Independent research · In progress
Token-Level Causal Binding Sparsity
How much of a language model’s earlier context actually influences its next-token generation?
RESEARCH / 2026
Hello
Researcher!
AI is getting more capable. But can we make it reliable, interpretable, and aligned?
I explore research questions in language models, multimodal AI, and autonomous systems — with a focus on understanding, evaluating, and improving them.
- ◈ LLM interpretability
- ◌ Multimodal AI
- ⌘ Agentic systems
- ◉ AI evaluation
01
ONGOING QUESTIONS
Current
Investigations
Ongoing research questions, methods, and progress toward more capable, reliable, and understandable AI systems.
Research direction
Detecting Strategic Failure in LLM Agents
Can poor strategic decisions be detected before an agent’s eventual outcome makes a failure obvious?
Illustrative decision signals
Research plan
Prompt Entailment & Hierarchies
How can relationships between prompts be represented and evaluated more systematically?
Conceptual prompt hierarchy
02
WORKING SOFTWARE
Research
Systems
Technical projects, benchmarks, and systems built to explore ideas through working software.
TAFE-ID
Document-tampering localization with an open record of custom training experiments.
NirikshaID Bench
A reproducible VLM and OCR benchmark for structured extraction, tamper detection, calibration, and failure analysis.
MirageEval
An environment for evaluating whether uncertainty signals predict factual errors in LLM outputs.
GEOFORGE
An interactive exploration of agent-based scientific discovery in a living terrain.
ARES
An autonomous guidance research system for simulated Mars entry, descent, and touchdown.
03
OPEN RESEARCH RECORD
Notes &
Publications
Working notes and research plans. Formal publications will appear here when available.
Research notes · In progress
Token-Level Causal Binding Sparsity
An ongoing question about context influence during generation.
Research plans · Research plan
Agent evaluation directions
A working direction for evaluating decisions before final outcomes.
Research notes · Concept
Prompt entailment & hierarchies
A planned investigation into systematic prompt relationships.
No formal publications listed yet.
LET’S INVESTIGATE
Let’s build a more capable and reliable AI.
I’m interested in research collaboration around LLM evaluation, interpretability, multimodal robustness, and autonomous decision-making.