Paradigm Shift

Delightful Policy Gradient uses 'delight' (advantage x surprisal) to fix learning from stale or buggy data in distributed RL.

Continued Fraction Neural Networks (CFNN) introduce a rational inductive bias that handles singularities with 10-100x fewer parameters than standard MLPs.

Network-of-Thought (NoT) moves LLM reasoning from linear chains and trees to complex directed graphs, significantly improving multi-hop QA.

Proposes 'semantic sections' as a replacement for global feature vectors to interpret LLMs in complex, non-linear representation spaces.

Introduces Bayesian scattering as a mathematically grounded, non-learned baseline for image uncertainty quantification.

A red-teaming protocol that uses RL-driven 'profit' objectives to find structural exploits in AI agents instead of just prompt-injection vulnerabilities.

Pretrained Diffusion Transformers (DiTs) possess an intrinsic 'synchronization gap' where different features commit at specific, depth-localized layers.

The 'routing paradox' proves that selective attention requires the very pairwise computations it aims to replace, explaining why pure recurrent models fail at associative recall.

VAE tokenizers in Latent Diffusion Models create 'overly compact' manifolds that cause variance collapse, leading to unstable generative sampling.

CounterScene endows generative world models with explicit counterfactual reasoning for safety-critical driving evaluation.

Proposes multi-cluster memory for test-time adaptation, proving that a single unstructured memory pool is fundamentally insufficient for non-i.i.d. data streams.

Reframes plasticity loss in Reinforcement Learning as an optimization problem where networks get trapped in local optima of previous tasks.

Repurposes a 2B-parameter latent video transformer as a differentiable physics simulator for urban wind flow optimization.

Proposes replacing flat conversation histories with a tree-based architecture to solve 'logical context poisoning.'

Replaces self-attention with Reaction-Diffusion PDEs as the predictive engine for world models.

Reconceptualizes human-agent interaction as dynamically generated software rather than just chat.

ADARUBRIC generates task-specific evaluation rubrics on the fly, significantly outperforming static rubrics in human correlation and agent training outcomes.

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

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.

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.

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

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

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

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

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

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

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

Proposed a test-time scaling paradigm for image restoration that allows compute-to-quality trade-offs during inference.

Identifies that the direction of log-probability change is more critical than magnitude for improving LLM reasoning via RL.

Identifies 'Visual Anchor Collapse' in DPO-aligned VLMs and introduces an asymmetric constraint to prevent models from ignoring visual evidence in favor of language priors.

Bypasses Reinforcement Learning during the exploration phase by using uncertainty-guided tree search to discover informative data.

UNITE enables single-stage joint training of the tokenizer and the diffusion model from scratch, removing the need for frozen VAEs.

LassoFlexNet matches or beats leading tree-based models on tabular data while maintaining Lasso-like interpretability through per-feature embeddings and a group Lasso mechanism.

Introduces a statistical alternative to the standard frequency-based BPE tokenization used in nearly all modern LLMs.

Formally proves that a causal Transformer is mathematically equivalent to a stateless Differentiable Neural Computer.

Solves the compositional generalization failure of neural networks (0% to 100% accuracy) by embedding algebraic semiring constraints.

Challenges the 80-year-old assumption that neurons must use weighted summation as their primary aggregation mechanism.

Introduces Hyperagents: self-referential systems where the meta-level modification logic is itself an editable program.

Fine-tunes Large Vision Language Models for medical tasks using only image-description pairs, bypassing the need for expensive expert-curated instructions.

Formalizes the 'Neural Uncertainty Principle,' linking adversarial vulnerability in vision and hallucinations in LLMs to a shared geometric and information-theoretic origin.

A massive field study (9,000+ users) proves that algorithmic shifts can reduce affective polarization without sacrificing user engagement.

Enables zero-shot humanoid robot interaction by generating robot-centric 'dream' videos instead of relying on human-to-robot motion retargeting.

Replaces fixed context compression ratios with a performance-floor constraint to ensure reliable LLM deployment.

FIPO overcomes reasoning length stagnation in LLMs by using Future-KL divergence to create dense rewards, extending Chain-of-Thought lengths to over 10,000 tokens.

Breaking the 'capability ceiling' in LLM post-training by replacing full-history dependencies with explicit Markov states.

Identifies 'critical times' in diffusion generation where targeted guidance pulses significantly improve image control.

Derives a variational ELBO for the Joint-Embedding Predictive Architecture (JEPA), unifying it with generative modeling.

Integrates Kolmogorov-Arnold Networks (KANs) into causal generative modeling to produce human-readable symbolic structural equations.