adu2021
- 1.2k skills
- 0 followers
- 1 day ago last updated
- ▌ Ttt E2e Long Context · adu2021Enable long-context modeling via test-time training with meta-learning. Inner loop continues training on context, compressing information into weights rather than KV cache, outer loop optimizes initialization—maintaining full-attention quality with RNN-like constant inference latency across 8K-128K token contexts.
- ▌ Turbodiffusion Video · adu2021Achieve 100-200× video generation speedup via algorithm-system co-optimization. Combines sparse attention acceleration (SageAttention + trainable Sparse-Linear Attention), step distillation, W8A8 quantization, and custom CUDA kernels—maintaining quality through modular optimizations that compose for cumulative gains.
- ▌ Vfm Visual Tokenizer · adu2021Use frozen vision foundation models like DINOv2 and CLIP as image tokenizers for autoregressive generation. Region-adaptive quantization identifies semantically coherent areas and reduces redundancy. Achieves 256-token encoding (vs. 576), 3× AR model speedup, state-of-the-art 1.36 gFID on ImageNet while eliminating classifier-free guidance.
- ▌ Longcat Flash Prover · adu2021Integrate agentic tool interaction (Lean4 compiler, syntax checkers) with curriculum-based RL for formal reasoning. Replace standard importance sampling with Hierarchical Importance Sampling Policy Optimization (HisPO): sequence-level masking removes train-inference discrepancies, token-level masking filters inconsistent tokens, staleness control manages policy drift. Achieves 97.1% auto-formalization (vs 83% baseline), 95.5% MiniF2F-Test (72 attempts vs 1,024+), 70.8% ProverBench.
- ▌ Safe Flow Q Learning · adu2021Train offline safe RL agents using Hamilton-Jacobi reachability principles to learn feasibility-gated policies. Combine reward and safety critics with flow-matching teacher policies, distill to one-step actors, and calibrate safety thresholds via conformal prediction—achieving near-zero constraint violations with 2.5× inference speedup.
- ▌ Adversarial LLM Judge · adu2021Uncover and fix reward hacking vulnerabilities in LLM-based judges. Simple tokens like punctuation or generic reasoning phrases trigger false positive rewards without substantive content. Defend using data augmentation with truncated model outputs as adversarial negatives, creating robust Master Reward Models resistant to superficial inputs.
- ▌ Aria Intention Reward · adu2021Reduce policy gradient variance in language agent training by aggregating rewards in semantic intention space, enabling 9.95% average performance gains across downstream tasks without exponential action space explosion.
- ▌ Autoregressive U Nets · adu2021Autoregressive U-Net operating directly on raw bytes with hierarchical multi-scale pooling for adaptive token embedding, eliminating fixed vocabularies.
- ▌ Coact 1 Coding Agents · adu2021Hybrid multi-agent architecture where orchestrator delegates tasks to GUI Operator or Programmer agent. Coding enables efficiency on computational tasks, achieving 60.76% on OSWorld with 33% fewer steps.
- ▌ Coda Dual Brain Agent · adu2021Train dual-brain agents with specialized planner (Cerebrum) and executor (Cerebellum) through decoupled RL, resolving planning-execution trade-off for scientific GUI agents
- ▌ Complex Image Editing · adu2021Decompose complex image editing instructions into simpler sub-tasks with automatically generated control guidance. Handles multi-object edits, preserves identity of surrounding regions, and eliminates manual mask creation.
- ▌ Dmlr Latent Reasoning · adu2021Improve multimodal reasoning at test-time through confidence-guided latent optimization without retraining. Iteratively refine learnable latent think tokens via policy gradient using confidence reward. Dynamically select and update relevant image patches based on internal confidence levels. Maintain high efficiency with all optimization in latent space.
- ▌ Eager Entropy Scaling · adu2021Monitor token-wise entropy to adaptively allocate compute during inference. Branch into multiple paths at high-entropy tokens, reducing token generation by up to 65% while improving accuracy by up to 37% on reasoning tasks.
- ▌ Evocua Agent Learning · adu2021Train autonomous agents to use computers by generating synthetic task experiences and iterating on them, achieving 56.7% success on OSWorld benchmarks through scalable experience-driven optimization. Use when you need agents that autonomously learn complex computer interaction patterns without manual task curation.
- ▌ Exp Bench AI Research · adu2021Evaluate AI systems' ability to conduct autonomous research experiments using EXP-Bench, a benchmark for multi-step scientific reasoning and iterative experimental workflows.
- ▌ Fission Grpo Tool Use · adu2021Train small models to recover from tool execution errors by treating errors as training signals. Use error simulators to generate recovery examples and fission failed trajectories into multiple recovery attempts.
- ▌ Fs Researcher Scaling · adu2021Scale research agent capability using persistent filesystem as external memory. Dual-agent architecture with context builder accumulating knowledge and report writer composing outputs enables computation scaling beyond context windows.
- ▌ Geometry Grounded Vlm · adu2021Extend vision-language models with 3D spatial understanding by adding geometric expert stream alongside semantic expert: predict pixel-aligned 3D point maps, surface normals, and camera poses from 2D images, enabling unified reasoning across 2D semantic and 3D geometric domains.
- ▌ Gigaevo LLM Evolution · adu2021Evolve Python algorithms and programs using LLMs as mutation operators combined with MAP-Elites quality-diversity search, achieving competitive results on geometric optimization and algorithmic problems by iteratively mutating code informed by historical performance and lineage context.
- ▌ Gui Test Time Scaling · adu2021Improve GUI agent planning and action grounding through test-time scaling and reinforcement learning. Sample and evaluate multiple action candidates, then use RL to precisely target visual interface elements.
- ▌ Hypergraph Memory RAG · adu2021Build hypergraph-structured memory systems for multi-step RAG that capture high-order relationships between facts, enabling stronger reasoning across long contexts. Use when combining multiple retrieved documents in complex reasoning chains that require understanding connections between pieces of information.
- ▌ Insight O3 Multimodal · adu2021Enable VLMs to perform generalized visual search—locating relational, fuzzy, and conceptual regions from free-form language descriptions. Introduces O3-Bench benchmark with high-density composite charts/maps, uses RL-trained vSearcher for spatial localization, improving frontier models (GPT-5-mini 39%→61.5%) without architecture changes.
- ▌ Lattice 3d Generation · adu2021Semi-structured latent representation combining efficiency of VecSet with spatial structure guidance via voxel queries and rotary positional embeddings, enabling strong test-time scaling (6K to 30K tokens) and improved model scaling without sparse components.
- ▌ Learning 4d Reasoning · adu2021Enable VLMs to perform dynamic spatial reasoning (DSR) by extracting 4D priors from videos and using Geometry Selection Modules (GSM) for selective injection. Provides DSR-Train dataset (50K QA pairs) and benchmark with six reasoning types, balancing geometric specialization with general video understanding—improving VLM 4D reasoning without degradation on general tasks.
- ▌ Livr Visual Reasoning · adu2021Enable richer visual reasoning in multimodal models via learnable latent tokens trained with visual bottlenecking. Introduces special tokens that implicitly learn task-relevant visual abstractions without explicit supervision, through attention masking forcing visual information through latents—improving vision-heavy tasks without task-specific annotations.
- ▌ MCP Bench Tool Agents · adu2021Evaluate LLM agents on realistic tool-use tasks via 28 live MCP servers with 250 tools, assessing fuzzy tool discovery, multi-step planning, and cross-domain workflow coordination
- ▌ Moe Routing Alignment · adu2021Prevents MoE router instability during RL training by recording and replaying inference-phase routing distributions back into training. Reduces training-inference routing divergence and KL divergence, enabling stable MoE RL scaling without sacrificing training speed.
- ▌ Opendev Coding Agents · adu2021Designs terminal-based AI coding agents through workload-specialized model routing, where distinct models handle planning, thinking, critique, and execution tasks. Implements extended ReAct loop with context compaction and approval gates for safe command execution.
- ▌
- ▌ Pisa Sparse Attention · adu2021Accelerate 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.
- ▌ Plan Budget Reasoning · adu2021Decompose complex queries into sub-questions and allocate computational budgets adaptively based on estimated difficulty, achieving 70% accuracy improvements and 39% token reduction without retraining.
- ▌ Rl Verifiable Rewards · adu2021RLVR extends reasoning capabilities by proving answer-only rewards implicitly incentivize correct intermediate reasoning via the Logic Prior principle.
- ▌ Stagevar Acceleration · adu2021Accelerate visual autoregressive (VAR) image generation 3.4× without retraining by analyzing generation stages. Exploits semantic irrelevance in detail-refinement stages where classifier-free guidance becomes redundant and features exhibit low-rank structure—enabling dimensionality reduction while preserving output quality.
- ▌ Starflow Latent Flows · adu2021Learn to implement Transformer Autoregressive Flows for efficient high-resolution image synthesis using latent space normalization and maximum likelihood training.
- ▌ Svg T2i Vfm Diffusion · adu2021Train text-to-image diffusion models directly in frozen DINOv3 feature spaces, eliminating VAE-based compression. Enables high-resolution synthesis by leveraging VFM representations as native latent manifolds with unified cross-modal transformers.
- ▌ Temporal Diffusion Lm · adu2021Leverages temporal dynamics in diffusion models by aggregating predictions across denoising steps for improved inference quality without retraining.
- ▌ Test Driven AI Agents · adu2021Applies test-driven development to agent prompts by iteratively refining prompts against behavioral test suites until compliance is achieved. Enables measurable agent behavior validation through semantic mutation testing and specification evolution.
- ▌ Toolsafe Agent Safety · adu2021Detects and prevents unsafe tool invocations in LLM agents through proactive step-level guardrails, reducing harmful tool calls by 65% while improving task success rates by 10% against prompt injection attacks.
- ▌ Vla 4d Spatiotemporal · adu2021Enhance VLA models with spatiotemporal awareness by embedding both 3D spatial coordinates and temporal sequences: predict actions that include temporal parameters (duration) alongside spatial movements, achieving 97.4% robotic manipulation success by grounding reasoning in coherent 4D representations.
- ▌ Youtu Agent Framework · adu2021Build scalable LLM agent frameworks with automated generation of tools, prompts, and configurations, plus continuous optimization through experience accumulation and reinforcement learning—achieving 71.47% on WebWalkerQA and 72.8% on GAIA.
- ▌ Look Where It Matters · adu2021Add tool-calling interface for on-demand high-resolution crop retrieval, enabling VLMs to first observe low-resolution global view then selectively request detailed crops. Achieves 80.3% of full-res performance (vs 80.46%) with only 36% of visual tokens, reducing wall-clock latency 4.4× (from 2.71s to 0.61s) through KV-cache reuse without architectural changes.
- ▌ Ace Confidence Penalty · adu2021Asymmetric Confidence-aware Error Penalty (ACE) dynamically penalizes overconfident mistakes in RL training, improving reasoning quality without requiring additional computation.
- ▌ Agent Early Experience · adu2021Bridge imitation learning and experience-driven RL by collecting state-based supervision from agents' own actions. Trigger: improve agent generalization when expert demonstrations are limited and environments lack dense rewards.
- ▌ Agent R1 End To End Rl · adu2021Train LLM-based agents with end-to-end RL by extending MDPs to handle tool invocation and environmental stochasticity—enable dense process rewards for intermediate steps and masked policy gradients for learnable actions.
- ▌ Composition Rl Prompts · adu2021Convert easy, high-accuracy training prompts into harder compositional problems by sequentially chaining multiple prompts together. Use Composition-RL to maintain effective learning signals during RL training when many prompts achieve near-perfect accuracy, enabling curriculum learning through progressive compositional depths.
- ▌ Cot Dataset Cold Start · adu2021Enable open-source reasoning model development with a 100K-instance Long CoT Collection, scaling from 1K o1 seed samples through guided synthesis with GPT-4o, achieving 2-3× RL performance gains.
- ▌ Cove Tool Use Training · adu2021CoVe synthesizes high-quality tool-use training data using explicit task constraints as both generation guidance and verification validators, enabling effective agent training without manual curation.
- ▌ Cua Skill Computer Use · adu2021Build desktop agents via reusable, parameterized skills encoding human computer-use knowledge. Skills combine execution graphs (handling UI variations) with composition graphs (chaining strategies). 57.5% success on WindowsAgentArena.
- ▌ Darwin Godel Evolution · adu2021Enable autonomous agent self-improvement through evolutionary mutation of agent codebases, using LLM-generated variants and empirical validation to discover beneficial modifications like enhanced tools and context management.
- ▌ Deepseek V3 2 Frontier · adu2021Combines DeepSeek Sparse Attention (DSA) achieving O(Lk) complexity, scalable RL framework allocating 10%+ compute to post-training, and large-scale agentic task synthesis with 1,800+ environments. Matches proprietary models in reasoning; DeepSeek-V3.2-Speciale achieves olympiad gold.
- ▌ Dppo Divergence Policy · adu2021Replace PPO's heuristic ratio-based clipping with Divergence Proximal Policy Optimization (DPPO) that directly constrains policy divergence using either Total Variation or KL, enabling lightweight approximations (Binary, Top-K) for vocabulary-scale computations while improving stability and efficiency.
- ▌ Dr LLM Dynamic Routing · adu2021Use per-layer routers trained with Monte Carlo Tree Search to dynamically skip, execute, or repeat transformer layers for each token. Saves ~5 layers per example while maintaining accuracy on diverse benchmarks.
- ▌ Duetsvg Multimodal Svg · adu2021Generate SVGs through simultaneous image and SVG token generation with internal visual guidance. DuetSVG overcomes text-only limitations by leveraging visual predictions to enhance SVG coherence—ideal when visual quality and geometric correctness matter.
- ▌ Dynamic Concept Models · adu2021Implement hierarchical language modeling that compresses variable-length token sequences into high-capacity semantic concepts, achieving +2.69% benchmark improvements while reducing inference FLOPs by reallocating compute to concept-level reasoning. Use for efficiency-critical deployments where reasoning quality can be improved while maintaining computational budget.
- ▌ Essential Web Taxonomy · adu202124-trillion-token web dataset with 12-category taxonomy labels enabling efficient curation of specialized datasets through SQL filtering and small annotation models.
- ▌ Grao Unified Alignment · adu2021Unifies supervised fine-tuning and reinforcement learning through GRAO framework that combines multiple-output generation with group direct alignment loss for improved preference learning.
- ▌ Knot Forcing Animation · adu2021Enable real-time portrait animation via causal AR video generation with temporal coherence. Sliding window with global reference frame caching, temporal knot module overlapping adjacent chunks, global context running ahead—preventing error accumulation while maintaining streaming efficiency and bidirectional-quality consistency.
- ▌ Latent Thoughts Tuning · adu2021Enable models to reason in continuous latent space via context-prediction fusion, combining hidden state context with vocabulary embeddings to maintain scaling across model sizes.
- ▌ Lavit Visual Reasoning · adu2021Improves vision-language model distillation by aligning latent visual reasoning trajectories between teacher and student, enabling 3B parameter models to outperform larger open-source and proprietary systems with +16.9% gains on reasoning tasks.
- ▌ Leank Kv Cache Pruning · adu2021Learning-based approach to reduce key-value cache memory during inference through static channel-wise sparsity masks. Achieves 70% K cache and 16-18% V cache memory reduction with 1.3x attention speedup.
- ▌
- ▌ Magistral Reasoning Rl · adu2021Build reasoning capabilities through pure RL without distilled traces, achieving 50% AIME accuracy improvement via scalable asynchronous training with novel reward shaping for multilingual consistency.
- ▌
- ▌ Modomod Multimodal Rl · adu2021Optimize data mixtures across diverse vision-language domains when applying RL with verifiable rewards to multimodal LLMs, balancing task-specific performance with generalization.
- ▌ Moe Sparsity Reasoning · adu2021Determine optimal MoE sparsity by separating memorization and reasoning trade-offs: active FLOPs improve reasoning while total parameters improve memorization, requiring joint optimization
- ▌ Mu Parametrization Moe · adu2021Apply μ-parametrization to Mixture-of-Experts architectures to enable reliable hyperparameter transfer across model sizes, eliminating costly retuning when scaling to trillion-parameter systems.
- ▌ Multi Task Grpo Robust · adu2021Enable balanced multi-task GRPO training via robustness-aware optimization and improvement-aware task reweighting, dynamically adjusting task weights based on both reward and loss trajectory improvement, achieving 6-28% worst-task improvements while maintaining competitive average accuracy.
- ▌ Nag Diffusion Guidance · adu2021Apply training-free negative guidance in diffusion models by extrapolating in attention space with L1-based normalization, restoring suppression of unwanted attributes across architectures and modalities.
- ▌ Over Searching Control · adu2021Diagnose and prevent over-searching—unnecessary search invocations that degrade LLM performance and waste tokens. Framework introduces Tokens Per Correctness (TPC) metric and OverSearchQA dataset to train models that know when NOT to search. Teaches agents to recognize answerable vs. unanswerable queries and selectively invoke search only when needed, improving both accuracy and efficiency.
- ▌ Primo R1 Vision Critic · adu2021Transform 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.
- ▌ Puzzle Curriculum Grpo · adu2021Post-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.
- ▌ Radial Attention Video · adu2021Accelerate video diffusion models using sparse radial attention that exploits energy decay patterns. Achieves 3.7× speedup on long videos while maintaining quality through O(n log n) complexity instead of O(n²).
- ▌ Reasoning Via Sampling · adu2021Improves base model reasoning through iterative sampling without training or fine-tuning. Uses MCMC-inspired sampling to extract latent reasoning from pretrained models, achieving RL-comparable gains on math, coding, and QA tasks while preserving diversity.
- ▌ Rl On Pretraining Data · adu2021Scale LLM training using RL on unlabeled pre-training corpora without human annotation. Derive reward signals directly from text segments to optimize both autoregressive generation and in-context reasoning across knowledge and mathematical domains.
- ▌ Rlkv Cache Compression · adu2021Use reinforcement learning to identify reasoning-critical attention heads and apply selective KV cache compression, reducing cache by 20-50% while preserving reasoning quality. Maintains speedups up to 1.21x with minimal performance loss.
- ▌ Rstar2 Agent Reasoning · adu2021Train efficient 14B-parameter agents via GRPO with resample-on-correct rollout strategy and Python code execution, achieving state-of-the-art reasoning through autonomous exploration and iterative refinement
- ▌ Saffron Safety Scaling · adu2021Implement Saffron's multifurcation reward model approach to achieve efficient inference-time safety scaling, improving robustness against prompt injection attacks while reducing computational overhead.
- ▌ Sage Long Video Agents · adu2021Enable agents to reason over long videos through multi-turn reasoning for complex questions and single-turn answering for simpler queries. Equip orchestrator VLM with multiple tools (web search, speech transcription, temporal grounding). Use synthetic data generation and multi-reward RL achieving 6.1% improvements on open-ended tasks.
- ▌ Spice Corpus Self Play · adu2021Enables continuous self-improvement through corpus-grounded self-play. Challenger mines difficult examples from document corpus for Reasoner to solve. External corpus prevents task stagnation that plagues closed-loop self-play. Achieves 8.9% math, 9.8% general reasoning improvements.
- ▌ Standing Committee Moe · adu2021Reveal that Mixture-of-Experts models harbor a 'Standing Committee' of consistent expert coalitions handling majority computational load across domains. Challenges specialization assumptions and suggests training approaches like load-balancing losses may work against natural optimization.
- ▌ Stepsize Learning Rate · adu2021Improve training efficiency under iteration budgets using the Unified Budget-Aware (UBA) schedule, a theoretically grounded learning rate approach governed by a single hyperparameter φ that balances adaptability across network architectures.
- ▌ Token Order Prediction · adu2021Improve LM performance with token order prediction (TOP) auxiliary loss using learning-to-rank instead of exact multi-token prediction, achieving gains across math, code, and NLP tasks
- ▌ Tool Agent Calibration · adu2021The Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents. From arXiv:2601.07264
- ▌ Treecua Gui Automation · adu2021Scale GUI automation by organizing trajectories into tree structures for reuse and branching exploration, reducing data cost while maximizing step-level diversity through adaptive topology.
- ▌ Web Agent Reward Model · adu2021Create principle-guided reward models for web automation agents that evaluate reasoning process quality rather than just outcomes. Implement domain-specific principles (HTML understanding, interaction patterns, state tracking) to guide agent behavior in web navigation and task completion.
- ▌ Zip2zip Adaptive Vocab · adu2021Reduce token count by 15-40% at inference through context-adaptive compression, merging frequent token sequences into hypertokens using online Lempel-Ziv-Welch compression without retraining entire models.
- ▌ 2xplat Two Expert 3dgs · adu2021Replace monolithic 3D Gaussian Splatting with two-expert architecture separating geometry (pose) estimation from appearance synthesis. Converges 30× faster (5K vs 150K iterations) while matching pose-dependent methods. Works best for multi-view reconstruction when geometry and appearance have conflicting optimization dynamics. Trigger: When doing feed-forward 3DGS and need faster convergence without sacrificing quality.
- ▌ Adversarial Flow Models · adu2021Combines adversarial training with optimal transport constraints enabling deterministic, transport-optimal one-step image generation with FID 2.38 on ImageNet-256. Apply when you need fast generative models without teacher-student distillation complexity.
- ▌ Agentic Search Async Rl · adu2021Enables long-horizon agentic search extending beyond 100 tool calls through scalable asynchronous RL training with autonomous QA dataset synthesis.
- ▌ Arc Encoder Compression · adu2021Reduces inference cost by compressing context into continuous representations using a separate encoder. Generates 4-8x fewer representations than token embeddings while maintaining model performance. Works with any decoder LLM without modification or fine-tuning.
- ▌ Cartridges Long Context · adu2021Train reusable pre-computed KV cache representations of large text corpora for efficient retrieval, achieving 38.6x memory reduction and 26.4x throughput improvement.
- ▌ Codev Verilog Reasoning · adu2021Generate Verilog hardware code from natural language using reasoning-enhanced LLMs, combining rule-based testbench generation with round-trip data synthesis and adaptive DAPO reinforcement learning for reliable hardware design.
- ▌ Concerto Joint Learning · adu2021Learns richer spatial representations by training on both 2D and 3D data simultaneously. Combines 3D intra-modal self-distillation with 2D-3D cross-modal joint embedding, achieving 14.2% and 4.8% improvements over single-modality baselines in scene understanding and geometric consistency.
- ▌ Continue Thinking Token · adu2021Train a specialized continue-thinking token via reinforcement learning to extend reasoning depth during inference, achieving superior performance over fixed-token baselines.
- ▌ Critique Grpo Reasoning · adu2021Improve LLM reasoning by combining numerical and natural language critique feedback in online RL for policy refinement.
- ▌ Deep Forcing Long Video · adu2021Maintains half of sliding window as attention sinks with dynamic temporal RoPE alignment plus importance-aware KV cache pruning, enabling 12× extrapolation beyond training length (60+ seconds from 5-second training) without fine-tuning.
- ▌ Deepcode Agentic Coding · adu2021Transform research specifications into production-grade codebases through strategic information management and autonomous agent orchestration. DeepCode surpasses PhD experts and commercial tools—critical when you need scientific code reproducibility at scale.
- ▌ Densegrpo Flow Matching · adu2021Improve diffusion model alignment by assigning step-wise rewards during denoising instead of terminal rewards. Fixes sparse reward signal mismatch in multi-step generation processes through ODE-based reward estimation.
- ▌ Gere Continual Learning · adu2021Prevents catastrophic forgetting in continual LLM learning using threshold-based margin loss with fixed general replay samples from pretraining data.
- ▌ Group Rank Reranking Rl · adu2021Improve ranking quality via groupwise reranking with RL—process document groups jointly for within-group comparisons using GRPO with composite rewards (recall, ranking metrics, distribution alignment).