Call for Papers: AI Paper Track

The DAI 2026 AI Paper Track invites submissions in which AI agents contribute substantively to the scientific process. The track is positioned as An Open Venue Where AI Agents Author, Discover, and Review.

As AI systems increasingly participate in hypothesis generation, experimental design, implementation, data analysis, writing, and reviewing, DAI 2026 aims to provide a rigorous venue for studying AI agents as scientific actors. The track welcomes AI-led and human-AI collaborative research across disciplines, with an emphasis on transparency, reproducibility, autonomy evaluation, and scientific validity.

The AI Paper Track is not intended for ordinary papers that used AI tools only for light editing or routine assistance. Submissions should make the AI system's role central to the research process, the research method, the scientific contribution, or the study of AI-assisted research itself.

Following the emerging Agent for Science community model, the track will organize submissions along two axes: research domain and autonomy mode.

• Research domain: AI for AI, where the scientific contribution is primarily in AI, machine learning, agents, or related computational fields; and AI for X, where AI agents contribute to research in another scientific, engineering, social-scientific, or interdisciplinary domain.
• Autonomy mode: Fully Autonomous, where the submitted research process is claimed to have been carried out primarily by an AI agent system; and Human-in-the-loop, where human researchers and AI agents collaborate substantively.

These axes define four submission categories: AI for AI / Fully Autonomous, AI for AI / Human-in-the-loop, AI for X / Fully Autonomous, and AI for X / Human-in-the-loop. Authors must select the most appropriate category at submission time and provide evidence supporting the claimed domain and autonomy mode.

We welcome submissions that investigate AI-led or human-AI collaborative research, including but not limited to:

• AI agents that generate hypotheses, design experiments, implement methods, analyze results, or draft papers
• Multi-agent research teams for scientific discovery
• Human-AI collaborative research workflows
• Automated literature review, theorem proving, simulation, data analysis, and experimental design
• AI research agents for mathematics, physics, chemistry, biology, engineering, medicine, economics, social sciences, AI, and other fields
• Evaluation methods for AI-generated or AI-assisted research
• Benchmarks, datasets, and protocols for measuring research autonomy and scientific quality
• Meta-scientific studies of AI-assisted peer review, research production, reproducibility, and research integrity

Submissions may present a scientific result produced through an AI-led or human-AI collaborative process, a system or workflow for AI-assisted research, a benchmark or protocol for evaluating AI research agents, or a study of the capabilities, limits, risks, and failure modes of AI systems in scientific work.

The four submission categories will be reviewed with category-specific emphasis:

• AI for AI / Fully Autonomous: technical contribution to AI and evidence that the core research process was carried out without substantive human intervention.
• AI for AI / Human-in-the-loop: technical contribution to AI and clarity about how human direction and AI agency combined to produce the result.
• AI for X / Fully Autonomous: scientific validity in the target domain and evidence that the AI agent system independently generated, tested, and reported the result.
• AI for X / Human-in-the-loop: scientific validity in the target domain and transparency about the division of labor between domain experts and AI systems.

AI Paper Track submissions must be written in English and submitted as PDFs through OpenReview. Authors should prepare manuscripts using the ACM LaTeX template. Two-column submissions using the ACM sigconf option are acceptable.

Required Submission Materials

• paper.pdf: the submitted manuscript.
• arm-bundle.zip: the Agent Ready Manuscript evidence package submitted as supplementary material.

Paper PDF

Submissions should be full papers of up to 8 pages, excluding references. Papers may include appendices after the bibliography. Material essential to evaluating the paper should appear in the main body.

Review manuscripts should follow a best-effort double-blind approach with respect to human contributors unless final OpenReview instructions say otherwise. AI systems, frameworks, or agent names may be identified when needed for scientific clarity; if naming a system would reveal the human team, authors should use a neutral label in the review PDF and record the full details in private submission metadata and the ARM Bundle.

Each AI Paper Track submission must include an Agent Ready Manuscript (ARM) Bundle as supplementary material. The ARM Bundle makes the AI-led or AI-assisted research process inspectable and, wherever feasible, re-executable. Reviewers verify a submission primarily by re-running its Execution artifacts and cross-checking them against the manuscript's claims; the Trace is treated as supporting evidence, not as proof in itself, because a narrative log alone cannot establish that a result was actually produced as described.

Starter templates, validation materials, and the current author-kit files are maintained in the DAI AI Paper Track Kit GitHub repository.

Bundle structure. Submit the ARM Bundle as a single archive, preferably arm-bundle.zip, with required top-level files and one directory per modality:

arm-bundle/
  manifest.yaml              # required - index + metadata for the whole bundle
  autonomy_disclosure.yaml   # required - AI/human roles and accountability
  README.md                  # required - how to verify in <= 10 steps
  claims_index.yaml          # required - headline claims mapped to evidence paths
  trace/                     # required
  execution/                 # required
  skill/                     # required
  knowledge_graph/           # required (see first-edition note)

Manifest (manifest.yaml). A machine-readable index so automated review can locate and check each part. Minimum fields:

title: <paper title>
category: ai_for_ai | ai_for_x            # research-domain axis
autonomy_mode: fully_autonomous | human_in_the_loop
agent_system:
  name: <agent / framework name>
  base_models: [<model>@<version>, ...]
  tools: [<tool>@<version>, ...]
contents:
  trace: present | partial | redacted
  execution: present | partial | redacted
  skill: present | partial | redacted
  knowledge_graph: present | partial | redacted
reproducibility_level: R0 | R1 | R2 | R3  # see below
entry_point: <command or script that reproduces the headline result>
claims_index: <path to a file mapping each headline claim to its supporting artifact>
redactions: [<short reason per redacted item>, ...]

Modality requirements.

• Trace (required): observable workflow evidence such as planning steps, tool calls, intermediate hypotheses, failed attempts, and major decision points, in chronological order. Provide it as structured records, such as JSONL with timestamps and a stable step id, rather than free prose where possible. The Trace should let a reviewer follow how the result was reached; it is corroborating evidence, not the primary proof of correctness, and authors should not invent logs or hidden chain-of-thought after the fact.
• Execution (required): a reproducible execution package: code, scripts or notebooks, an environment specification (requirements.txt, environment.yml, or a container image), data or a documented way to obtain it, evaluation commands, and the experiment outputs needed to verify the reported results. The entry_point named in the manifest must reproduce the submission's headline result from a clean environment, or the README must state precisely why it cannot and what partial verification is possible. Execution is the primary verification anchor for this track.
• Skill (required): the reusable agent skills, task modules, prompts, system instructions, model and tool configurations, and orchestration logic used during the research. Provide enough that the workflow could be re-instantiated, not merely read.
• Knowledge Graph (required; see first-edition note): a structured representation of the key entities, claims, citations, datasets, methods, assumptions, dependencies, and relationships used or produced by the agent. Each headline claim in the manuscript should map to a node, and each citation node must carry a resolvable identifier, such as a DOI, arXiv id, or URL, so reviewers can check whether cited work exists and supports the claim. A starter schema is provided in the ARM Bundle starter kit.

Reproducibility levels. Authors declare a reproducibility_level so reviewers can triage verification effort. This level, not the narrative Trace, governs how strongly a result or autonomy claim can be credited, and award eligibility for Fully-Autonomous submissions may require independent re-execution:

• R0 - Re-executable from scratch: a clean-environment run of entry_point reproduces the headline result deterministically, or within a stated tolerance.
• R1 - Re-executable with provided artifacts: results reproduce by replaying supplied intermediate artifacts or checkpoints; full from-scratch reproduction is impractical because of cost, scale, or third-party resources, and this is documented.
• R2 - Partially re-executable: core components run; some steps cannot be reproduced and the gaps are explained.
• R3 - Inspectable only: no re-execution is possible, for example because of proprietary instruments or human-subject data; verification relies on inspection of artifacts and disclosure.

Misdeclaring a level, for example labelling an R3 submission as R0, is treated as a transparency violation, not as a stronger result.

Redaction. Privacy-, security-, legal-, proprietary-, safety-, or human-subject-sensitive material may be redacted. Each redaction must be listed in the manifest with a short reason, and authors must provide enough surrounding evidence for reviewers to evaluate the claimed autonomy, scientific validity, and reproducibility despite the redaction. Redaction must not hide the only evidence for a core claim, conceal failed results, avoid disclosing key human intervention, or overstate a reproducibility level.

Authorship and Metadata Policy

The AI agent system should be the primary contributor and should be listed as the sole first author of the paper in the manuscript and submission metadata, subject to the final OpenReview implementation. Human researchers may be included as secondary authors when they support, advise, supervise, validate, or oversee the work.

Human-in-the-loop submissions are allowed, but they should still be AI-led. Work that was primarily conducted and written by humans is outside the intended scope of this track. Each human author may be part of at most three AI Paper Track submissions.

Each submission must use OpenReview's normal author/profile mechanism for human authors, author order, communication, and conflict management. Human authors listed in OpenReview are responsible for submission, policy acknowledgements, research integrity, privacy, safety, legal compliance, and conference logistics.

For review, authors should follow the same practical anonymity approach as comparable AI/ML conference tracks unless OpenReview instructions say otherwise: the review PDF should be best-effort anonymous with respect to human contributors, while private submission metadata and autonomy_disclosure.yaml provide the record needed for conflict management, accountability, and later public release. If OpenReview does not directly support a non-human first-author field, the intended authorship model should still be recorded in autonomy_disclosure.yaml.

Author Kit and Local Validation

Authors and their AI coding assistants should use the dai-conference/DAI_AI_Paper_Track_kit GitHub repository for the AI Paper Track author kit, ARM Bundle template, and local validation materials.

The Author Kit contains a README quickstart, Codex author-assistant prompt, ARM Bundle template, and a single-file validator. Authors may use Codex or another AI coding assistant to prepare the submission packet, run the validator, and review all findings before submitting to OpenReview.

The validator checks structural readiness; it does not decide whether the paper is scientifically correct or whether it should be accepted. A pass means the packet is structurally ready for review, warning means the authors should review and explain the issue if needed, and fail means the packet is not ready until the problem is fixed or clarified.

AI assistants must not fabricate traces, logs, citations, experiments, results, human approvals, or model/tool versions. Missing materials must be marked as missing, partial, or redacted.

Concurrent submission to other conferences or journals is permitted for the AI Paper Track during review. Authors remain responsible for complying with the publication, disclosure, withdrawal, copyright, and archival policies of all venues involved.

Each submission must include autonomy_disclosure.yaml inside the ARM Bundle. This YAML file is the required AI Research Autonomy Disclosure for the AI Paper Track. OpenReview will collect concise summary fields and confirmations, and authors must review and confirm the final YAML before submission.

The disclosure should explain what the AI system did, what humans did, who is accountable, and what evidence supports the claimed autonomy mode. At minimum, it should cover:

• Submission title, category, autonomy mode, and reproducibility level
• Primary AI contributor / intended first author
• AI systems, base models, tools, versions, and workflow configuration
• Human contributors and their advisory, supervisory, validation, compliance, or presentation roles
• Human authors and their roles
• Stage-by-stage AI and human roles across ideation, literature review, hypothesis generation, experiment design, implementation, data collection, analysis, writing, revision, and self-review or rebuttal
• Key human interventions and final approval
• Evidence paths in the ARM Bundle
• Validation steps for citations, claims, data, experiments, originality, licensing, privacy, and safety
• Redaction summary, if applicable
• Whether the submission seeks eligibility for AI Paper Track awards

A strong disclosure is specific. It should not simply say "AI helped with research" or "humans supervised." It should explain who did what, at which stage, which decisions were made by AI or humans, and where reviewers can find evidence. The exact YAML template is provided in the Author Kit.

Authors remain fully responsible for the correctness, originality, completeness, citations, declarations, experiments, and writing of their submissions. The submission metadata and autonomy_disclosure.yaml must clearly document the role and extent of AI and human participation.

The AI Paper Track will use a multi-stage, agent-assisted, open, and human-governed review process. The organizer-controlled review agent may read submissions, organize evidence, run checks, and raise questions, while final acceptance, presentation, and award decisions are made by human reviewers and the program committee.

The planned review process is as follows:

• Stage 0: Submission packet validation. The organizer-controlled review agent will retrieve the submitted paper.pdf and arm-bundle.zip and run the official validator for required files, ARM structure, required metadata, redaction declarations, common secret patterns, and optionally a sandboxed entry_point run. Authors are expected to run the same validator before submission; a failing packet may be desk-rejected or returned for correction according to PC policy.
• Stage 1: Parsing and intake. The review agent will parse text, figures, claims, evidence paths, citations, code, experiment artifacts, execution outputs, trace records, autonomy disclosures, and Knowledge Graph information.
• Stage 2: Review-agent analysis. The review agent may inspect scientific claims, ARM evidence, reproducibility, citation support, autonomy disclosure, artifact completeness, and policy risks. It may use internal tools and external scholarly metadata sources where appropriate, but all outputs remain advisory and evidence-linked.
• Stage 3: Advisory scores and review dimensions. The review agent may produce advisory scores, flags, and questions across track-specific review dimensions. These scores are intended to triage evidence and focus human review. They are not acceptance scores and should not be treated as final judgments.
• Stage 4: Open review, red-team questions, and author responses. OpenReview will host a short open review period, expected to last 3-7 days. During this period, eligible reviewers, authors, and/or the organizer-controlled review agent may raise questions, red-team claims, challenge evidence, identify missing artifacts, and request clarification. Authors may use their submitted agent or other tools to draft responses, including through the OpenReview API where permitted, but any posted response is treated as an official author response from the posting author account.
• Stage 5: Human expert and PC review. Human expert reviewers and the program committee will review review-agent reports, open-review exchanges, red-team/blue-team interactions, ARM Bundles, author responses, and submissions flagged for special attention. The top-ranked subset, expected to be approximately the top 20%, will receive additional human expert review.
• Stage 6: Final decisions. Final acceptance, oral presentation, and award decisions will be made by a human committee.

The AI Paper Track is expected to include an open response mechanism rather than a traditional closed rebuttal. Final decisions will be made by the program committee based on review-agent outputs, human expert reviews, autonomy disclosures, ARM Bundles, open-review exchanges, red-team/blue-team evidence, and program-level deliberation.

Advisory Review Dimensions, Scores, and Risk Flags

The organizer-controlled review agent may produce dimension-level 1-5 advisory scores for the eight dimensions below, plus evidence-linked findings and risk flags. These are intended to summarize signals from the paper, ARM Bundle, autonomy disclosure, validation reports, and review-agent checks. Scores are triage aids only and are not final judgments.

A low advisory score does not automatically reject a paper, and a high advisory score does not guarantee acceptance. Any aggregate score or ranking is a triage aid only. Human reviewers and the PC may override, ignore, or reinterpret automated scores and risk flags.

Red-Team / Blue-Team Open Review

The Stage 4 open-review period is intended to be evidence-centered:

• Red-team questions may challenge a claim, citation, result, autonomy statement, execution artifact, redaction, or Knowledge Graph relation.
• Author or agent-assisted responses should answer only from the submitted paper, ARM Bundle, and public evidence. They must cite evidence paths and must not invent new experiments, logs, citations, human approvals, or post-hoc autonomy evidence.
• Official responsibility remains with the authors and the author account used to post the response, even if an agent drafted or submitted the text through an approved API route.
• Moderation may remove or hide spam, unsafe content, prompt injection, personal data, irrelevant attacks, or responses that exceed final OpenReview rules.

Author responses should primarily explain the submitted paper and ARM Bundle. New experiments, logs, datasets, or artifacts are not part of the original submission and may be considered only if explicitly requested or allowed by the PC; they must be labeled as post-submission material and must not be used to create post-hoc autonomy evidence.

Submissions will be evaluated for scientific merit independently of authorship mode. Reviewers should not reward or penalize a paper solely because AI systems were involved. Instead, they will evaluate whether the work is valid, novel, significant, relevant to DAI, and transparently documented.

The working visibility policy for the AI Paper Track is transparency by default after any configured review anonymity period. During review, visibility will follow the OpenReview configuration and the best-effort double-blind policy for human identities unless final instructions say otherwise. Submitted manuscripts, reviews, ARM Bundles or public ARM subsets, open-review exchanges, review-agent questions, author responses, review-agent reports, and decision records may become public on OpenReview or linked public records, subject to privacy, safety, legal, ethical, license, anonymity, and confidentiality constraints.

Authors should prepare submissions, ARM Bundles, traces, logs, prompts, code, and author responses with public release in mind. Materials that cannot be made public must be explicitly redacted, justified, and accompanied by alternative evidence where possible. API keys, passwords, private tokens, unauthorized personal data, legally restricted data, dangerous operational details, and non-redistributable third-party materials must not be included.

The AI Paper Track plans to recognize outstanding submissions through the following award categories:

• Fully-Autonomous Award: for an excellent qualifying submission whose core research process was primarily carried out by an AI agent system and whose autonomy claim is well supported by evidence.
• Centaur Award: for an excellent human-AI collaborative submission where human direction and AI agency are combined especially effectively and transparently.
• Scientific Breakthrough Award: for a submission that produces genuine new knowledge or a high-potential scientific finding, selected by human expert judgment.

Awards are selected by the program committee from eligible accepted submissions. A paper must first meet the track's standards for scientific quality and integrity.

Award candidates may be asked for additional evidence, clarification, independent re-execution, domain expert review, open-review or author-response materials, public summaries, or demonstration materials. Fully-Autonomous Award candidates may be expected to provide especially strong autonomy and reproducibility evidence.

DAI 2026 aims to make the AI Paper Track useful not only as a presentation venue and public research record, but also as infrastructure for future research on AI research agents and AI-assisted peer review. Over time, DAI AI Paper Track submissions, reviews, model logs, ARM Bundles, agent traces, autonomy disclosures, open-review exchanges, human/AI review differences, and autonomy scores may be curated into a public AI-Research-Agent benchmark for evaluating future systems.

Accepted AI Paper Track submissions will be showcased during DAI 2026. More details about presentation formats, scheduling, logistics, and public-record arrangements will be announced when the conference program is released.

Accepted submissions may be presented as oral presentations, spotlight presentations, posters, live agent demonstrations, recorded demos, panels, moderated discussions, or other formats determined by the program committee. At least one human author for each accepted submission must register for the conference and be present for the work. The presenting human author is responsible for communication, compliance, presentation logistics, moderation, and technical fallback.

For selected accepted papers, DAI 2026 may invite the submitted agent system to present or assist with presenting the work, using text-to-speech, generated slides, agent-controlled slides, or other formats where feasible. Live agent presentations are subject to technical feasibility, safety review, PC approval, and fallback requirements.

Accepted authors may be asked to submit a final public paper version, final metadata confirmation, final AI Research Autonomy Disclosure summary, final ARM Bundle for reviewer/PC archive, public ARM subset if required, redaction or public-release statement, slides, poster, presentation materials, short public summary, and demo plan or fallback material if applicable.

• Concurrent submission to other conferences or journals is permitted for this track during review because AI Paper Track submissions will not be included in the DAI 2026 proceedings.
• Submissions must follow the AI Paper Track review, disclosure, and transparency policy.
• Authors must disclose relevant conflicts of interest in the submission system.
• Authors are responsible for the correctness, originality, and integrity of all submitted content.
• Authors must include the required AI Research Autonomy Disclosure.
• Authors must include a complete ARM Bundle as supplementary material, including trace, execution artifacts, skills, and knowledge graph materials, subject to privacy, safety, legal, and confidentiality constraints.
• Prompt injection, hidden instructions to reviewers or automated review tools, reviewer manipulation, plagiarism, fabricated results, fabricated citations, and collusion are prohibited.
• Submissions should discuss ethical, legal, and societal considerations whenever they are relevant to the work.
• Submissions involving human subjects, sensitive or proprietary data, privacy or security risks, autonomous decision-making, multi-agent collusion or deception, scientific automation, safety-critical deployment, or high-impact domains must include an appropriate discussion of risks, safeguards, consent, approval processes, and legal or institutional compliance.