01EMNLP 2026 · First author
SCIT: Testing Causal Cache Carriers in Latent Chain-of-Thought Models
A causal audit for identifying cache objects that carry counterfactual computation.
Latent CoTCausal interventionKV cache
View research noteClose note+
- Question
- Which cached objects actually carry a model’s counterfactual computation?
- Approach
- Source–recipient counterfactuals, K/V component replacement, semantic controls, and matched corruption.
- Evidence
- A competence-gated carrier map across latent-CoT models, with value-cache suffix trajectories prominent in the examined GPT-2 cells.
- Boundary
- Cells that fail the base-task competence gate receive no mechanism claim.
02ICLR 2027 · In preparation
Transport vs. Destructive Sensitivity
Separating answer transport from destructive sensitivity in causal-cache interventions.
Causal cacheRelayReplication
View research noteClose note+
- Question
- When an intervention changes an answer, did it transport computation or merely damage it?
- Approach
- A two-estimand framework with crossed overwrites and independently trained checkpoints.
- Evidence
- Prospective transport patterns and an isolated three-token relay, with directional replication on CODI-Llama-1B.
- Boundary
- Transportability is not treated as proof of the model’s natural computation path.
03EMNLP 2026 · Accepted
FCPRAG: Fusion-Controller Parametric Retrieval-Augmented Generation
A lightweight controller for stable multi-passage LoRA injection under retrieval uncertainty.
RAGLoRA fusionUncertainty
View research noteClose note+
- Question
- How can parametric RAG combine passage adapters without brittle, uniform fusion?
- Approach
- Predicts fusion weights, gates, and temperatures, with a conservative fallback under uncertainty.
- Evidence
- Evaluated on four QA datasets and three LLM backbones; reported gains reached 4.65% on 2WikiMultiHopQA and 7.55% on CWQ.
- Boundary
- Gains are reported for the evaluated retrieval and backbone settings, not as universal improvement.
04NeurIPS 2026 · Submitted
Target Selection Margin for Causal SAE Feature Selection
Target-specific SAE feature selection evaluated through directional interventions.
Sparse autoencodersInterventionFeature selection
View research noteClose note+
- Question
- Does a selected feature transfer from teacher-forced scoring to held-out interventions?
- Approach
- A target–distractor probability margin with label-shuffle controls and cross-model ablations.
- Evidence
- Compared target-aware, random, and generic utility selectors across Qwen, Gemma, and Llama.
- Boundary
- Teacher-forced margins are reported separately from visible generation changes.
05ICML 2026 · Submitted
REWIND: Improving LLM Reasoning Quality at Inference Time
An inference-time controller that locates risky spans, backtracks, and selectively retains better trajectories.
Inference-time controlReasoningEvaluation
View research noteClose note+
- Question
- Can a model repair a reasoning trajectory without restarting the whole response?
- Approach
- Entropy-gradient localization, semantic-boundary backtracking, suffix regeneration, and MH-style acceptance.
- Evidence
- Evaluated across five reasoning benchmarks and three model families, with the clearest gains on difficult mathematics.
- Boundary
- Quality gains are considered together with additional inference cost.
06MICCAI 2024 · Sole author
MARIO Challenge: Medical Image Prediction
OCT sequence diagnosis and progression prediction for neovascular AMD.
Medical imagingOCTRobust ML
View research noteClose note+
- Question
- How can temporal OCT changes support reliable diagnosis and patient-level risk prediction?
- Approach
- ConvNeXt V2 features, Siamese OCT-DiffNet, CLAM-SB multiple-instance learning, and weighted sampling.
- Evidence
- Two sole-authored LNCS papers, third place, Best Method, and an F1 score above 0.84.
- Boundary
- Challenge results are presented as application-specific evidence, not clinical validation.