Research

Research

I study how compound AI systems are structured, how they adapt through experience, how their reliability changes over time, and what their internal representations reveal.

Compound AI and LLM-agent architecture
Reliable long-horizon behavior
Prompt-only and inference-time adaptation
Agent evaluation and failure analysis
Mechanistic interpretability of reasoning
Cost-aware inference and experimentation
Research infrastructure and reproducibility

Theme

7 of 7 projects

Adaptation2026
FORGE: Self-Evolving Agent Memory With No Weight Updates
Prompt-only reflection becomes more reliable when useful discoveries are selected and broadcast across a population.
Contribution
A population-based protocol that converts failed trajectories into reusable prompt memory and broadcasts champion memories to every agent.
Result
1.7–7.7× improvement over zero-shot and 29–72% over isolated Reflexion across four model families, with no weight updates.
ACM proceedings paperACM CAIS 2026
- Adaptation
- Agent architecture
1.7–7.7×
over zero-shot
29–72%
over isolated Reflexion
4
model families
Details
Architecture2026
Context, Reasoning, and Hierarchy in Compound LLM Agents
What an agent sees can matter more than how long it deliberates.
Contribution
A controlled cost–performance study of three design axes (context, reasoning, and hierarchy) for compound LLM agents in an adversarial POMDP.
Result
Structured state abstraction improved return by up to 76%. Distributed deliberation made it up to 3.4× worse across every tested model family.
ACM proceedings paperACM CAIS 2026
- Agent architecture
- Evaluation
3,475
episodes
up to 76%
context improvement
up to 3.4×
worse under deliberation cascade
Details
Evaluation2026
Evaluation of Multi-Turn Consistency in LLM Agents
Reliability is temporal: models fail at different stages and produce systematically different narratives before doing so.
Contribution
A time-to-event evaluation methodology and a seven-category failure-rationale taxonomy for long-horizon LLM agents.
Result
Across 84,540 trajectories and eight model families, models show near-flat, early-spike, and bimodal hazard regimes with distinct pre-failure narratives.
Workshop paperICLR 2026 Workshop on LLM Logical Reasoning
- Evaluation
84,540
trajectories
8
model families
13,780
labeled rationales
- Paper (opens in a new tab)
Details
Interpretability2026
Discovering Cross-Language Reasoning Invariance in LLMs
Geometric alignment across languages does not automatically imply functional interchangeability.
Contribution
Geometry-invariant sparse autoencoders (GI-SAE) paired with causal intervention to test whether reasoning features are shared across languages.
Result
Identifies convergent, saturated, and low-sharing regimes in which geometric similarity and causal interchangeability come apart.
Workshop paperICML 2026 Workshop on Mechanistic Interpretability
- Interpretability
GI-SAE
method
Causal
intervention design
3
sharing regimes
- Paper (opens in a new tab)
Details
Evaluation2025
Delay-of-Gratification Multi-Agent Survival Micro-Benchmark
A small, controllable environment can expose long-horizon failure modes that large benchmarks average away.
Contribution
Built the benchmark gym, experimental setup, and data collection.
Result
Accepted to the NeurIPS 2025 Workshop on Multi-Turn Interactions in LLMs.
Workshop paperNeurIPS 2025 MTI Workshop
- Evaluation
- Agent architecture
MDP/POMDP
formulation
Multi-agent
social exposure
- Paper (opens in a new tab)
- Poster (opens in a new tab)
Applied ML2025
Bandit-Driven Prompt Selection for Low-Cost LLM Essay Scoring
Adaptive prompt selection can cut inference cost without giving up grading reliability.
Contribution
Built the bandit agent, experimental setup, and data collection.
Result
78.4% reduction in LLM calls.
Applied-ML paperEDM 2025 WLIEDM
- Applied ML
- Adaptation
78.4%
fewer LLM calls
4
grading recipes
- Paper (opens in a new tab)
- Code (opens in a new tab)
Applied ML2025
Infant Care Video Dataset and Transformer Benchmark
A privacy-compliant clinical video dataset can support strong intervention classification with video transformers.
Contribution
Created the dataset and fine-tuned and benchmarked video transformers.
Result
Up to 94% top-1 accuracy on neonatal interventions.
Dataset paperCOMPSAC 2025 MediComp
- Applied ML
4,144
videos
12
classes
94%
top-1 accuracy
- Paper (opens in a new tab)
- Dataset (opens in a new tab)

Contact

Building reliable AI systems requires both research and engineering.

I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.

Email IgorView CV LinkedIn

Navigation

FORGE: Self-Evolving Agent Memory With No Weight Updates

Context, Reasoning, and Hierarchy in Compound LLM Agents

Evaluation of Multi-Turn Consistency in LLM Agents

Discovering Cross-Language Reasoning Invariance in LLMs

Delay-of-Gratification Multi-Agent Survival Micro-Benchmark

Bandit-Driven Prompt Selection for Low-Cost LLM Essay Scoring

Infant Care Video Dataset and Transformer Benchmark

Building reliable AI systems requires both research and engineering.

Send a message