optimization was occasionally unreliable, requiring conversion to binary judgments and hard repair-iteration caps; some unfixable code is simply discarded.
  • Human evaluation sets are small (43 and 45 samples) and pairwise, limiting statistical power.
  • BFCL evaluation deviates from the official leaderboard setup (DuckDuckGo replaced with Google Search API), complicating direct comparison.
  • NexA4A-generated custom tools are often simple code snippets with limited real-world API interaction; authentic grounding depends on the curated MCP tool subset being representative.
  • Full results section was truncated in the provided text; some benchmark details (e.g., Terminal Bench 2, BaxBench methodology) are only briefly described.

Relevance to Harnesses / Meta-Harnesses

NexAU is itself a meta-harness: a runtime layer that decouples agent definition (YAML schema) from agent execution, unifying heterogeneous frameworks into a single reproducible data-generation substrate—exactly the abstraction that harness research seeks. NexA4A extends this by making the harness itself generatable: it is a harness factory that synthesizes new harness configurations from natural language, closing the loop from human intent to executable environment without manual scaffolding. NexGAP demonstrates that wrapping such a meta-harness in a quality-control pipeline (structured taxonomy, inverse-frequency sampling, batch-iterative trajectory judging) is necessary to produce training-grade data at scale, directly informing how harness builders should think about data fidelity and reward hacking in automated pipelines. Together, the three components offer a concrete blueprint for scaling harness construction—declarative composition, automated topology generation, and LLM-based quality assessment—that is broadly applicable beyond agentic training.