Intelligence & Reliability
Latent-space computation, causal credit assignment, feature detection, and runtime assurance for embodied AI.

RESEARCH · BUILD · CREATE
Beijing · China University of Geosciences, Beijing
Undergraduate in Computer Science and Technology · Class entering in 2025
I study representation, reasoning, and reliability in AI. My recent work focuses on automated research, search, and continual improvement.
Beyond research, I write long-form fiction, develop learning materials, and stay curious about music, history, and philosophy.

RESEARCH × CREATIVE WORK
Asking questions. Building tools. Imagining worlds worth entering.
Explore my workLatent-space computation, causal credit assignment, feature detection, and runtime assurance for embodied AI.
Exploring latent computation, causal credit, and the limits of model learning.
Connecting question discovery, search, structural reorganization, and experimental validation.
AN OPENING
Research and creation begin with a willingness to look again.
An opening · A 15-second film
A personal AI workspace that brings memory, sustained work, and creation into one place.
Read the announcementCS336 Lecture 1, a return to the Transformer paper, gradient descent by hand, and all ten PS01 consolidation PDFs.
Read the update →A three-month initiative to deepen our mathematical understanding of AI through close reading, first-hand derivation, open-source practice and community dialogue. Build understanding through practice. Let real questions lead the next step.
Read the announcement →InkTide brings inspected paper lessons, author evidence, original prose and figure design into a connected writing workflow, with a separate review of the result.
Read the updateBegin with a question. Let evidence guide exploration and criticism drive evolution. Ariadne v7.5 is a first step toward an autonomous research machine, developing directions worth testing.
Read the updatePast release: Daedalus v7.4: structural thinking across fields, research tools, constructive review and continued exploration within AutoResearch.
Read the updateThe latest public work adds fair UR5e verification timings, pre-motion VLA integration checks, and a hardened writer-entry boundary with independently inspectable signed records.
Read the updateThe new local workbench authorizes exact native VLA requests and compares independently verified execution records. A 40-pair engineering study retains both the matching actions and the one camera-feedback difference.
Read the updateVerifier-Bit Ledger is under review at ICLR, with an arXiv preprint forthcoming. DFO is a paper with submission to IEEE TEVC forthcoming. The portfolio also includes seven AutoResearch workstreams and original music for CIVILIZATION; the web CV and bilingual PDF share the current profile.
View the updated CVThe Innovation Generator subproject of AutoResearch connects source evidence, candidate actions and observed feedback. This update records literature exploration, a controlled computational diagnostic, and improvements to source acquisition and stop closure.
Read the updateV3 through V7 now have independent series introductions; minor versions retain their philosophy, architecture, changes, walkthroughs and source anchors.
Read the update
Compositional computation in latent representations
Explores compositional computation in latent representations, focusing on reasoning depth, state stability and learning.
Explore research
Final-action validation and verifiable execution
Final-action checks, incremental certificate reuse, pre-motion integration checks and signed execution records.
Explore research
Verifier-in-the-loop reasoning
How much improvement can verifier feedback buy? An information ledger connects uplift ceilings, ghost baselines and executable audits.
Explore research
Automated research and recursive improvement
Connecting questions, literature, search, refutation, and frozen experiments so that sustained exploration accumulates evidence future work can build on.
Explore research
INDEPENDENT INITIATIVE · LAUNCHED 2026.10.18
Deeper foundations. Further possibilities.
A three-month initiative to deepen our mathematical understanding of AI through close reading, first-hand derivation, open-source practice and community dialogue. Build understanding through practice. Let real questions lead the next step.

One AI, yours alone.
Explore Vesper
Evidence-guided search, substantive revision and testable research dossiers within AutoResearch.
Explore the current release
Connecting research facts, arguments, paragraphs, and figures so that manuscripts and scientific visuals can be cross-checked.
Explore projectThe contribution-review and writing module in AutoResearch, connecting lens routing, independent critique, evidence profiles, and deterministic checks with research taste and author-guided writing.
Explore the moduleA structured route through source discovery, record processing, staged screening, thematic synthesis, and citation review, supporting evidence preparation within AutoResearch.
Explore the module
InkTide brings inspected paper lessons, author evidence, original prose and figure design into a connected writing workflow, with a separate review of the result.
Explore InkTideThe ability to execute local operations does not imply the ability to discover a long-horizon program independently.
Read the essayA good research workflow should also preserve why an idea failed.
Read the essayUndergraduate studies in Computer Science · Entered 2025
Studying AI, while continuing to learn through literature, music, and expression.

LANCER / HUMAN × AI
Across Codex and Claude Code: a thousand-agent scale of practice, hundreds of agents within a task, and work connecting tools, engineering, and stories.
Read my story of collaboration ↗LET’S CONNECT
I welcome conversations about AI, automated research, literature, and sharing knowledge.
Get in touch