VeriTrace Steers Research Agents With Three Feedback Loops
Cambridge researchers give a deep research agent explicit feedback loops. With the same Qwen3.5-27B model as its rivals, it beat the strongest by 1.49 points overall and 4.22 on insight.
TL;DR
- VeriTrace, from the University of Cambridge, keeps a research agent's working picture of a question in a graph that three explicit feedback loops revise [^1].
- With the same Qwen3.5-27B model as its rivals, it beat the strongest one by 4.22 points on DeepResearch Bench's insight score [^1].
Background
Deep research agents search the web over many rounds and write a long report from what they find [^1]. Because the material outgrows the model's context window, recent systems keep an intermediate layer between rounds, such as an evolving outline with an evidence bank, or to-dos with an acceptance checklist [^1]. The authors argue that these systems leave the layer's evolution to the model's implicit reasoning, so mixed-quality information can contaminate it and errors spread along its dependencies [^1].
What happened
Researchers from the University of Cambridge's Department of Engineering, the Technical University of Munich and the Alan Turing Institute treat that layer as the agent's mental model of the task and store it as a "cognitive graph" [^1]. Nodes are the concepts under inquiry, edges are the questions that link them, and each node records acceptance criteria, findings and a quality history [^1]. Three loops regulate the graph. Interpretive update classifies each new finding against what is already there, as criterion-satisfying, redundant, contradictory or unexpected, instead of simply appending it [^1]. Deviation feedback scores each search result against what the planner expected, separately for relevance and source credibility, and routes the next search to one of five strategies, such as cross-checking against authoritative sources or switching source type when a page is blocked [^1]. Schema revision restructures the graph itself when repeated feedback shows the framing is wrong, while preserving the evidence collected so far [^1].
To judge pages, a reader model rates each one from 1 to 5 on currency, relevance, authority, accuracy and purpose [^1]. The final report is written from the graph and a separate evidence store of verbatim quotes with their source URLs. These records enter the store without further LLM processing, so every citation can be traced back to its page [^1].
For a fair test, the team reran three open systems, WebWeaver, EnterpriseDR and FS-Researcher, on the same Qwen3.5-27B backbone [^1]. On DeepResearch Bench, 100 research queries in Chinese and English scored against expert reference reports, VeriTrace reached 52.28 overall against 50.79 for WebWeaver, the strongest of the three, and 55.86 on insight against 51.64 [^1]. On DeepConsult, 102 queries judged pairwise by Gemini 2.5 Pro, its win rate was 81.1 percent, against 75.2 for WebWeaver [^1]. With DeepSeek-V4 models, VeriTrace reached 55.77 overall, which the authors call the strongest reproducible open-source result on DeepResearch Bench [^1].
Why it matters
The paper's central claim is that part of the apparent need for bigger models in deep research comes from missing regulation [^1]. The controlled rerun fits that reading: FS-Researcher scored 53.94 overall with Claude Sonnet 4.5, but 48.03 on Qwen3.5-27B [^1]. The ablations point the same way. Removing all four of VeriTrace's mechanisms dropped it to 49.80, below the WebWeaver baseline. Without deviation feedback alone, the planner made 1.31 times as many search calls and still scored lower, because it could only add similar searches until the budget ran out [^1]. The authors stress that they do not claim scale is unimportant, or that regulation replaces it [^1].
The results have limits the authors name. Evaluation covers only DeepResearch Bench and DeepConsult, the restructuring triggers are heuristic, the graph is built from scratch for every task, and only non-frontier Qwen and DeepSeek configurations were tested [^1]. Both benchmarks are scored by Gemini 2.5 Pro as judge [^1]. At Alibaba Cloud rates, a query cost $0.63 with the Qwen3.5-27B configuration and $1.96 with DeepSeek [^1]. The authors also warn that outputs should be treated as drafts that need human verification before use in medicine, law or finance [^1].
Practical example
Say Jana, an analyst at a small consultancy, asks a research agent how the world's wealthiest governments invest. The paper traces exactly this question, query 53 of DeepResearch Bench, through VeriTrace [^1].
The first plan has one node for the wealthiest governments, feeding two others: how they operate their wealth and how they allocate assets [^1]. The first search brings back 13 findings, but the pages bundle both topics by investing entity, not by topic [^1]. Instead of grinding on, the planner restructures. Four entity-focused nodes replace the two topic nodes, eight planned search tasks shrink to four, and the 13 findings stay in place [^1].
For the node on NBIM, pages such as its own annual report on the Government Pension Fund Global score high on relevance and credibility, so the router chooses to exploit that source family and narrows the next search to NBIM's official exclusion guidelines [^1]. The finished report says that companies with more than 30 percent of revenue from thermal-coal extraction are subject to exclusion, and that paragraph's citations point to NBIM's own guidelines and a Norwegian Ministry of Finance regulation [^1]. When the writing stage tried to attach a citation from outside the section's evidence, a filter stripped it [^1].
Without schema revision, the same query ran the original eight-task plan until the budget ran out [^1]. Even with the full system, Jana checks the cited sources before the report reaches a client, as the authors advise for finance [^1].
Related gear
We recommend this book because it explains how applications built on foundation models are designed and evaluated, including agents and retrieval, the building blocks of deep research systems like VeriTrace.