AI & Machine Learning

Proposes Modulated Hazard-aware Policy Optimization (MHPO) to solve the instability and mode collapse common in GRPO-based reinforcement learning.

AwaRes enables low-resolution Vision-Language Models to retrieve only the high-resolution image crops needed for a specific query via tool-calling.

Minimum-Action Learning achieves a 10,000x reduction in noise variance for symbolic physical law identification from observational data.

Learns task-specific dense reward functions directly from images using vision foundation models, without requiring privileged simulator states.

Uses SMT solvers to formally verify the physical consistency of tree-based ML models across their entire input domain.

Provides a systematic profiling of VLM inference bottlenecks and releases 'recipes' that cut time-to-first-token by up to 93%.

Provides a formal proof and empirical evidence that Transformers can learn symbolic rules entirely absent from training, debunking the 'stochastic parrot' interpolation-only hypothesis.

Introduces HopChain, a framework for synthesizing multi-hop vision-language reasoning data that yields generalizable gains across 20+ diverse benchmarks.

Identifies a fundamental conflict in Direct Preference Optimization (DPO) for unified models, where image generation quality resists alignment while understanding improves.

Mathematically proves that the Transformer architecture is functionally equivalent to a Bayesian Network performing loopy belief propagation.

Democratizes dexterous robot data collection by enabling high-fidelity 21-DoF teleoperation using only a standard smartphone.

Reveals that cross-lingual knowledge failure in large reasoning models is primarily a script-translation barrier rather than a linguistic or reasoning deficit.

Shows that 'Mid-Training' on high-quality reasoning data is the primary driver of model capability, whereas RL only succeeds as a sparse refinement step.

Leverages cross-lingual inconsistencies to pinpoint exactly which experts in a Mixture-of-Experts (MoE) model store specific factual knowledge.

Exposes 'hidden clones' in VLM ensembles, where models from the same family share correlated errors that naive voting mechanisms fail to detect.

Proposes REAL, a Reinforcement Learning framework tailored for regression and ordinal scoring rather than simple binary accuracy.

Introduces a framework for LLM agents to autonomously evolve their policies and skill libraries during system idle time without retraining downtime.

A backbone-agnostic denoising objective that allows small GNNs to outperform large models pretrained on much larger supervised datasets in physical sciences.

Achieves high-performance online continual learning without the massive memory overhead of traditional experience replay buffers.

Internal activation probing detects LLM 'rationalization' more reliably than monitoring the model's own Chain-of-Thought (CoT).

A dynamic data pruning framework that cuts dense retriever training time by 50% while actually improving retrieval accuracy.

Automates the generation of synthetic machine learning challenges to train agents that can genuinely learn research skills from doing.

Alignment processes induce a 'normative bias' that makes LLMs worse at predicting real human behavior in strategic scenarios.

Enables reliable, training-free emotion steering in speech-generative audio models via direct manipulation of specific emotion-sensitive neurons.

A formal, graph-native memory architecture that treats agent memory as a versioned asset, dramatically outperforming Gemini 2.5 Pro on complex recall.

A framework to quantify and fix 'task steerability,' the common failure of robots to respond to new instructions while mid-task.

Achieves up to a 1,000x gain in RLHF data efficiency by using information-directed exploration and epistemic neural networks.

Introduces a reward framework that reduces LLM reasoning verbosity by optimizing for 'Information Density' via entropy reduction per step.

Shifts retrieval from static contrastive vector alignment to dynamic reasoning trajectories using a generative model (T1) and GRPO.

Identifies that reasoning-induced safety failures occur *during* Chain-of-Thought and proposes a shift to 'decide-then-reason' architectures.

Generates 9 million grid points of 3D spatiotemporal physical fields in seconds, a 10,000x speedup over traditional physics simulations.

Proposes a world model that jointly generates appearance and binocular geometry using an epipolar-aware attention mechanism.

Introduces FineViT and a 450M local caption dataset to solve the 'coarse perception' bottleneck in current CLIP-based encoders.

Provides a sheaf-theoretic proof that local causal consistency in generative models does not guarantee global counterfactual coherence.

Replaces quadratic self-attention with $O(N \log N)$ phase-native coupling for time-series, enabling massive context windows.

Introduces a paradigm for vision-language navigation that uses ubiquitously available semantic floor plans as global spatial priors.

Embeds invisible, agent-specific 'watermarks' into token distributions to enable forensic attribution and topology reconstruction in multi-agent systems.

Achieves an 80% reduction in Chain-of-Thought (CoT) tokens while slightly increasing reasoning accuracy.

Extends LLM context from 32K to 128K by teaching models to selectively skip global attention for ~80% of tokens.

Reduces hallucinations by teaching models 'epistemological humility'—the ability to admit they don't know something—using synthetic non-existent terms.

Develops a zero-watermarking framework that survives AI editing by leveraging invariant relations between image patches.

Unifies large-scale search, recommendation, and reasoning into a single self-contained LLM by treating item IDs as a distinct modality.

Video fine-tuning consistently degrades static image understanding in multimodal LLMs, revealing a zero-sum trade-off between spatial and temporal capabilities.

Introduces a Prompt-Free Universal Region Proposal Network (PF-RPN) that identifies objects in any domain without needing text or image exemplars.

FrescoDiffusion enables coherent, 4K image-to-video generation using a training-free, tiled diffusion method with precomputed latent priors.

Knowledge-Aware Active Learning (KA2L) uses latent space probing to identify what an LLM doesn't know and generates targeted synthetic questions.

Dense retrieval architectures are fundamentally flawed at detecting negation and contradictions due to 'Semantic Collapse' in vector space.

Edit-As-Act reframes 3D scene editing as a goal-regressive planning problem using symbolic action languages rather than purely generative pixel manipulation.

ARES demonstrates high-fidelity data reconstruction from large Federated Learning batches without requiring any architectural modifications to the model.

Mechanistic probing reveals a directional asymmetry in how LLMs encode hierarchy: hypernymy is redundant and resilient, while hyponymy is fragile and compact.