TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios

TL;DR

TRIP-Bench introduces a realistic long-horizon benchmark for interactive agents in real-world travel planning, emphasizing global constraint adherence, multi-tool orchestration, and adaptation to evolving user behavior. It combines a large-scale, tool-augmented environment with 40 travel rubrics and 18 tools, plus a diverse user-simulator and automatic evaluation. To address optimization in long-horizon interactions, the authors propose GTPO, an online multi-turn reinforcement learning method that uses Global Instruction Normalization, Turn-wise Reward Differencing, and Turn-level Reward Normalization to stabilize training; applied to Qwen2.5-32B-Instruct, GTPO outperforms baselines like GRPO and Gemini-3-Pro on TRIP-Bench. The work offers a practical framework for robust long-horizon planning and tool coordination, with implications for deploying real-world autonomous agents in complex, evolving user settings.

Abstract

As LLM-based agents are deployed in increasingly complex real-world settings, existing benchmarks underrepresent key challenges such as enforcing global constraints, coordinating multi-tool reasoning, and adapting to evolving user behavior over long, multi-turn interactions. To bridge this gap, we introduce \textbf{TRIP-Bench}, a long-horizon benchmark grounded in realistic travel-planning scenarios. TRIP-Bench leverages real-world data, offers 18 curated tools and 40+ travel requirements, and supports automated evaluation. It includes splits of varying difficulty; the hard split emphasizes long and ambiguous interactions, style shifts, feasibility changes, and iterative version revision. Dialogues span up to 15 user turns, can involve 150+ tool calls, and may exceed 200k tokens of context. Experiments show that even advanced models achieve at most 50\% success on the easy split, with performance dropping below 10\% on hard subsets. We further propose \textbf{GTPO}, an online multi-turn reinforcement learning method with specialized reward normalization and reward differencing. Applied to Qwen2.5-32B-Instruct, GTPO improves constraint satisfaction and interaction robustness, outperforming Gemini-3-Pro in our evaluation. We expect TRIP-Bench to advance practical long-horizon interactive agents, and GTPO to provide an effective online RL recipe for robust long-horizon training.

Figures (6)

• Figure 1: Overview of TRIP-Bench. Left: data construction via rubric-to-constraint generation, progressive modification-chain synthesis, and complexity-conditioned task curation. Right: long-horizon evaluation pipeline where a travel agent iteratively plans with a unified suite of tools and is assessed by rule-based and turn-level metrics under diverse user-simulator interactions.
• Figure 2: Overview of our training pipeline. Left: 120k trajectories are sampled from synthesized prompts, repaired with three rounds of error feedback, and filtered to obtain high-quality rollouts for SFT. Right: GTPO optimizes on groups of multi-turn rollouts by (i) global instruction-wise normalization, (ii) turn-wise reward differencing, and (iii) per-turn reward normalization.
• Figure 3: Left: Performance vs resource use. Three scatter plots: performance vs #turns (left), output tokens per 10k (middle), and avg reasoning cost (USD, log; right). Models: thinking vs non-thinking; open- vs closed-source by marker shape; dashed lines are trend fits.
• Figure 4: Left: Breakdown of the top-15 highest-error constraints by domain and constraint type (Global vs. Pointwise) in the multi-turn setting. Right: DeepSeek-V3.2-Thinking score rates per rubric under single-turn and multi-turn (regular vs. no-issue) settings.
• Figure 5: Pass-k Performance Results.