The Archive

Theoretical analysis of Delightful Policy Gradient, which gates advantage-based updates to escape suboptimal policy corners faster than softmax policy gradient.

Empirical study of supervised fine-tuning on procedural skills across Qwen3.5 0.8B-4B models shows W-shaped pre-training trajectory and uniform SFT gains.

YFPO leverages neuron activation patterns for preference optimization in mathematical reasoning, using model internals to guide reward signals instead of external preference data.

Proposes segment-level supervision for LLM-based Lean 4 theorem proving, balancing dense local signals of step-level training with coherence of whole-proof generation.

Proposes Hierarchical-Cluster SOINN classifier for class-incremental learning that models classes as manifolds rather than point collapse, addressing Neural Collapse theory gaps.

Earphone-based passive biometric authentication system using in-ear accelerometers for heartbeat identification, unrelated to AI frontier topics.

Reddit user benchmarks open-weight LLMs (Qwen 3.6, Zaya1) on chess visualization task; Qwen 35B-A3B achieves near-perfect SVG chessboard generation.

Model for emulating individualized chess player styles and decision-making, addressing limitations of skill-level generalization in superhuman chess engines.

Proteus red-teaming framework studies adaptive leakage of LLM agent skills via iterative adversarial revision, addressing real deployment risks beyond single-shot audits.

Mechanism for collaborative ML ensures fair data valuation and incentivizes truthfulness via game-theoretic rewards.

Qwen-Scope: open-source suite of sparse autoencoders for mechanistic interpretability across 7 Qwen models.

Stratechery opinion: Elon Musk's dual involvement in SpaceX and xAI mirrors Anthropic's structure; argues Musk should focus xAI on B2B services.

Layer-wise relevance propagation extends attribution methods to EEG-based foundation models for interpretability verification.

Sobolev-regularized MMD gradient flow with global convergence guarantees for distribution matching.

FATE: on-policy framework for agentic safety alignment via failure trajectory learning without safety-utility trade-off.

Adaptive TD-Lambda extends temporal difference learning to multi-agent RL with large joint action spaces.

Task-aware contrastive learning for automatic modulation classification via intra-instance consistency.

LOFT: parameter-efficient fine-tuning framework using low-rank orthogonal transformations with task-aware subspace selection.

Information-theoretic foundation for self-supervised clustering via KL-divergence optimization with mode-collapse constraints.

FIS-DiT: training-free frame interleaving acceleration for video diffusion transformers in few-step inference regimes.

IPI-proxy toolkit enables red-teaming web-browsing AI agents against indirect prompt injection attacks embedded in whitelisted domain HTML.

Anchor-guided variance-aware reward modeling resolves non-identifiability in preference learning by augmenting pairwise comparisons with coarse response-level labels.

ZipRerank achieves efficient listwise multimodal reranking for long documents by reducing visual tokens and eliminating autoregressive decoding.

Concordance comparison method assembles local grammars for Portuguese named entity recognition via pairwise grammar analysis.

EvoNav uses LLMs to evolve reward functions for robot navigation via reinforcement learning, automating design of navigation policies.

Semantic consensus framework enables federated fine-tuning of LLMs through model behavior rather than parameter aggregation, supporting heterogeneous architectures.

UniVLR unifies text and vision in multimodal LLMs by rendering reasoning traces as shared visual workspace, improving latent reasoning efficiency.

Bistable Memory Recurrent Units (BMRU) improve gradient flow and stability for ultra-low power sequence models via hardware-software co-design.

Self-distilled trajectory-aware Boltzmann modeling closes training-inference gap in diffusion language models via multi-step denoising trajectories.

GEAR enables fine-grained credit assignment in RL-trained LLM agents via adaptive-granularity advantage reweighting at token and segment levels.