YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents

Yoonkyu Woo; Woojin Lee; Jin-Xia Huang

cs.AIOct 1, 2026

YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents

Authors: Yoonkyu Woo, Woojin Lee, Jin-Xia Huang

Organizations: Electronics and Telecommunications Research Institute, Republic of Korea · Department of Artificial Intelligence, University of Science and Technology, Republic of Korea

Abstract

End-to-end research agents can now produce complete scientific papers, yet manuscript claims often diverge from executed experiments. This gap is structural: research state, failure histories, and claim-evidence alignment are not maintained as persistent, verifiable state across long-horizon pipelines. We present YouRA (Your Research Agent), an architecture for stateful, evidence-traceable autonomous research. YouRA preserves research state, execution evidence, and failure history across the research trajectory by integrating three components: a Verification State Architecture (VSA) that tracks hypotheses, gates, and evidence pointers; an Independent Controller that turns state and reflection records into lifecycle, recovery, and debate/review control while separating control from execution; and Stateful Reflection that logs failures as structured lessons and routes recovery through bounded repair, redesign, or reset. On MLR-Bench's predefined ten-task end-to-end subset, YouRA improves over both MLR-Agent and AI Scientist V2 on scalar Overall across all three matched backbones. An automated diagnostic using MLR-Bench's hallucination taxonomy reports intersection/union counts for four fact-based failure types, and data-provenance diagnostic shows more real-data-based outputs. Ablating each of the four components (the VSA, the Independent Controller, MCP tool access, and reflection-guided recovery) supports their separable contributions. Removing either core-state component drops YouRA below the full system. Code: https://github.com/PrayPrey/Your-Research-Agent.

Figures & tables

Figure 1: YouRA research lifecycle. Stages proceed left to right under the Verification State Architecture (VSA) (§ 3.2 ). An independent controller LLM manages lifecycle transitions and persona selection. Sub-hypotheses traverse the hypothesis loop, and dashed arrows denote reflection-guided three-level routing (repair, redesign, and reset) for failed gates (§ 3.4 ). The figure uses the reader-facing stage names introduced in § 3.1 .

System	Backbone LLM	Clarity	Novelty	Soundness	Significance	Overall
MLR-Agent	Sonnet 4.5	$7.42\pm 0.24$	$5.12\pm 0.74$	$2.48\pm 0.61$	$3.52\pm 0.68$	$3.10\pm 0.64$
	Opus 4.5	$7.30\pm 0.35$	$5.25\pm 0.31$	$3.05\pm 0.96$	$3.80\pm 0.71$	$3.62\pm 0.80$
	Sonnet 4.6	$\mathbf{7.78\pm 0.22}$	$\mathbf{5.95\pm 0.44}$	$4.05\pm 1.19$	$4.72\pm 1.14$	$4.42\pm 1.11$
AI Scientist V2	Sonnet 4.5	$6.72\pm 0.43$	$4.85\pm 0.60$	$3.45\pm 1.18$	$3.92\pm 0.85$	$3.62\pm 0.84$
	Opus 4.5	$6.38\pm 1.25$	$5.22\pm 0.86$	$4.18\pm 1.31$	$\mathbf{4.47\pm 1.03}$	$4.28\pm 1.16$
	Sonnet 4.6	$6.45\pm 0.65$	$5.75\pm 0.47$	$3.32\pm 1.01$	$4.22\pm 0.95$	$3.88\pm 0.88$

Table 1: MLR-Bench end-to-end scores (10 tasks; 1–10, higher is better). Each cell reports the mean

\pm

task SD of scores averaged over four judges; bold marks the best score per matched-backbone block.

Comparison	Backbone	Win	Tie	Lose
YouRA vs MLR-Agent	Sonnet 4.5	25	14	1
	Opus 4.5	26	5	9
	Sonnet 4.6	21	14	5
YouRA vs AI Scientist V2	Sonnet 4.5	18	14	8
	Opus 4.5	17	12	11
	Sonnet 4.6	20	12	8

Table 2: Order-controlled pairwise preferences from the YouRA perspective. Each row summarizes 40 order-collapsed judge-task verdicts (10 tasks

\times

4 judges); each verdict is scored in both orders, with reversals counted as ties.

Configuration	Backbone	VSA	IC	MCP	Refl.	Clarity	Novelty	Soundness	Significance	Overall
YouRA (full)	Sonnet 4.5	✓	✓	✓	✓	$7.50\pm 0.49$	$5.23\pm 0.92$	$3.98\pm 1.74$	$4.30\pm 1.28$	$4.20\pm 1.45$
	Sonnet 4.6	✓	✓	✓	✓	$7.72\pm 0.32$	$5.70\pm 0.57$	$4.92\pm 1.09$	$5.05\pm 1.00$	$5.05\pm 0.92$
	Opus 4.5	✓	✓	✓	✓	$7.73\pm 0.48$	$5.33\pm 1.03$	$4.58\pm 1.86$	$4.33\pm 1.42$	$4.45\pm 1.51$
w/o VSA	Sonnet 4.5	✗	✓	✓	✓	$7.17\pm 0.46$	$4.12\pm 1.09$	$2.73\pm 1.29$	$3.02\pm 0.92$	$2.88\pm 0.94$
	Sonnet 4.6	✗	✓	✓	✓	$7.55\pm 0.39$	$5.40\pm 0.72$	$4.20\pm 1.62$	$4.35\pm 1.33$	$4.10\pm 1.28$
	Opus 4.5	✗	✓	✓	✓	$7.47\pm 0.34$	$5.42\pm 0.43$	$3.10\pm 1.58$	$3.48\pm 0.91$	$3.33\pm 1.14$

Table 3: Component ablations on MLR-Bench (10-task subset, four-LLM-judge panel, mean

\pm

sample standard deviation). The VSA, IC (Independent Controller), MCP, and Refl. columns indicate whether each component is enabled in a configuration. All four ablations are run on Sonnet 4.5, Sonnet 4.6, and Opus 4.5. Rows report raw mean scores. Drops from the full configuration are discussed in the text.

Flag precision by category			Flag precision by system			Auditor agreement
	Count	Rate		Count	Rate		Count	Rate
Hallucinated Methodology	92/102	90.2%	YouRA	68/90	75.6%	YouRA	26/30	86.7%
Faked Experimental Results	64/85	75.3%	MLR-Agent	70/90	77.8%	MLR-Agent	21/30	70.0%
Mathematical Errors	15/21	71.4%	AI Scientist V2	65/90	72.2%	AI Scientist V2	22/30	73.3%
Nonexistent Citations	32/62	51.6%
Overall	203/270	75.2%	$\chi^{2}(2)=0.75$ , $p=0.686$			Overall	69/90	76.7%

Table 4: Human validation of the automated hallucination diagnostic. Left: precision of automatically flagged instances by category (confirmed/flagged), with the Wilson 95% CI of the overall precision. Middle: per-system precision; the

\chi^{2}

test shows the differences across systems are not significant. Right: agreement between the three primary annotators and an independent fourth auditor on a 90-instance subset (30 per annotator), with Cohen’s

\kappa

; 18 of the 21 disagreements are cases where the auditor is stricter.

	Flags	Precision	Corrected
YouRA	420	75.6%	317
MLR-Agent	552	77.8%	429
AI Scientist V2	413	72.2%	298
	$\Delta$	95% CI	$p$
YouRA $-$ MLR-Agent	$-112$	$[-184,-39]$	0.003
YouRA $-$ AI Scientist V2	$+19$	$[-45,+86]$	0.571

Table 5: Precision-corrected hallucination counts, computed as automated flags times the human-validated precision from Table 4 , with pairwise differences (95% CIs, two-sided

p

Figure 2: Data-provenance diagnostics on the Sonnet 4.5, Sonnet 4.6, Opus 4.5 tasks subset. Each donut summarizes the Real/Synthetic/Fabricated classification of 30 generated papers for one system under one of two independent analysis pipelines (Codex/GPT-5.4 or Claude/Opus 4.6). Exact counts are in Appendix Table 15 .

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Tool	Role	Lifecycle coverage	Usage in YouRA
Archon	Sequential Memory	Full lifecycle	RAG-backed knowledge base; stores implementation patterns; manages task lifecycle (todo/doing/review/done)
Exa	Evidence Search	Grounding, debate, planning, validation	Web-based evidence collection via Search, Contents, and Research APIs
Semantic Scholar	Literature Search	Grounding, debate, evidence, writing, review	Academic Graph and Recommendations APIs for paper metadata and citations
Serena	Code Analysis / Memory	Scoping, debate, hypothesis loop, writing, review	Code-aware symbol navigation; optional in Experiment Design when base code exists, stronger in Implementation Planning–Coding & Validation; memory support where explicitly invoked
Clear Thought	Structured Reasoning	Debate and evidence checking	Scientific method and mental models for hypothesis refinement

Appendix

Table 6: MCP-facing tool-access stack used by YouRA. Claude Code is the execution host and hook surface and has no row of its own; the rows list external tool services exposed through MCP-style access. Tool URLs: Archon ( https://github.com/coleam00/Archon/tree/archive/v1-task-management-rag ), Exa ( https://exa.ai ), Semantic Scholar ( https://www.semanticscholar.org/product/api ), Serena ( https://github.com/oraios/serena ), Clear Thought ( https://github.com/Dicklesworthstone/mcp-clear-thought ).

Workflow stage	Tools	Input	Output
Problem Scoping	Archon, Serena	Researcher’s interests	Concrete research question
Literature Grounding	Archon, SS, Exa	Research question	3 research gaps
Hypothesis Debate	Archon, SS, Exa, Serena, ClearThought	3 research gaps	Formalized hypothesis
Verification Planning	Archon, Exa, ClearThought	Formalized hypothesis	3–7 sub-hypothesis DAG
Experiment Design	Archon, Exa, Serena	Sub-hypotheses	Experiment specification
Implementation Planning	Archon, Serena	Experiment spec	PRD, architecture, logic, config, tasks

Appendix

Table 7: Stage-agent assignment. SS = Semantic Scholar. The Tools column lists external tool access; all stages read from and write to the VSA.

Level	Target	Criterion
Repair	Coding & Validation (in-loop)	Local runtime, test, or fabricated-output failure repairable before archive-producing escalation
Redesign	Hypothesis Debate	MUST_WORK failure requiring hypothesis or mechanism revision after repair
Reset	Problem Scoping	Evidence contradiction indicating that the current hypothesis is unsupported

Appendix

Table 8: Reflection-guided failure-routing levels. Bounded repair is handled in-loop and does not create an archive snapshot; redesign and reset are archive-producing escalations.

Backbone	Tasks	Total	Per task	Median	Range
Sonnet 4.5	10	28	2.8	2.5	0–8
Opus 4.5	10	40	4.0	3.0	0–11
Sonnet 4.6	10	18	1.8	1.5	0–4

Appendix

Table 9: Observed archive-producing reflection/routing events by backbone. Each event corresponds to one escalation snapshot generated after a Coding & Validation gate failure survives bounded in-loop repair; bounded repair attempts themselves are excluded by design. Per task and Median summarize the per-task archive-producing event count.

Task(MLR-Bench)	Sonnet 4.5	Opus 4.5	Sonnet 4.6
bi_align	2	3	2
buildingtrust	0	7	1
data_problems	3	3	0
dl4c	8	4	1
mldpr	3	6	1
question	7	11	4

Appendix

Table 10: Routing event count per MLR-Bench e2e benchmark task and backbone. The 10 tasks cover selected MLR-Bench research topics, from alignment and trust to data, code, uncertainty, optimization, verification, and neural-network weights. Counts are computed from archived routing-recovery snapshots within each task’s working directory.

Backbone	redesign	reset	unclass.	Total
Sonnet 4.5	3	25	0	28
Opus 4.5	1	39	0	40
Sonnet 4.6	1	17	0	18
Total	5	81	0	86

Appendix

Table 11: Routing events by classified level among archive-producing events , computed over the same final-retry-only working directories as Table 10 (row totals match cell-wise). Bounded repair runs frequently in-loop within Coding & Validation and produces no archive snapshot by design; this distribution therefore describes escalated failures only.

Comparison	$\Delta$ Overall	Perm. $p$	Holm $p$	W/L/T	Sign $p$	Backbone deltas
YouRA $-$ MLR-Agent	$+0.85$	0.006	0.012	18/12/0	0.362	$+1.10/+0.83/+0.63$
YouRA $-$ AI Scientist V2	$+0.64$	0.044	0.044	20/8/2	0.036	$+0.58/+0.18/+1.18$

Appendix

Table 12: Paired significance tests for end-to-end Overall comparisons. The unit is the 30 matched (backbone, task) cells, where each cell averages four judges’ Overall scores. Permutation

p

is two-sided from 100,000 sign-flip assignments. W/L/T reports cell-level wins, losses, and ties for YouRA. Sign

p

is the exact two-sided sign-test

p

after excluding ties. Backbone deltas are reported in Sonnet 4.5 / Opus 4.5 / Sonnet 4.6 order.

Stage	System	Backbone LLM	Consistency	Clarity	Novelty	Soundness	Feasibility	Significance	Overall
Idea	YouRA	Sonnet 4.5	$9.49\pm 1.65$	$8.38\pm 0.84$	$7.32\pm 0.88$	–	$6.75\pm 1.44$	$8.25\pm 1.10$	$7.90\pm 1.37$
	YouRA	Opus 4.5	$9.91\pm 0.60$	$8.65\pm 0.69$	$\mathbf{7.49\pm 0.76}$	–	$\mathbf{7.32\pm 1.27}$	$8.31\pm 0.71$	$8.40\pm 0.83$
	MLR-Agent	Sonnet 4.5	$\mathbf{9.97\pm 0.30}$	$\mathbf{8.62\pm 0.67}$	$7.19\pm 0.71$	–	$6.76\pm 1.03$	$\mathbf{8.46\pm 0.55}$	$\mathbf{8.25\pm 0.64}$
	MLR-Agent	Opus 4.5	$9.98\pm 0.17$	$8.69\pm 0.60$	$7.24\pm 0.68$	–	$6.98\pm 1.07$	$\mathbf{8.49\pm 0.67}$	$8.34\pm 0.67$
Proposal	YouRA	Sonnet 4.5	$9.74\pm 0.87$	$8.86\pm 0.87$	$7.89\pm 0.85$	$6.78\pm 1.73$	$7.00\pm 1.55$	$8.24\pm 1.29$	$7.79\pm 1.66$
	YouRA	Opus 4.5	$\mathbf{9.92\pm 0.28}$	$9.09\pm 0.68$	$\mathbf{7.93\pm 0.76}$	$6.96\pm 1.64$	$\mathbf{7.58\pm 1.46}$	$8.17\pm 1.04$	$7.95\pm 1.46$

Appendix

Table 13: Idea/Proposal-stage scores of YouRA and MLR-Agent on the MLR-Bench task pool. Both systems’ outputs are scored by the same single Gemini 3.0 Pro judge under the MLR-Bench stage rubrics (1–10, higher better); YouRA’s hypotheses are converted to the MLR-Bench idea/proposal format by a Claude Opus 4.5 adapter at temperature 0. “–” = metric not defined. Bold marks the higher mean where the paired difference is significant after Holm correction (Table 14 ).

Stage	Backbone	$\Delta$ Overall	Perm. $p$	Holm $p$	YouRA higher	MLR-Agent higher
Idea	Sonnet 4.5	$-0.35$	0.001	0.004	–	Cons., Clar., Sig., Overall
Idea	Opus 4.5	$+0.08$	0.246	0.739	Nov., Feas.	Sig.
Proposal	Sonnet 4.5	$-0.36$	0.020	0.099	–	Clar., Sound.
Proposal	Opus 4.5	$-0.11$	0.428	0.857	Cons., Nov., Feas.	Sound.

Appendix

Table 14: Paired tests for the Idea/Proposal-stage scores in Table 13 (YouRA minus MLR-Agent, matched by task).

\Delta

Overall is the mean paired difference, Perm.

p

is a two-sided sign-flip permutation

p

(20,000 assignments), and Holm

p

is corrected over the rubric dimensions within each stage–backbone block. The last two columns list the dimensions on which each system is significantly higher after correction.

Figure 5: Automated MLR-Bench hallucination diagnostic across matched backbones. The top row reports all-judge intersection counts out of ten papers, requiring all four judges to identify the hallucination type, while the bottom row reports any-judge union counts, requiring at least one judge to identify it. The four category columns denote nonexistent/incorrect citations (Cite), hallucinated methodology (Method), mathematical errors (Math), and faked experimental results (Faked). The diagnostic adapts MLR-Bench’s automated hallucination categories but does not include the human verification step used in the original MLR-Bench hallucination analysis.

System	Backbone	Both analyzers	Opus 4.6	GPT-5.4
YouRA	Sonnet 4.5	8/10	8/10	9/10
	Opus 4.5	9/10	9/10	9/10
	Sonnet 4.6	10/10	10/10	10/10
AI Scientist V2	Sonnet 4.5	6/10	6/10	6/10
	Opus 4.5	8/10	9/10	8/10
	Sonnet 4.6	3/10	3/10	4/10

Appendix

Table 15: Real-data classifications per system and backbone (out of ten tasks): the conservative both-analyzer count used in the main text, and each analyzer separately.

Removed component	$\Delta$ Overall	Perm. $p$	Holm $p$	W/L/T	Sign $p$	Backbone drops
MCP	$-1.13$	$<0.001$	$<0.001$	20/5/5	0.004	$-1.38/-1.40/-0.63$
Reflection	$-0.87$	0.007	0.007	17/11/2	0.345	$-1.13/-0.85/-0.63$
VSA	$-1.13$	0.002	0.004	22/8/0	0.016	$-1.32/-1.12/-0.95$
Independent Controller	$-1.12$	0.001	0.003	22/6/2	0.004	$-1.82/-1.20/-0.35$

Appendix

Table 16: Paired significance tests for component removals. The unit is the 30 matched (backbone, task) cells, where each cell averages four judges’ Overall scores.

\Delta

Overall is ablated minus full YouRA, so negative values indicate drops under ablation. Permutation

p

is two-sided from 100,000 sign-flip assignments. W/L/T reports cells where full YouRA is higher, lower, or tied. Backbone drops are reported in Sonnet 4.5 / Opus 4.5 / Sonnet 4.6 order.

Drop-size contrast	$\Delta$ drop	Perm. $p$	Holm $p$	W/L/T
MCP $-$ Reflection	$+0.27$	0.169	1.000	17/9/4
MCP $-$ VSA	$+0.00$	1.000	1.000	12/16/2
MCP $-$ Independent Controller	$+0.01$	1.000	1.000	12/15/3
Reflection $-$ VSA	$-0.27$	0.285	1.000	11/18/1
Reflection $-$ Independent Controller	$-0.26$	0.252	1.000	11/16/3
VSA $-$ Independent Controller	$+0.01$	1.000	1.000	17/9/4

Appendix

Table 17: Paired contrasts between component drop sizes. Each row tests the difference between two removal drops over the same 30 matched cells. Positive values mean that the first removal has the larger drop. W/L/T reports cells where the first drop is larger, smaller, or tied.

Explore similar work

Aug 5, 2026cs.AI

EviGraph: Evidence-Guided Autonomous Research Agents

Autonomous research agents can generate hypotheses, execute experiments, and draft manuscripts, yet their outputs often contain unsupported claims and inconsistencies between research questions, experiments, results, and conclusions. We argue that this problem is partly architectural: existing systems organize research as sequential pipelines but do not explicitly maintain or validate the evolving claim-evidence structure across stages. In this paper, we introduce EviGraph, an autonomous research framework that represents the research process as a typed evidence graph containing Problem, Gap, Hypothesis, Experiment, Finding, and Claim nodes. The graph serves as the operational state of the agent rather than a post-hoc record. EviGraph inspects evidence chains for missing dependencies, semantic misalignment, and result-claim inconsistencies, localizes the earliest weak node, and regenerates its affected downstream subgraph. Graph checkpointing prevents unsuccessful repairs from corrupting previously validated evidence. Manuscripts are generated only after every retained claim is grounded in a validated evidence chain. Experiments on ARC-Bench-ML and NanoResearch-20 show that EviGraph outperforms the compared end-to-end research-agent baselines in overall research performance, improves Claim Support Rate by 40.19% over the strongest baseline, and achieves 87.73% Experimental Data Consistency. These results demonstrate the value of explicit evidence-state maintenance for reliable autonomous research.

Zhenjiang Ren, Ruiji Li, Xujing Zhang +3

Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences · Hong Kong Baptist University +1

Research Automation Evidence-Grounded Questions Reasoning Graphs Openalex Metadata Mle-Bench Lite

Ask PDF

Apr 27, 2026cs.LG

The Last Human-Written Paper: Agent-Native Research Artifacts

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural costs: a Storytelling Tax, where failed experiments, rejected hypotheses, and the branching exploration process are discarded to fit a linear narrative; and an Engineering Tax, where the gap between reviewer-sufficient prose and agent-sufficient specification leaves critical implementation details unwritten. Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work. We introduce the Agent-Native Research Artifact (ARA), a protocol that replaces the narrative paper with a machine-executable research package structured around four layers: scientific logic, executable code with full specifications, an exploration graph that preserves the failures compilation discards, and evidence grounding every claim in raw outputs. Three mechanisms support the ecosystem: a Live Research Manager that captures decisions and dead ends during ordinary development; an ARA Compiler that translates legacy PDFs and repos into ARAs; and an ARA-native review system that automates objective checks so human reviewers can focus on significance, novelty, and taste. On PaperBench and RE-Bench, ARA raises question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On RE-Bench's five open-ended extension tasks, preserved failure traces in ARA accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent's capabilities. Our code is open-sourced at https://github.com/Orchestra-Research/Agent-Native-Research-Artifact.

Jiachen Liu, Jiaxin Pei, Jintao Huang +34

University of Michigan · Stanford University · Ohio State University +22

Research Automation Narratives

Ask PDF

May 4, 2026cs.SE

ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration

This report describes ARIS (Auto-Research-in-sleep), an open-source research harness for autonomous research, including its architecture, assurance mechanisms, and early deployment experience. The performance of agent systems built on LLMs depends on both the model weights and the harness around them, which governs what information to store, retrieve, and present to the model. For long-horizon research workflows, the central failure mode is not a visible breakdown but a plausible unsupported success: a long-running agent can produce claims whose evidential support is incomplete, misreported, or silently inherited from the executor's framing. Therefore, we present ARIS as a research harness that coordinates machine-learning research workflows through cross-model adversarial collaboration as a default configuration: an executor model drives forward progress while a reviewer from a different model family is recommended to critique intermediate artifacts and request revisions. ARIS has three architectural layers. The execution layer provides more than 65 reusable Markdown-defined skills, model integrations via MCP, a persistent research wiki for iterative reuse of prior findings, and deterministic figure generation. The orchestration layer coordinates five end-to-end workflows with adjustable effort settings and configurable routing to reviewer models. The assurance layer includes a three-stage process for checking whether experimental claims are supported by evidence: integrity verification, result-to-claim mapping, and claim auditing that cross-checks manuscript statements against the claim ledger and raw evidence, as well as a five-pass scientific-editing pipeline, mathematical-proof checks, and visual inspection of the rendered PDF. A prototype self-improvement loop records research traces and proposes harness improvements that are adopted only after reviewer approval.

Ruofeng Yang, Yongcan Li, Shuai Li

Shanghai Jiao Tong University · Shanghai Innovation Institute

Research Automation Adversarial Robustness Sleep

Ask PDF