§ Research

Analysis of emerging patterns in AI mathematical reasoning capabilities and predictability of model performance on unseen problems.

Proposes Sensorimotor Alignment as offline proxy for evaluating robotic foundation models by testing convergence between vision, language, and motion representations.

Introduces Direct Pass@k optimization for pruning candidate reasoning solutions during test-time inference, balancing solution quality and diversity.

Analyzes phase transitions in approximation bounds for partially trainable neural networks versus fixed random features under spectral decay assumptions.

Proposes selective auditing framework for online monitoring of AI decisions using confidence scores to route to human review versus automatic acceptance.

Introduces SpectralCanary for auditing text-to-image model dataset use by embedding low-frequency signals detectable in generated outputs.

Evaluates cost-aware abstention policies for LLM-based IT ticket routing, comparing classifiers on real 47,837-ticket dataset with asymmetric error costs.

Proposes Memory-Enhanced Test-Time Adaptation (METTA) for meta-RL using dual-timescale memory to improve fast task adaptation from limited interactions.

Proposes low-precision storage for n-gram embedding tables in language models, targeting 196B-parameter tables in DeepSeek-V4.1-Flash without performance loss.

Identifies and fixes two numerical stability bugs in softmax/cross-entropy implementations: Softmax Collapse and Softmax Cancellation.

CausalCoT enforces causal transparency in chain-of-thought reasoning by decomposing problems into DAGs of sub-questions, preventing post-hoc justifications.

LLM agent debugs FineWeb2 Slovak corpus via tool-use auditing, finding 1.4M machine-generated documents (5.32%) for $2.21-$2.84 per run.

BeNice addresses looped architecture inference memory via KV-cache sharing across recurrences, investigating shareability mechanisms.

RLVR checkpoint analysis on OLMo-3 empirically tests whether reinforcement learning induces instrumental convergence (power-seeking, deception).

Bilingual LMs capture L1-driven cross-linguistic influence in L2 reading, validated on 12-language eye-tracking cohorts.

OP-Print demonstrates membership inference attacks against on-policy self-distillation, exposing privacy risks from coupled training views.

Polynomial-time algorithms for learning discrete distributions under person-level differential privacy match information-theoretic optimality.

Evidence-aware evaluation protocol for clinical RAG system on breast lesion corpus tests retrieval chunk size, overlap, and reranking.

CRANE enables post-hoc counterfactual explanations for multimodal classifiers via nearest-neighbor retrieval, agnostic to architecture.

Simulated control evaluations (SCEs) use LM-simulated tool responses to broaden safety testing beyond purpose-built control environments.

Diffusion models for inverse problems fail on ambiguous measurements; proposes prior-consistent DAPS to avoid measurement-guidance favoring non-prior modes.

Action learning recovers stochastic control policies from trajectory data, enabling unified objective for reasoning about system dynamics and optimization.

Empirical study: users defer to confident but wrong LLMs over correct information; demonstrates trust calibration collapse independent of model sycophancy.

Ablates learned unmasking policies in masked diffusion LMs; identifies label source, policy inputs, and capacity as key determinants of speed-accuracy tradeoff.

TTR-Compressor uses test-time regression for sequence compression in LLMs; learns end-to-end via backprop, computes coefficients in closed form at inference.

Quantifies social-media exposure via LLM-inferred user profiles; measures accuracy and distance to social engineering attacks post-API landscape.

Identifies token collapse pathology in graph soft prompting (G-Retriever) across three benchmarks and model families; graph tokens lose discriminative signal.

Advantage-Based On-Policy Distillation weights teacher actions by counterfactual student advantage; outperforms standard OPD by distinguishing preference from utility.

Diffusion language models improve adversarial prompt attacks on black-box LLMs by generating natural, harder-to-detect malicious inputs via learned text decoding.

Algorithmic stability framework decomposes catastrophic forgetting in continual learning into intra-task and cross-task components without prescribing architecture.

Privacy auditing method for deployed models without retraining, using score transport to handle distribution shift in membership inference attacks.

Twin-world credit assignment isolates policy decisions from random sampling noise in episodic reinforcement learning with deterministic components.

PYTHIA framework applies transfer learning and foundation models to material defect and adsorption site optimization with reduced sample requirements.

LACToS method for localized anomaly detection in images, handling multimodal in-distribution data beyond industrial defect contexts.

Empirical benchmark of recent LLMs vs. fine-tuned encoder-decoder models on knowledge-graph-to-text generation, narrowing performance gap.

Diona regression method estimates state derivatives at query points with admissible bounds via local constrained optimization instead of global training.

Autoregressive factorization with parallel gradient boosting for nonparametric high-dimensional probability distribution estimation.

Autoregressive models can leak future tokens during training via architectural bugs, artificially inflating eval metrics while harming generation quality.

Framework for continual learning from user corrections in interactive systems via typed feedback interfaces.

Mechanistic interpretability reveals causal subspaces controlling metaphor representation in LLMs across architectures.

Production case study: moving compliance rules from prompts to structured retrieval context improves LLM pharmaceutical claims generation.

User simulators balancing complex intent adherence with natural conversational behavior for multi-turn AI assistant evaluation.

Music generation method expands piano MIDI to pop production via flow-matching without paired piano-pop training data.

Neuro-symbolic household food recommender balancing dietary constraints and fairness via constrained fair division.

Physics-informed reduced-order models for thin panel dynamics in supersonic/hypersonic flow using Ritz-Galerkin methods.

Bayesian optimization method for high-dimensional latent spaces using affine surrogates with limited evaluation budgets.

Small open-weight LLM trained via RL to select tools from corpora of 10k+, matching proprietary model performance on ToolBench.

Motion-guided VLM for efficient video anomaly detection using temporal gradient magnitude to trigger inference only when needed.

LLM-based framework for translating structured clinical data into patient-readable health narratives while preserving clinical semantics.

Transformer-based neural operator (RheOPT) pretrained on Newtonian flows for data-efficient few-shot prediction of non-Newtonian fluid dynamics.