Nanjing University's DRIFT Catches AI Research Agents Going Wrong 30 Points Sooner
Researchers from Nanjing University introduce TELBench and DRIFT to pinpoint exactly where deep-research agents go wrong mid-trajectory, not just whether they get the final answer right.
Takeaways
- New benchmark TELBench provides 1,000 labeled agent trajectories for span-level error localization, split into easy/hard difficulty tiers. Dataset on HuggingFace.
- DRIFT auditing framework tracks agent claims, checks evidence support, and traces dependencies to pinpoint root-cause error spans — improving F1 by up to 30 percentage points over bare LLM prompting.
- 36.9% of successful trajectories contain process errors, proving final-answer accuracy is a misleading proxy for agent reliability.
- Decision-making and finalization are the riskiest stages (60.5% and 51.8% normalized error rates), not retrieval, which is frequent but low-risk.
- Scaling model size alone doesn't fix trajectory diagnosis — larger Qwen variants showed no consistent improvement on TELBench's hard split.
- Code and data are open: DRIFT is available at GitHub and runs on any frontier LLM without fine-tuning.
When a deep-research agent gives you a wrong answer, the real question isn't what went wrong , it's where. A new paper from Nanjing University's NJU-LINK team tackles this head-on, introducing a benchmark and an auditing framework designed to catch the exact moment an agent's reasoning goes off the rails. The work, titled "Where Do Deep-Research Agents Go Wrong?", argues that final-answer evaluation is fundamentally incomplete, and that the field needs process-level diagnostics to build truly reliable agents.
The Blind Spot in Agent Evaluation
Deep-research agents , systems that autonomously search the web, use tools, inspect evidence, and synthesize answers , have become a serious capability frontier. Benchmarks like GAIA, BrowseComp, and WebArena have pushed evaluation toward long-horizon tasks. But they all share a common flaw: these evaluations remain primarily outcome-centered, assessing final task completion but failing to localize where a research trajectory first becomes unreliable.
The problem runs deeper than just missed answers. Evaluation based on final answers can tell us whether an agent succeeded, but not which part of the trajectory made the result unreliable. And the raw logs that could reveal the answer are a mess: they are long, heterogeneous, and framework-specific. Asking an LLM to just "find the error" in a full trajectory is also unreliable , it might flag harmless exploration as a mistake, or miss the one early unsupported claim that quietly poisoned everything downstream.
There's also a subtler insight the paper surfaces: process errors and final-answer failures are not the same thing. Among all the trajectories analyzed, 36.9% of successful trajectories also contain process errors, meaning agents can stumble mid-reasoning and still luck into a correct answer. Conversely, 97.3% of failed trajectories contain at least one annotated error span. This disconnect makes final-answer metrics a noisy signal for reliability.
Building the Dataset: 2,790 Real Trajectories
The team didn't simulate agent failures , they collected real ones. They collected 2,790 real trajectories from two agent frameworks, three backbone models, and three benchmarks, converted raw logs into semantic spans, and annotated harmful error spans through LLM-assisted expert review.

The data pipeline involved several key steps:
- Trajectory collection: Runs across GAIA-val, XBench, and BrowseComp-test using GPT-5, Gemini-2.5-Pro, and Claude-Sonnet-4.5 under two agent frameworks (MiroFlow and OAgent).
- Span segmentation: Raw logs are converted into "semantic spans" , contiguous chunks of execution around a coherent local objective, like planning, retrieval, verification, or finalization. This is coarser than individual tool calls but precise enough to localize the first bad commitment.