AI & Machine Learning

PivotRL identifies 'pivot' turns in agent trajectories where actions matter most, enabling compute-efficient reinforcement learning that matches end-to-end RL at 4x lower cost.

Discovers 'silent commitment failure,' where some model architectures produce confident, incorrect outputs with zero detectable warning signals before execution.

Provides a causal explanation for 'embedding collapse' in Transformers, linking it to the concept of semantic shift rather than just text length.

KG-Hopper enables 7B-parameter models to outperform 70B systems on complex Knowledge Graph reasoning by embedding the entire multi-hop process into a single 'thinking' stage.

Introduces Cross-Context Verification (CCV) to detect benchmark contamination, finding that contamination is binary: models either recall solutions perfectly or lack reasoning entirely.

DSPA performs preference alignment at inference time by steering Sparse Autoencoder (SAE) features, bypassing the need for expensive weight-update training.

DRTriton uses large-scale synthetic data and curriculum RL to automatically generate highly optimized Triton kernels, significantly outperforming top-tier LLMs.

Introduces git-inspired primitives to enable truly asynchronous and non-interfering multi-agent software engineering collaboration.

Demonstrates that learning systems can stably converge to incorrect solutions when feedback reliability is unobservable.

Achieves state-of-the-art open-vocabulary segmentation using a training-free, purely geometric projection and propagation method.

Reveals that 'erasing' concepts from video diffusion models only suppresses output rather than removing the underlying representations.

Solves the 'recursive drift' problem in self-improving LLMs by using symbolic verification to gate training data quality.

Introduces a counterfactual framework for precise individual credit assignment in collaborative multi-agent LLM systems.

Provides the first unified theoretical formalism for hierarchical memory systems used by long-context language agents.

Proves an information-theoretic lower bound showing that embedding hidden payloads in LLM text must increase its Kolmogorov complexity.

Transitions MLLMs from reactive planning to 'mental navigation' by forcing the construction of hierarchical cognitive maps from egocentric video.

Enables merging independently trained specialist models (e.g., Vision-LLM and Audio-LLM) into a single multimodal model without any paired training data.

Standard entropy-based uncertainty quantification (UQ) fails in RAG because the 'induction heads' that copy correct answers also trigger 'entropy neurons', causing false uncertainty signals.

Rule-State Inference (RSI) inverts the standard ML paradigm by treating known regulatory rules as priors and inferring the latent state of compliance and drift, rather than approximating rules from noisy data.

GSB-PPO lifts proximal policy optimization from discrete action steps to full generation trajectories by framing it as a Generalized Schrödinger Bridge.

Auditing 'Silicon Bureaucracy' reveals that LLM benchmark scores are often inflated by contamination-related memory reactivation rather than genuine generalization.

SparseVoxelDet is the first fully sparse object detector for event cameras that never instantiates a dense tensor, achieving 858x GPU memory compression.

HumanOmni-Speaker achieves end-to-end speaker diarization and lip-reading by compressing high-frequency motion residuals into just 6 tokens per frame.

PRM-as-a-Judge shifts robotic evaluation from binary success/failure to a dense, potential-based progress metric system.

Depth-Recurrent Transformers decouple computational depth from parameter count, revealing a 'computational frontier' where performance on reasoning tasks snaps from zero to perfect based on iteration steps.

The 'Mirage' study demonstrates that frontier MLLMs generate detailed reasoning traces and clinical findings for images they were never actually shown.

Confidence-Evidence Bayesian Gain (CEBaG) provides deterministic hallucination detection for medical VQA without requiring 10-20 stochastic generations.

FIM-Merging provides a theoretical framework for layer-adaptive model merging using the Fisher Information Matrix to bound merging error.

Challenges the gold standard of Upper Confidence Bound (UCB) exploration in diversity-aware bandit tasks.

Identifies structured table data as a primary driver for scaling long-context reasoning in LLMs.

Achieves zero-shot, zero-training collaborative navigation between humanoid and quadruped robots.

Enables high-performance Zeroth-Order (ZO) fine-tuning of LLMs by leveraging online curvature signals.

Reduces token consumption in interleaved multimodal reasoning by over 72% using dynamic visual thoughts.

Introduces a training-free method to visualize and validate the invariances of any feature extractor using diffusion priors.

Hypothesizes and demonstrates a unified Gaussian latent geometry connecting vision encoders and generative models.

Eliminates the need for strictly aligned image pairs in infrared and visible image fusion.

Solves the structural redundancy problem in symbolic regression by collapsing expression DAG isomorphisms.

Reduces human annotation requirements for NLP model testing by up to 95%.

Reveals that frozen LLMs contain person-specific 'neural signatures' that can predict individual brain activity.

Introduces a robust framework for optimal Mixture-of-Experts (MoE) architecture design across six orders of magnitude in compute.

Synergizes prompt optimization with policy optimization to overcome the 'sparse reward' problem in complex reasoning tasks.

Demonstrates that the two standard mathematical interpretations of Temporal Difference (TD) error diverge in deep reinforcement learning.

Identifies the 'golden subspace' for test-time adaptation, enabling extreme efficiency in online model updates.

Uses the chronological visitation order of medical scans as a self-supervised signal for disease progression modeling.

Achieves a 50x reduction in visual tokens for Video-LLMs while preserving over 90% of baseline performance.

Open-sources a high-fidelity foundation model that jointly generates synchronized video and audio using a unified single-stream Transformer.

Introduces a learnable bridge between GELU and ReLU activations to enable deployment-friendly piecewise-linear networks.

Achieves a 75x parameter reduction in 3D medical image segmentation by hybridizing Mamba and Transformer modules.

Decouples high-level reasoning from low-level motor control in robotics using a visual prompting interface.

Releases the first large-scale family of learned sparse retrieval (LSR) models specialized for code (up to 8B parameters).