Outputs are lossy
A final string collapses a long, structured computation into one token. Two runs can agree on the answer and think in completely different ways.
A research agenda · Fudan University + Shanghai Innovation Institute
The model’s white box is the next data layer.
Beyond final outputs, we read a model’s internal signals — activations, MoE routing, neurons — and its run-level process signals, and turn them into measurable, controllable data across inference, evaluation, and model construction.
CckFdu·white-box core + process telemetry
Read the inside. Not just the answer.
Led by Kang Chen
Ph.D. researcher at Fudan University (advised by Prof. Yixin Cao) · first or co-first author on all nine studies below · prior algorithm-research internships at SenseTime, Baidu ERNIE, and Alibaba Taotian.
The insight
Outputs are the black box. The real signal is inside — and along the way.
Almost all of AI reads a model by what it finally says. But the richest, least-tapped signals are how it computes and how a run unfolds: which neurons fire, which experts route — and where a trajectory commits, traps, or recovers. Both are measurable beyond the final answer. Six studies turn them into a substrate, united by one question — not by one identical pipeline.
White-box core · internal signals Activations · MoE routing · role-conditioned neurons NAD · RAD · SliceGraph · ARM
Process telemetry · run-level signals Partial traces · agent action–observation · trap & repair events TAAR · TraceGraph
One question: beyond the final answer, which signals make reasoning measurable and controllable?
Measurable — and mostly label-free
Activation keys and routing traces need no gold answers; the run-level layer adds only a light diagnostic probe. The signal scales with compute, not annotation.
One substrate, many uses
The same signals feed a control loop, a process-evaluation layer, and cheaper model construction — one core, three directions the nine studies open up.
The engine
One substrate. A control loop, an evaluation layer, and model construction.
From signal to control
How the signal becomes control.
Read a model’s internal and run-level signals as it thinks, score each sampled trajectory, and act — continue, stop, restart, or select. Each of the nine studies establishes one piece of this loop; together they sketch its shape.
Model runtime → Activation / routing / trace hooks → White-box telemetry → Agreement · trap · route-basin → Stop · restart · select → Cost per correct task ↓
16 sampled runs · live agreement example values
Actions this step
- stop 3 low-agreement runs
- restart 1 trapped run, before the trap
- select the densest route basin
Three directions
Where the work goes next.
- 1Inference-time controlstop / restart / select — more correct answers per unit of compute
- 2Evaluation & diagnosticssee the process, not just the score
- 3Model construction & repairbuild & fix models without retraining
The three feed one another: control produces process data, process data sharpens evaluation, and evaluation tells us which parts of a model to rebuild. Below, the research behind each.
Direction 1 · Inference-time control
Turn compute into correctness — not just more tokens.
Internal agreement and control decide when to trust a run, when to stop, and when to restart — reading no answer strings. The same signal makes test-time compute pay off instead of piling up.
Up to 98%fewer generated tokens (NAD)NAD · AIME · early-stopping · research result
+4.6pton an 8B base model (TAAR)TAAR · base frozen; diagnostic policy trained separately
32tokens to a correctness signal (NAD)NAD · first early signal
Proven in:
Direction 2 · Evaluation & diagnostics
See how models think — not just what they answer.
Raw signals become structured, process-level data: atlases of reasoning routes, decision landscapes, trap maps. A layer of evaluation and diagnostics a leaderboard number can’t give — built from signals that are usually discarded.
60,622reasoning trajectories mapped (SliceGraph)SliceGraph · 954 cells · 6 models
85.5%same-answer runs take different routes (SliceGraph)SliceGraph · blinded-validated atlas
+3.1ppSWE-bench Verified, per provider (TraceGraph)TraceGraph · +3.8 pp on common-fired instances
Proven in:
Direction 3 · Model construction & repair
Build and repair models — without retraining.
Reading the white box lets us compose and fix models cheaply: pick and graft the right internal circuits instead of running gradient descent. Better checkpoints, no training run.
Beatsthe oracle expert selector, training-free (ARM)ARM · Qwen3-8B & Qwen2.5-7B pools
2model families, one merge recipe (ARM)ARM · cross-family replication
0gradient steps — pure activation surgery (ARM)ARM · training-free merge
Proven in:
The work
Nine papers, one question.
Nine papers on arXiv, each with real numbers, blinded validation, and reproducible methods. Every claim on this page links to the paper that proves it — read them yourself.
Peer-reviewed: NAD · ICML 2026 Spotlight ARM · EMNLP 2026 Main TAAR · ACL 2026 Findings
NADwhite-box coreICML 2026 · Spotlight
Predicts correctness from internal-activation consensus, reading no answer strings.
up to 98% fewer generated tokens (AIME early-stop) · signal by 32 tokens Read the paper → RADwhite-box coreSelects the answer by MoE-routing agreement — a signal, not a verifier; works where answer-string voting is undefined.
on par where voting works · available where voting breaks Read the paper → TAARprocess telemetryACL 2026 · FindingsDiagnoses “thinking traps” and restarts before them — base weights frozen; a lightweight diagnostic policy is trained separately.
+4.6 pts on an 8B base model · base frozen Read the paper → SliceGraphwhite-box coreAn activation-key atlas of many CoT runs — same answer, different reasoning routes.
60,622 trajectories · 85.5% multi-route Read the paper → TraceGraphprocess telemetryPools agent runs into one decision landscape; its trap map lifts SWE-bench Verified resolved rate.
+3.1 pp per provider (+3.8 on common-fired) Read the paper → ARMwhite-box coreEMNLP 2026 · MainTraining-free agent merging via answer-free activation overlap — build models without retraining.
beats the oracle expert selector · training-free Read the paper → JARwhite-box coreA 64-axis readout of latent reasoning state, built label-free — and native routing reconstructs it at no cost.
r = 0.69–0.86 reconstruction · 95–100% of the gain, for free Read the paper → RISAprocess telemetryRouting as a behavioral coordinate for software agents: explore away from your own history, commit where peers agree.
44.9% → 48.2% on SWE-bench Verified · no judge, no extra runs Read the paper → d_effprocess telemetryA spectral read of a rollout cohort: it concentrates, differentiates, then regroups — a reproducible arc.
prominent peak in 98.5% of 3,105 cohorts · label-free Read the paper →9papers · white-box core + process telemetry
60k+internal-signal trajectories (the core is label-free)
Answer-freeat inference — the selectors read no answer strings
Why this scales
Labels don’t scale. Runs do.
The data writes itself
Internal-signal and process data are produced by the act of thinking — not bought, not annotated. Every sampled run adds cross-model, cross-task, cross-failure observations, so the dataset grows with compute instead of with annotation budgets.
The hard part is instrumentation
Reading activations, routing and role-conditioned neurons across model families — and normalizing them into one comparable representation — is where most of the difficulty lives. Getting that right once is what let the same substrate carry six different studies.
Grounded in nine peer-reviewable papers from Fudan University and the Shanghai Innovation Institute. Every number on this page is real and reproducible — the evidence is on arXiv.
Let’s talk
Let’s build the white-box layer of AI.
If you work at the frontier of model interpretability, reasoning, and inference, we’d like to talk — research collaborations and partnerships welcome.
Read the inside. It is the next data layer.