2026-04
Workflow
An executable research workflow paradigm
The bottleneck in research is often not a lack of ideas, but the absence of one shared system that holds problem definition, evidence organization, and execution rhythm together. For me, workflow is part of research quality, not an accessory.
Why workflow now matters more
I no longer rely on the idea that insight will appear simply by reading more papers. A more effective approach is to treat research as a maintained system where problem framing, literature mapping, experiment execution, writing structure, and failure retrospectives are persisted outside memory.
My preferred sequence is: start from observed phenomena, identify which failure modes are worth studying, build a focused literature map that shows why prior methods are insufficient, write claims and evidence into a matrix, and then move to implementation and experiments.
How this is made operational
I keep a source-of-truth note for active tasks, a review note for rejected assumptions, and an execution plan with concrete steps. Dashboard-style tools connect long-term targets and short-term actions; they provide external memory and visibility for decision-making.
This workflow is practical, not symbolic. Without clear structure around failures, claims, and execution, work can oscillate between “the method is elegant” and “the problem was never precisely defined.”
Core principles I keep
- Problem first, not method vocabulary.
- Literature serves argument quality, not volume.
- Experiments validate claims, not just populate tables.
- Start writing early so structure guides implementation.
- Keep reviewable traces for key decisions.
2025-09
Paper
Why AV-RISE changed how I think about robust audiovisual speech
AV-RISE was not only another multimodal model. It reframed robustness as a representation-learning problem when modalities are inconsistent or partly missing.
The core question
In real-world audiovisual speech understanding, the challenge is not just adding modalities. It is deciding what to trust when audio or video is unreliable, unsynchronized, corrupted, or mutually contradictory.
Many systems assume fusion is always a gain if both signals are present. In realistic environments, noise, occlusion, pose variation, and subtle mismatches can make fusion a liability instead of a benefit.
What AV-RISE taught me
AV-RISE taught me that denoising is not a pre-processing afterthought; it is part of representation learning itself. A hierarchical view on modality reconciliation allows models to separate signal, noise, and helpful complementary cues before final decisions.
In later emotion tasks, my question became less “can we fuse more?” and more “what evidence does each modality provide for this sample?”
How it changed later work
- I treat “stronger fusion” as a testable hypothesis, not an always-default choice.
- I check whether a model remains robust when evidence is incomplete.
- Whether for emotion understanding or agents, I care more about recovery under degraded evidence.
2026-03
Agent
What a stronger emotion agent should have
I am not building a polished chat-style companion. I am building a system that tracks user state, calibrates uncertainty, clarifies before replying, and respects boundary constraints.
Why I keep the term precise
“Emotion Agent” can easily become a copy of supportive language. That is not enough. The useful problem is whether the system can maintain stable affective judgments and intervention policy under uncertainty and long interactions.
So I view it as a set of measurable capabilities rather than a personality style. The core is intervention logic over time: whether to reflect, clarify, validate, reframe, suggest, or abstain.
The minimum four-layer stack
I currently split this into belief tracking, clarification gating, multi-act stance planning, and versioned skill evolution. The first layer estimates emotion, need, preference, memory, risk, conflict, and uncertainty. The second controls next actions from reflect/clarify/validate to suggest/abstain.
Many systems fail because they enter a wrong interpretation too early and then keep reinforcing it. A mature agent should verify before committing, especially when user intent is ambiguous.
What I am not building
- Not a broad companion shell.
- Not a clinical support system beyond safe boundaries.
- Not open-ended personality simulation; I focus on calibrated perception and intervention.
2026-04
Models
Beyond emotion models: conflict, uncertainty, and repair
Emotion models and large benchmarks have improved a lot. The next step I care about is not only larger models, but better mechanisms for uncertainty and error correction.
Why “bigger” is not the only target
We already have strong universal baselines. Average gains are useful, but they do not explain hard failures in the same way conflict-aware evaluation does.
Many failures are not due to missing evidence but due to a premature default answer before unresolved conflict and ambiguity are recognized.
Current bottleneck
I care more about explicit arbitration and corrective mechanisms than one-shot monolithic inference. On difficult samples, fluent output can still be unreliable without structured repair.
This is the main thread behind ongoing work: when is the general path enough, and when should selective repair trigger? The value is in deciding this explicitly, not only growing model capacity.
Questions to pursue
- How to detect early that a default reasoning path is unreliable.
- How to repair only on hard cases without harming easy cases.
- How to measure conflict, uncertainty, and repair as explicit outcomes.