Trellis CHEN Haoyang 陈昊阳Homepage

User-directed context for LLM sensemaking

Trellis.

Choose what your next answer builds on.

An investigation rarely follows one straight line. Trellis turns conversation turns into an editable graph: explore alternatives separately, bring useful branches together, and inspect which earlier turns reach the LLM.

Follow an investigation See the workspace
From a chat to a working graphFigure 1
Trellis replaces a linear conversation with a directed acyclic graph. Branches and merges support exploration and integration; a context brief and structural refactoring support overview and revision.
Each node holds a conversation turn. Connections determine which turns are inherited as context for continuing the conversation.

The graph changes what the LLM receives.
It is both a map of the conversation and a control for its context.

01 / Follow an investigation

Two lines of inquiry. One deliberate synthesis.

A journalist is investigating AI and the future of work. Organizational change and worker experience need room to develop separately—then a way to inform each other.

AI & the future of workIllustrated scenario
Shared starting questionAI and the future of work
Branch AOrganizational redesign
Branch BWorker experience
Follow-up AIndicators of redesign
Follow-up BChanges in roles and skillsContinue here
Bring the branches togetherLabor-management relations
Included contextContinue hereOutside context

01 / Branch

Separate the questions without losing the starting point.

Continue from the worker-experience branch. The shared question and this branch’s turns form the context for the next response. The organizational branch stays outside it.

02 / Merge

Compare before bringing ideas together.

Both branches now inform a question about labor-management relations. Trellis separates shared history from branch-specific material, then lets the user merge directly or ask AI to blend it with an instruction.

03 / Overview

See the whole inquiry—and the part you are using.

The context brief links its statements to source nodes. Here, the brief includes both perspectives, while the next response still builds on the worker-experience branch. Select a brief sentence to locate its source.

04 / Refactor

Bring useful context into the current path.

Suppose the organizational evidence would help the worker-experience inquiry. Preview a supporting connection, then decide whether to include that branch. Existing downstream responses need review after their context changes.

Context for the next turn

3 turns
  1. AI and the future of work
  2. Worker experience
  3. Changes in roles and skills

The other branch remains available on the graph.

A guided illustration based on the manuscript’s future-of-work scenario (§4.5). The node labels and brief are shortened for this page; no live model is called.

02 / Inside the workspace

Read a turn. Rework its connections. Inspect its context.

Three coordinated views keep the content, conversation structure, and working context together.

Figure 2: the Node Inspector shows an editable prompt and response on the left. The DAG Canvas shows connected conversation turns in the center. The Context Inspector on the right contains a source-linked brief, structural refactoring proposals, and a minimap.
Figure 2 · The original Trellis interface. The active path is highlighted; surrounding branches remain available for inspection and reuse.
A

Node Inspector

Read or edit the selected prompt and response. Continue from the point that matters to the current question.

B

DAG Canvas

Branch, merge, cut, and reconnect conversation turns. The selected node anchors the context inherited by the next turn.

C

Context Inspector

Use source-linked summaries and a minimap to stay oriented. Preview AI-proposed structural changes before accepting them.

Integration

A merge preserves where ideas came from.

The comparison view separates shared history from the two branches. Users can inspect overlap, divergence, and complementary information before creating a multi-parent turn.

View the compare-and-merge interface ↗

Revision

A structural edit changes future context.

Refactoring can bring supporting turns into a path or remove an unhelpful connection. Users inspect the proposed graph first; changed ancestry marks affected descendants as stale.

View the refactoring interface ↗

03 / What the study found

People used the graph to diverge and converge.

16 experienced LLM users compared Trellis with a linear chat interface in a counterbalanced, within-subjects study. Both began with the same five-turn conversation about AI and the future of work.

16experienced
LLM users
56.4%

Content beyond the main thread

Content produced in newly created branches, reflecting exploration of separate lines of inquiry.

Interaction logs · §6.1.2
2.4

Merges per person

Users brought branches back together. 76.9% of merges used AI Blend & Merge.

Mean operations · §6.2.2
57.3%

Refactoring suggestions accepted

Users applied proposals selectively, rather than accepting every suggested change.

Recommendation acceptance · §6.4.2

“it matches how I think—first diverge, then converge.”

P3 on integrating branches · §6.2.3

More structure also asks for more judgment.

Participants described branching as a way to separate lines of thought and avoid interference. Refactoring was harder to use: it could interrupt a clear line of thinking, and participants wanted better explanations of what a proposal would change.

Interviews and observed use · §6.1, §6.4

What this study establishes. Reported sensemaking support and observed use on one text-based research topic. It does not establish more accurate conclusions, broader domain performance, or benefits for group collaboration.

The research

Trellis: Interpretable DAG Scaffolding for Non-Linear LLM-Assisted Sensemaking

The conversation becomes a structure people can inspect and revise—so the next question builds on the context they choose.

← Back to research projects
Paper figure