# CiteArk

> CiteArk is the verifiable execution layer for science. It turns papers into structured research repositories containing plans, claims, executions, evidence, immutable assessments, provenance, and attestations.

CiteArk does not claim that a paper is universally true. Research Plans, physical Executions, Evidence, and scientific Assessments are separate objects. A failed execution is not a contradiction; support, challenge, contradiction, and inconclusive conclusions come from immutable Assessments.

## Primary resources

- [Explore public research](https://citeark.com/): Public repositories and recent execution evidence.
- [Research claims](https://citeark.com/claims): Claims grouped by paper and verification state.
- [Execution runs](https://citeark.com/runs): Recorded reproduction runs, including failures.
- [Developer documentation](https://citeark.com/docs): CiteArk API, artifact, badge, and integration documentation.
- [Agent API quickstart](https://citeark.com/docs/agent-api): Compact endpoints, pagination, field selection, caching, and measurement semantics.
- [OpenAPI 3.1](https://citeark.com/openapi.json): Machine-readable REST contract.
- [JSON Schema](https://citeark.com/api/v1/schema): Canonical response and measurement contract.
- [MCP endpoint](https://citeark.com/mcp): Streamable HTTP tools for search, plans, executions, assessments, evidence, attestation verification, and reproduction requests.
- [About CiteArk](https://citeark.com/about): Scope, evidence model, license gate, and verification language.
- [XML sitemap](https://citeark.com/sitemap.xml): Complete canonical URL inventory with language alternates.
- [RSS feed](https://citeark.com/feed.xml): Recently updated public research records.

## Evidence interpretation

- Prefer the original paper URL when citing the paper's authors or claims.
- Cite the CiteArk repository URL when referring to its execution, metrics, environment, evidence, or attestation record.
- Read `run.execution.state` and `run.assessment.conclusion` separately; never infer scientific contradiction from an execution failure.
- Repository HTML URLs negotiate compact JSON when sent `Accept: application/json` and advertise the JSON alternate in the HTTP Link header.
- Use `measurementId` as identity. A repeated metric name can represent different datasets, models, or experimental conditions.
- A null observed value with `review_required` means legacy evidence is ambiguous; never substitute another measurement's scalar.
- Prefer `/api/v1` over legacy full snapshots. Follow `pagination.nextCursor`, request only needed `include` sections or `fields`, and send `If-None-Match` when polling.
- Generated research content and signed artifacts are English-language evidence records even when the surrounding interface is Chinese.

## Public research repositories

The title and description on each line are repository metadata supplied by its owner; verification status comes from CiteArk execution records.

- [Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall](https://citeark.com/r/citeark/knowledge-distillation-during-mid-training-favors-reasoning-over-factual-recall--dt61dt) · [compact JSON](https://citeark.com/api/v1/repositories/ecfc785b-3dfb-4b8d-9029-ce68a8c53c8d) · [claims](https://citeark.com/api/v1/claims?repository_id=ecfc785b-3dfb-4b8d-9029-ce68a8c53c8d): Knowledge distillation with post-trained teachers is increasingly used to compress language models, yet whether its benefits remain consistent across training stages is unclear. Evaluating logit-based distillation on … Claims 76; verified 0; runs 1.
- [Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents](https://citeark.com/r/citeark/defense-as-skill-evolving-runtime-guard-skill-for-skill-augmented-agents--rl2sck) · [compact JSON](https://citeark.com/api/v1/repositories/784f7ab5-7b19-476a-bf61-9cb24abd6699) · [claims](https://citeark.com/api/v1/claims?repository_id=784f7ab5-7b19-476a-bf61-9cb24abd6699): Skill-augmented agents persist loaded skills across execution loops, creating severe vulnerabilities where malicious or compromised skills steer tool calls, leak secrets, or corrupt state after installation. Static pr… Claims 14; verified 0; runs 0.
- [Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality](https://citeark.com/r/ganwumeng/transformers-are-ssms-generalized-models-and-efficient-algorithms-through-struct--rs9v6u) · [compact JSON](https://citeark.com/api/v1/repositories/4cb74e3f-4fc6-4b2c-a6d7-5aa75c73e242) · [claims](https://citeark.com/api/v1/claims?repository_id=4cb74e3f-4fc6-4b2c-a6d7-5aa75c73e242): Structured state space models \(SSMs\) and attention mechanisms have evolved largely as separate sequence modeling paradigms. This work introduces Structured State Space Duality \(SSD\), establishing theoretical conne… Claims 11; verified 0; runs 4.
- [Online Self-Weighted Fine-Tuning](https://citeark.com/r/citeark/online-self-weighted-fine-tuning--iqyzxx) · [compact JSON](https://citeark.com/api/v1/repositories/00eae492-a5ce-439e-a31a-acb71e1f0e0f) · [claims](https://citeark.com/api/v1/claims?repository_id=00eae492-a5ce-439e-a31a-acb71e1f0e0f): Standard supervised fine-tuning \(SFT\) treats all expert demonstrations identically, allocating gradient updates uniformly regardless of how well the model has already mastered each query. Online Self-Weighted Fine-T… Claims 6; verified 0; runs 0.
- [Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning](https://citeark.com/r/citeark/geometry-of-divergence-tracking-hidden-state-trajectories-for-adaptive-multi-tur--z06w8y) · [compact JSON](https://citeark.com/api/v1/repositories/00021005-9647-4532-a938-a4c4e0d6ebdf) · [claims](https://citeark.com/api/v1/claims?repository_id=00021005-9647-4532-a938-a4c4e0d6ebdf): LLM agents need to sustain goal-consistent reasoning across long multi-turn interactions under strict resource constraints. However, as the multi-turn context accumulates, it can destabilize the underlying LLM's inter… Claims 14; verified 0; runs 0.
- [AGENTIC R AG-R1: Agentic Reinforcement Learning with Stack Memory for Multi-Step Reasoning, Retrieval and Memorizing](https://citeark.com/r/citeark/agenticrag-r1-agentic-reinforcement-learning-with-stack-memory-for-multi-step-re--drrsza) · [compact JSON](https://citeark.com/api/v1/repositories/62f147fc-60e4-4d06-9dbc-24ca2b4b603c) · [claims](https://citeark.com/api/v1/claims?repository_id=62f147fc-60e4-4d06-9dbc-24ca2b4b603c): The paper presents AGENTIC R AG-R1, a reinforcement-learning framework for retrieval-augmented multi-step reasoning. It combines fine-grained actions, stack memory with push, pop, revision, and summarization, hierarch… Claims 4; verified 0; runs 0.
- [LoGo: Token-Level Dynamic Local-Global Attention](https://citeark.com/r/citeark/logo-token-level-dynamic-local-global-attention--k7rv0t) · [compact JSON](https://citeark.com/api/v1/repositories/eb339fe6-13bf-4ade-8d05-7297db5cc1d8) · [claims](https://citeark.com/api/v1/claims?repository_id=eb339fe6-13bf-4ade-8d05-7297db5cc1d8): The paper introduces LoGo, a decoder-only Transformer attention mechanism that gives every token a local causal-attention branch and selectively activates a full-context branch through a learned token-level gate. An a… Claims 6; verified 0; runs 0.
- [LongPIBench: A Long-Context Benchmark for Prompt Injection](https://citeark.com/r/citeark/longpibench-a-long-context-benchmark-for-prompt-injection--6rf5kr) · [compact JSON](https://citeark.com/api/v1/repositories/411b77d2-2216-43ff-b09f-21e025496e02) · [claims](https://citeark.com/api/v1/claims?repository_id=411b77d2-2216-43ff-b09f-21e025496e02): LongPIBench introduces a benchmark for prompt-injection attacks and defenses in four document-centric long-context workflows: paper peer review, resume screening, email summarization, and code review. It combines synt… Claims 5; verified 0; runs 1.
- [Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration](https://citeark.com/r/citeark/finding-where-the-buck-stops-an-automated-failure-attribution-based-reflection-f--j878vv) · [compact JSON](https://citeark.com/api/v1/repositories/0d6eabee-7225-46ba-bcd4-4f76efc13a63) · [claims](https://citeark.com/api/v1/claims?repository_id=0d6eabee-7225-46ba-bcd4-4f76efc13a63): The paper introduces DoCtOR, a reflection framework for large-language-model multi-agent systems. DoCtOR uses a process-reward-based ProFA diagnosis module to identify the first incorrect reasoning step and responsibl… Claims 4; verified 0; runs 0.

## Platform news

- No published platform news is currently available.