Skip to content

Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

🕒 Published (v1): 2026-06-29 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Agents-A1 is a 35B MoE agentic model that matches or exceeds trillion-parameter models (Kimi-K2.6, DeepSeek-V4-pro) by scaling the agent horizon rather than model parameters. It uses a three-stage pipeline — full-domain SFT, domain-specific teacher training, and multi-teacher on-policy distillation with salient vocabulary alignment — trained on trajectories averaging 45K tokens from a novel knowledge-action infrastructure.

Problem

Scaling long-horizon agent competence (planning, tool use, verification, recovery) faces two bottlenecks: (1) existing web corpora lack provenance, action traces, execution logs, and verifier signals needed for process-level supervision, and (2) heterogeneous abilities across domains (search, coding, scientific reasoning, tool use) emerge unevenly and interact in complex ways, making it difficult to unify them into a single deployable model. Prior work either scales parameters to absorb these abilities (expensive, hard to reproduce) or scales the horizon but lacks the infrastructure and training methodology to do so effectively.

Method

  • Knowledge-Action Infrastructure (KAG): Heterogeneous corpora are decomposed into five atomic abilities (information acquisition, tool calling, executable iteration, evidence verification, constraint tracking). These are organized as a typed 4-tuple \(\mathcal{G}_d = (\mathcal{C}_d, \mathcal{A}_d, \mathcal{O}_d, \mathcal{V}_d)\) — corpus, actions, observations, verifiers — with linked action records \((s_t, a_t, o_t, v_t)\) preserving process-level supervision. A proposer–solver–verifier self-play loop expands the graph, rejecting trajectories that fail verifiability, validity, process-informative, evidence-covering, or unambiguousness checks.
  • Three-stage training: (1) Full-domain SFT on multi-domain long-horizon data. (2) Domain-level teacher models trained per domain with targeted SFT/RL (search tasks get RL, science gets science-enhanced SFT, instruction-following and tool-calling get RL). (3) Multi-teacher on-policy distillation with Salient Vocabulary Alignment (SVA): for each position, both student and routed teacher distributions are renormalized over the teacher's top-\(k\) support, and a truncated reverse KL is minimized. A domain-normalized objective averages losses within each active domain then across active domains, preventing high-frequency domains from dominating.
  • Multi-domain data pipeline: Covers long-horizon search (relation-chain generation via controlled random walks over wiki graphs, QA pair generation, real-Internet trajectory collection with 256K context), machine learning engineering (tree of executable solution nodes over Kaggle/MLE-Dojo tasks), scientific reasoning (problem construction and self-evolving KAG enhancement), instruction following, and tool calling — six domains total.

Key Contributions

  • Agents-A1, a 35B MoE model that matches or outperforms 1T-parameter models on long-horizon agentic benchmarks by scaling the horizon rather than parameters.
  • A long-horizon knowledge-action infrastructure (KAG) with self-play graph search and expansion, providing process-level supervision grounded in external knowledge, actions, observations, and verifier outcomes.
  • Domain-routed on-policy distillation with salient vocabulary alignment, reducing cross-domain conflict when unifying heterogeneous specialized teachers into a single student.

Results

  • Agents-A1 (35B) achieves leading results vs. Kimi-K2.6 and DeepSeek-V4-pro on: SEAL-0 (56.4), IFBench (80.6), HiPhO (46.4), FrontierScience-Olympiad (79.0), MolBench-Bind (56.8).
  • Remains highly competitive on: SciCode (44.3), HLE (47.6), BrowseComp (75.5).
  • Surpasses GPT-5.5 (xhigh) on FrontierScience-Olympiad (79.0 vs. 73.0) and IFBench (80.6 vs. 71.8).
  • Outperforms Kimi-K2.6 (1T) on 5/9 reported benchmarks and DeepSeek-V4-pro (1T) on 6/9.
  • Training trajectories average 45K tokens; maximum context window 256K tokens with no cap on tool calls per turn.

Limitations

  • The paper does not systematically ablate the contribution of the KAG infrastructure vs. the distillation method vs. teacher training — relative importance of each stage is unclear from presented results.
  • Domain routing requires explicit domain labels at inference time, adding complexity.
  • Trajectory generation relies on strong models (proposer/solver) and verifiers, which themselves require significant compute to build and run.
  • The gap to 1T models on some benchmarks (SciCode, HLE, BrowseComp) suggests the horizon-scaling approach does not uniformly close the parameter-scaling gap across all task types.
  • Self-play graph expansion may saturate or produce diminishing returns as the KAG grows; no analysis of scaling trends with respect to graph size.

Relevance to Agentic AI / LLM Agents

This work directly addresses the central challenge in agentic AI: how to build capable long-horizon agents without scaling to impractical model sizes. The KAG framework provides a principled template for constructing grounded, process-level supervision — a key enabler for training agents that plan, use tools, verify, and recover over extended trajectories. The multi-teacher distillation methodology is broadly applicable to any multi-domain agent system, and the finding that a 35B model can match 1T models on long-horizon tasks suggests that data infrastructure and training strategy may matter as much as model scale for agentic capabilities.