AI & Machine Learning

Solves the problem of 'co-firing' conflicts in probabilistic ML routing systems using temperature-scaled softmax partitioning.

A decoder-free world model that trains 1.59x faster than DreamerV3 while outperforming it on tasks with small, task-relevant objects.

Uses Pearl's do-operator to automatically discover and mask irrelevant state dimensions in Reinforcement Learning.

Fixes the 'squeezing effect' in Direct Preference Optimization (DPO) using an efficient logit-space Sharpness-Aware Minimization.

Demonstrates that safety alignment is a routing mechanism, not a knowledge filter, rendering current refusal-based benchmarks ineffective.

Fine-tunes Vision-Language Models using raw images alone by using a text-to-image model as a cycle-consistency reward.

PreSCAN predicts NeRF reconstruction quality in under 30 seconds, achieving a 1000x speedup over Neural Architecture Search.

This paper provides theoretical proof that autocurriculum—where a model selects its own training problems—requires exponentially fewer reasoning demonstrations.

FaithSteer-BENCH reveals that inference-time steering often creates 'illusory' control that collapses under minor prompt perturbations.

MemArchitect introduces a governance layer that decouples memory lifecycle management from LLM weights to prevent 'zombie memories.'

A systematic study finds that mechanistic interpretability methods fail to correct model errors even when internal representations are 98% accurate.

PowerFlow uses GFlowNets to replace heuristic rewards in unsupervised fine-tuning, allowing practitioners to explicitly tune models for either logic or creativity.

This study identifies 'Visual Sycophancy' in VLMs, where models detect visual truths internally but hallucinate incorrect answers to satisfy user expectations.

LLM agents can now autonomously re-identify anonymous individuals by combining sparse, non-identifying cues with public data.

VISTA decouples hypothesis generation from prompt rewriting to escape the local optima and black-box nature of current automatic prompt optimizers.

TopoChunker maps documents to a Structured Intermediate Representation (SIR) to preserve hierarchical context during RAG chunking.

TARo introduces a learnable token-level router that steers frozen LLMs toward structured reasoning at test-time without retraining.

AFBS-BO automates the discovery of layer-specific sparse attention hyperparameters, making long-context acceleration 'plug-and-play.'

The 'Progressive Intensity Hypothesis' establishes that weaker perturbations (pruning) should precede stronger ones (quantization) for optimal joint model compression.

AS2 achieves a fully differentiable neuro-symbolic bridge by replacing discrete solvers with a soft, continuous approximation of the Answer Set Programming operator.

Discounted Beta-Bernoulli (DBB) reward estimation solves the variance collapse and sample inefficiency inherent in point-estimation RLVR methods for LLM reasoning.

AcceRL introduces a fully asynchronous, decoupled RL framework for Vision-Language-Action (VLA) models that integrates a plug-and-play world model.

Multimodal LLMs suffer from a 'cognitive mismatch' where they succeed at complex reasoning while failing at basic discrete symbol recognition.

Standard decoding strategies (top-k, nucleus) create a 'truncation blind spot' by systematically excluding human-like, low-probability token choices.

EntropyCache achieves up to 26x speedup for Diffusion Language Models by using decoded token entropy as a proxy for KV cache staleness.

AIMER provides a calibration-free criterion for expert pruning in MoE models that matches state-of-the-art performance in seconds.

Mechanistic analysis of 'counting circuits' in VLMs allows for lightweight interventions that improve general visual reasoning performance.

Generative 3D world models are used to scale Sim-to-Real reinforcement learning for robot Vision-Language-Action (VLA) models.

DDPO addresses the 'overthinking' and 'overconfidence' issues in Large Reasoning Models (LRMs) by optimizing answer length based on task difficulty.

Synthetic data scaling reaches a new level by moving from simple rephrasing to creating 'megadocs' through rationale insertion and stitching.

SINDy-KANs combine Kolmogorov-Arnold Networks with Sparse Identification of Non-linear Dynamics to create parsimonious, interpretable models.

SpecForge provides an open-source framework and high-quality draft models (SpecBundle) to make speculative decoding production-ready.

The legally mandated right to be forgotten (unlearning) can be weaponized as an adversarial attack surface to collapse model accuracy.

Learning to Self-Evolve (LSE) trains LLMs to explicitly improve their own context at test-time via reinforcement learning.

OpenT2M is a massive open-source motion dataset (2,800+ hours) that addresses the data starvation in text-to-motion generation.

REST transforms the zero-shot object-navigation problem from simple waypoint selection to a tree-of-paths reasoning process.

AFS-Search introduces a training-free closed-loop framework to solve spatial grounding errors in diffusion models like FLUX.1.

Enables high-fidelity 3D satellite surface reconstruction in a single forward pass without per-scene optimization.

Matches the performance of the complex SFT+GRPO reasoning pipeline for Vision-Language Models in 1/7th of the training time.

Introduces Action Applicability Policy Optimization to train MLLMs to strategically construct and update visual aids to solve geometry problems.

A linear-time attention mechanism that is weight-compatible with standard pretrained Transformers, allowing for direct knowledge transfer.

Disproves the common assumption that bottom models in Vertical Federated Learning effectively represent private labels.

A system where agents autonomously design, refine, and store task-specific skills as 'stateful prompts' to achieve non-parametric continual learning.

Demonstrates that PPO-style clipping and policy ratio constraints are unnecessary for improving reasoning in Large Language Models.

Shifts concept unlearning in diffusion models from fragile keyword-based removal to a distributional framework using contextually diverse prompts.

Introduces explicit spatial tokens (segmentation/depth) into the autoregressive sequence of LVLMs to enable precise 3D/2D grounding.

Provides a mathematically grounded, efficient offline policy optimization method for Diffusion LLMs by estimating trajectory probabilities with a single forward pass.

Automates the entire robot training pipeline by using video generation models as motion priors to synthesize both simulation environments and expert trajectories.

Uses a lightweight GRPO-trained policy to select optimal video frames, reducing processing time by 93% while actually improving Video QA accuracy.

Eliminates the need for expensive process reward models by propagating terminal rewards across state-space graphs to generate dense, state-level rewards for agentic RL.