Claude Skills by ADu2021

Optimize multimodal LLMs by directly targeting perception errors using KL-divergence based perception loss, improving visual reasoning by 8-19% on vision-dependent tasks. Integrates perception-aware signals into policy gradients without relying solely on reward modifications.

Enable language models to reason over extremely long contexts (128K tokens) by encoding context into lightweight LoRA adapters during test time, achieving 20% performance improvements without full model retraining.

Method to identify and manipulate interpretable directions in model activation space corresponding to personality traits. Enables real-time monitoring of trait drift and preventive steering to mitigate unwanted behavioral shifts in language models.

Optimize sampling budget allocation for self-consistency inference by treating trajectory allocation as a crowdsourcing problem. Introduce self-consistency rate (agreement with infinite-budget consensus) as optimization target. Offline setting uses Optimistic Knowledge Gradient for Bayesian optimization; online setting uses difficulty grid discretization and greedy allocation. Achieves up to 75% budget reduction (offline) and 55% (online) while maintaining accuracy.

Physical principles are fundamental to realistic visual simulation, but remain a significant oversight in transformer-based video generation. This gap highlights a critical limitation in rendering rigid body motion, a core tenet of classical mechanics. While computer graphics and physics-based simulators can easily model such collisions using Newton formulas, modern pretrain-finetune paradigms discard the concept of object rigidity during pixel-level global denoising. Even perfectly correct m...

Apply reinforcement learning to flow-based VLA models by modeling denoising as an MDP and converting ODEs to SDEs, enabling efficient policy optimization for robotic control without expensive supervised fine-tuning data collection.

Accelerate diffusion transformers through training-free sparse attention combining exact computation for critical blocks with efficient approximation for non-critical ones. Achieves 2-3× speedup without retraining.

Enhance Masked Autoencoders through three algorithmic improvements: deeper decoders enabling semantic-focused encoding, larger masking blocks providing richer context, and multiple class tokens capturing diverse global properties. Combine with 2B web-crawled images and soft self-curation for competitive downstream performance.

Decompose complex queries into sub-questions and allocate computational budgets adaptively based on estimated difficulty, achieving 70% accuracy improvements and 39% token reduction without retraining.

Develop reasoning capabilities in multimodal models through interactive gameplay rather than supervised learning, achieving superior transfer to diverse reasoning tasks.

Generate spatially and temporally coherent multi-view video through autoregressive conditioning with camera-guided retrieval and progressive context scaling. Enable long-video generation maintaining spatio-temporal memory across viewpoint changes.

Improve LLM training stability and generalization by reparameterizing weight matrices as orthogonal transformations, achieving better perplexity than AdamW with fewer trainable parameters.

Use language models as optimizers to iteratively improve complex systems (prompts, code, agent policies) through noisy feedback and sampling. POLCA maintains a priority queue of candidates, explores with epsilon-nets, and learns meta-insights from trial history.

Learn generalizable reward models via unsupervised policy discrimination: pretraining models to distinguish between different policies enables efficient adaptation to human preferences and strong RLHF performance.

Remove deep layers during context encoding (prefill) while keeping them for token generation (decode). Identifies layer importance asymmetry via virtual gates; achieves 1.37x prefill speedup without retraining on any pre-trained model.

Improve speculative decoding throughput by employing position-specialized draft layers that handle position-specific error accumulation patterns.

Train LLM agents via multi-turn reinforcement learning by systematically optimizing environment complexity, reward signals, and policy initialization. Use curriculum learning, dense verified rewards, and domain-specific SFT for reliable agent convergence across TextWorld, ALFWorld, and SWE-Gym benchmarks.

Optimize text-to-image models using pairwise preference comparisons instead of pointwise rewards, eliminating normalization instability and reward hacking while providing fine-grained image quality differentiation

Implement Prefix Grouper to accelerate Group Relative Policy Optimization training by eliminating redundant prefix encoding, achieving up to 8x speedup for long-context scenarios.

Transform lengthy documents into fully narrated presentation videos with synchronized audio-visual delivery. Automatically segments content, generates visuals, synthesizes speech, and composes final video.

Bilevel min-max optimization where mask generator selects informative spans from pretraining data and mask predictor recovers them via chain-of-thought, enabling effective RL pretraining on noisy corpora without supervised fine-tuning.

Improve pretraining efficiency by refining noisy data through expert-guided programs: learn to generate deletion operations that clean documents, achieving 2.6-7.2% performance gains with fewer training tokens.

Understand when RL genuinely expands reasoning beyond pre-training through controlled experiments on synthetic tasks. Discover that RL works best at the edge of competence and process rewards reduce hacking—critical for designing effective reasoning model training.

Transform video multimodal models into active process critics for robotic tasks. Use RL to incentivize explicit reasoning about progress toward goals and anchor reasoning temporally between initial and current states.

Phone recognition (PR) serves as the atomic interface for language-agnostic modeling for cross-lingual speech processing and phonetic analysis. Despite prolonged efforts in developing PR systems, current evaluations only measure surface-level transcription accuracy. We introduce PRiSM, the first open-source benchmark designed to expose blind spots in phonetic perception through intrinsic and extrinsic evaluation of PR systems. PRiSM standardizes transcription-based evaluation and assesses dow...

Scale inference efficiency for discrete diffusion language models through hierarchical trajectory search with adaptive pruning and self-verified feedback. Achieve 3-4× speedup versus best-of-N with equal quality.

Unify semantic understanding and pixel-level detail in a single representation by decomposing features into frequency bands. Low frequencies encode semantics while high frequencies capture pixels—enabling one tokenizer for both understanding and generation through frequency-based modulation and semantic-wise alignment.

Optimize multi-step reasoning by treating candidate solutions as particles in a process-reward energy landscape. Use PRM step-level scores to guide stochastic refinement and population resampling, achieving directional error correction without hallucination amplification.

Leverage training-time privileged information (depth, saliency maps) to improve student detector performance without inference overhead. Model-agnostic methodology applicable across detection architectures with no increase in inference complexity.

Improves LLM reasoning by decomposing RL objectives into intermediate process rewards assigned to reasoning steps, improving both final accuracy and reasoning capacity without expensive Monte Carlo Tree Search.

Learn to ground LLM agent planning in real environment dynamics using Monte-Carlo Tree Search exploration combined with lightweight Monte-Carlo critics, reducing hallucination-driven planning failures in interactive tasks.

Enable models to refine outputs dynamically during generation based on internal signals, reducing token consumption by 41.6% while improving accuracy by 8.2%.

Apply semantic understanding to 3D Gaussian Splatting scenes through dense correspondence-guided pre-registration without render-supervised fine-tuning. Achieve semantic understanding in ~5 minutes using cross-view clustering and direct language feature fusion.

Improves LLM convergence and downstream task performance by introducing time-dependent scaling to residual connections, enabling shallow layers to learn first before deeper layers activate. Apply during model pretraining to achieve 0.4-4.86 perplexity reduction.

Use component-based markup with CSS-like styling to structure complex prompts, integrate diverse data types, and separate content from formatting for maintainable, version-control-friendly LLM applications.

Automatically discovers optimal in-context learning prompts through evolutionary token pruning that removes redundant demonstrations to create effective 'gibberish' prompts. Matches state-of-the-art optimization with low-data regimes. Use for automated prompt discovery without manual tuning or human expertise.

Detect when diffusion language models converge on correct answers before completing refinement steps using confidence gap monitoring, achieving 3.4x decoding speedup

Train LLMs to discover novel reasoning strategies beyond base model capabilities using prolonged RL with KL control, reference policy resets, and diverse task suites.

Replaces binary keep/drop masks with multi-level pooled key-value representations, allowing queries to access larger receptive fields under same compute budget through hierarchical aggregation without discarding information.

Post-train vision-language models using automatically-verifiable puzzle environments (Jigsaw, Rotation, PatchFit) with graded rewards. Implement exploration-aware curriculum combining difficulty weighting with solution-space diversity metrics. Track reasoning-answer consistency to prevent divergence during training.

Enable identity-aware video question answering with one-shot learning using Mixture-of-Heads enhanced ViLLM. Learns subject-specific features from single video through synthetic augmentation and progressive image-to-video training, enabling recognition of individuals in medical, smart home, and entertainment contexts.

Convert pretrained video diffusion models into pyramidal architectures via low-cost finetuning while preserving output quality. Explore step distillation for enhanced efficiency, enabling deployment of efficient inference without training from scratch.

Enable multimodal language models to autonomously generate and execute Python-based tools during visual reasoning, boosting performance on vision benchmarks by up to 31% through interactive problem-solving without relying on predefined tool sets.

Dramatically reduce training data requirements (to 12.5% of original) while improving model performance using joint sample and token pruning guided by Error-Uncertainty plane diagnostics. Asymmetric pruning preserves calibration signals while removing redundant tokens from misconception examples.

Combine NFVP4 quantization with LoRA to accelerate RL rollout phases while using quantization noise as implicit exploration bonus. Achieve 1.5x speedup and better strategy discovery through noise-enhanced policy entropy.

Stabilize LLM reasoning training by replacing mean-based advantage baselines with K-quantile baselines, preventing both entropy collapse and explosion while improving performance on mathematical benchmarks through response-level gating and asymmetric sample weighting.

Fine-tune quantized LLMs directly in low-precision discrete parameter space using evolution strategies with accumulated error feedback. Overcome gradient stagnation in quantized models by accumulating fractional updates using Delta-Sigma modulation, achieving significant improvements in INT4 quantized models without full-precision gradients.

Replace unreliable model-internal confidence signals with objective corpus statistics to decide when RAG retrieval is necessary. Pre-evaluates entity rarity in training data and verifies entity co-occurrence at runtime, triggering retrieval only when hallucination risk is high—improving reliability without per-model tuning.

QueryBandits adaptively learns per-query rewriting strategies to reduce LLM hallucinations, achieving 87.5% improvement without model retraining.

Leverages Qwen3 foundation models for text embedding and reranking via multi-stage training combining weakly-supervised pre-training on 150M synthetic pairs with supervised fine-tuning.