DeLoop Introduces Self‑Supervised Repeated Actions for Long‑Horizon Agents
agents gemini llama qwen
| Source: HF Papers | Original article
Researchers introduce DeLoop, a self‑supervision technique that curbs repeated tool calls—so‑called loops—in long‑horizon language‑model agents like Qwen, Llama and Gemini.
Language‑model agents that tackle long‑horizon tasks often fall into “loops”: they repeat a tool call they have already made and receive the identical observation, adding no new information. Researchers have found that such loops occur naturally in agents built on Qwen, Llama and Gemini, and become especially frequent when the agent’s history is compressed to fit within its context window.
The new paper *DeLoop: Repeated Actions as Self‑Supervision for Long‑Horizon Agents* proposes a remedy. DeLoop is a self‑supervised preference‑optimization technique that turns the agent’s own rollouts into training signals. A looping action is treated as a negative example, while the first subsequent non‑looping action in the same rollout serves as a positive one. By mining these signals, the method teaches the model to recognise and avoid redundant tool use without relying on external rewards, gold answers or a separate LLM judge.
Why this matters is twofold. First, eliminating loops cuts wasted computation and shortens execution time, a practical gain for any deployment that relies on multi‑step reasoning or tool integration. Second, the approach improves overall task success under context compression, a common bottleneck as agents scale to more complex, real‑world scenarios such as embodied or multimodal applications.
The next steps will reveal whether DeLoop can be incorporated into the training pipelines of the major LLM families that already exhibit looping behaviour. Watch for benchmark results that compare loop‑reduced agents against baseline models, and for announcements from model providers about integrating self‑supervised loop detection into future releases. If successful, DeLoop could become a standard component in the push toward more efficient, reliable autonomous agents.
Sources
Back to AIPULSEN