PANINI-DR Gives Deep-Research Agents a Verifiable Shared Workspace

Introduction

Deep research is more than issuing several search queries and asking a model to summarize the results. Difficult investigations require an agent to decide which subquestions to pursue, connect intermediate findings, retain the state of the work, and check evidence before an unsupported claim contaminates later reasoning. PANINI-DR approaches this problem as a persistent, inspectable, and verifiable investigation rather than a one-shot generation task.

How the system works

PANINI-DR is a multi-agent harness that separates global coordination from local research. A coordinator decomposes the main investigation, tracks dependencies between subquestions, and assigns follow-up work based on existing findings. Research agents focus on individual subquestions. A separate verification agent checks whether a proposed answer addresses the assigned question and whether the cited evidence actually supports it.

The agents interact through a Generative Semantic Workspace. Instead of retaining only a final response, the workspace records subquestions, their relationships, findings, and supporting evidence. Later assignments can use earlier findings, while verification feedback can trigger additional investigation when a result is incomplete or weakly supported. This persistent representation prevents research state from disappearing between model calls and makes it easier to identify where an error entered the process.

Evaluation results

The paper evaluates PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven deep-research harnesses. The systems use the same language-model backbone and frozen corpora, helping reduce the impact of differences in the underlying model or external information access.

On BrowseComp-Plus, PANINI-DR reaches 73.3% accuracy, 16 percentage points ahead of the next-best harness. Its advantage is particularly visible on difficult questions: it answers 26 of 50 hard cases correctly, nearly twice the count achieved by the next-best system. The improvement comes with higher inference cost, however, so the results should not be read as evidence that simply adding more agents guarantees better performance.

Ablation studies show that structured planning and persistent state both contribute to the gains. Performance also improves when every finding is checked against its cited evidence before being reused. The authors go beyond answer accuracy by examining whether a system’s saved evidence is sufficient to recover its answer. PANINI-DR achieves the highest answer recoverability on all three benchmarks, and complementary claim-support audits indicate that its answers are more consistently grounded in retained evidence.

Why it matters

The important contribution is not only better scores, but a clearer separation between being correct and being supported. In research settings that require review, auditing, or collaboration, a structured workspace provides an inspectable record of intermediate decisions and sources. It reduces the risk of a model producing a correct answer from memory while being unable to show how the investigation established it.

The trade-off is additional planning, verification, and inference overhead. In practical deployments, teams will still need to balance cost and latency against question difficulty and evidence requirements. Even so, PANINI-DR points toward a broader design principle for research agents: they should not merely search and write, but also maintain state, manage dependencies, and continuously establish where their conclusions come from.

Source: Hugging Face Daily Papers