adu2021
- 1.2k skills
- 0 followers
- 1 day ago last updated
- ▌ Prorl Reasoning Expansion · adu2021Train LLMs to discover novel reasoning strategies beyond base model capabilities using prolonged RL with KL control, reference policy resets, and diverse task suites.
- ▌ Rank One Safety Alignment · adu2021Apply rank-one weight modifications to amplify model safety via residual stream steering, requiring no fine-tuning and preserving utility on standard benchmarks
- ▌ Reasoning Path Confidence · adu2021Reduce LLM sampling costs by 50% while maintaining reasoning performance through Reasoning Path Confidence (RPC), which combines perplexity-guided pruning with self-consistency sampling.
- ▌ Recursive Language Models · adu2021Process prompts exceeding model context windows by recursively decomposing long inputs into manageable chunks and calling the model recursively on snippets—enabling inference on contexts 100x longer than native window while maintaining quality and improving over vanilla baseline approaches.
- ▌ Reinforcement Pretraining · adu2021Apply reinforcement learning to pre-training by framing next-token prediction as a reasoning task with verifiable rewards, achieving superior scaling compared to standard language modeling.
- ▌ Relay Gen Model Switching · adu2021Reduce inference cost by dynamically switching from large to small LLMs during reasoning generation. Large model handles demanding reasoning phases; small model completes consolidation and answer stages triggered by discourse cues. Achieves 2.2× speedup with minimal accuracy loss.
- ▌ Scale Selective Test Time · adu2021Decomposes math problems into sequential sub-problems, assesses difficulty, and allocates simple ones to fast System 1 reasoning while directing complex ones to deliberate System 2. Save 33-53% tokens while improving accuracy by up to 13.75 points on AIME.
- ▌ See Upo Sequence Level Rl · adu2021Train multi-turn AI agents with convergence guarantees using sequential backward-induction updates, eliminating the need for separate critic networks while maintaining theoretical optimality. Use for long-horizon agentic reasoning where monotonic improvement and global optimality are required.
- ▌ Seedvr2 Video Restoration · adu2021Achieves single-step video restoration at 1080p resolution with 4x speedup over multi-step diffusion approaches via adversarial training, adaptive window attention, and feature matching loss.
- ▌ Simplemem Lifelong Memory · adu2021Implement efficient memory systems for long-term LLM agent interactions using semantic compression, achieving 30-fold inference token reduction while improving F1 scores by 26.4%—enabling agents to learn from extended interaction histories without prohibitive context costs.
- ▌ Sparsemm Visual Attention · adu2021Discovers that <5% of attention heads process visual information in MLLMs; introduces SparseMM for asymmetric KV-cache allocation achieving 1.38x acceleration and 52% memory reduction.
- ▌ Swe Lego Issue Resolution · adu2021Achieve state-of-the-art software issue resolution through lightweight supervised fine-tuning with structured datasets and curriculum learning, plus test-time scaling—reaching 52.6% on SWE-Bench Verified and 58.8% with test-time strategies, outperforming complex multi-stage training.
- ▌ Swe Universe Environments · adu2021Automatically generate executable software engineering environments from GitHub pull requests at million-scale using an autonomous building agent. Detects superficial verification patterns to force genuine code execution testing.
- ▌ Synthagent Web Adaptation · adu2021Adapt web agents to new domains through targeted synthetic data generation and quality-aware refinement—identifying and correcting hallucinations while preserving task consistency to enable efficient adaptation with minimal human supervision.
- ▌ Tldr Thinking Reweighting · adu2021Compress reasoning models by dynamically re-weighting short-CoT (System-1) and long-CoT (System-2) training data, achieving 40% token reduction while maintaining accuracy.
- ▌
- ▌ Ttcs Test Time Curriculum · adu2021Improve model performance at test time by dynamically synthesizing curriculum of problem variants. Co-evolving synthesizer and solver agents create reinforcing feedback for continuous improvement without external labels.
- ▌ Universal Reasoning Model · adu2021Enhance Universal Transformers for complex reasoning through ConvSwiGLU modules integrating depthwise convolution into feed-forward blocks and truncated backpropagation through loops (TBPTL) restricting gradient computation to final iterations. Achieve state-of-the-art on ARC-AGI: 53.8% on ARC-1, 16.0% on ARC-2.
- ▌ Versatileffn Adaptive Ffn · adu2021Enable parameter-efficient computation through dual-pathway feed-forward networks. Create virtual experts via hidden subspace slicing (width-versatile) and recursive weight application (depth-versatile). Use difficulty-aware gating to balance pathways, adding computation not memory to fixed parameter budget.
- ▌ Video Deep Research Agent · adu2021Implements video deep research for multi-hop reasoning combining video analysis, web search, and evidence synthesis. Evaluates workflow vs agentic paradigms with 100-sample benchmark across 6 semantic domains, revealing goal drift and long-horizon consistency as core bottlenecks.
- ▌ Video Reasoning Grounding · adu2021Ground video reasoning in explicit visual evidence by highlighting timestamps, objects, and bounding boxes, making reasoning verifiable and improving accuracy through RL rewards for spatio-temporal alignment.
- ▌ Agent0 Symbiotic Evolution · adu2021Train agents from scratch without human-annotated data via symbiotic competition—curriculum agent proposes progressively harder tasks while executor agent learns to solve them, creating autonomous self-reinforcing loops.
- ▌
- ▌ Answer Matching Evaluation · adu2021Evaluate language models using open-ended answer generation and semantic matching instead of multiple choice, eliminating test-taking shortcuts and achieving near-perfect alignment with human judgment.
- ▌ Atr Grpo Policy Divergence · adu2021Replace ratio-based clipping in GRPO with KL-divergence constraints using the KL3 estimator, improving exploration and training stability with asymmetric clipping that requires no additional computation.
- ▌ Automated Tool Learning Rl · adu2021Improves LLM tool-use capabilities through automated environment construction that generates realistic feedback and verifiable rewards for RL-based training without external tools.
- ▌ Batch Speculative Decoding · adu2021Fixes batch speculative decoding ragged tensor problem where sequences in batches accept different token counts, desynchronizing state. EQSPEC guarantees output equivalence through proper synchronization. EXSPEC reduces overhead 40% via cross-batch scheduling. Enables efficient parallel decoding with 95% equivalence.
- ▌ Beyond Real Imaginary Rope · adu2021Improve long-context performance by incorporating imaginary components discarded in standard RoPE implementations. Use phase information from complex-valued attention for richer positional encoding—especially valuable as context length increases beyond normal ranges.
- ▌ Capmagine Visual Reasoning · adu2021CapImagine teaches models to explicitly imagine through text rather than latent reasoning, significantly improving visual reasoning performance.
- ▌ Clara Retrieval Generation · adu2021Unify retrieval and generation in RAG systems by compressing documents into shared continuous embeddings that serve both retrieval and generation: implement joint training with differentiable selection, achieving up to 16× context compression while improving generation quality.
- ▌ Clasp Layer Skip Inference · adu2021Accelerate LLM inference by dynamically skipping transformer layers based on in-context signals, without training auxiliary draft models or changing model weights.
- ▌ Cogvla Instruction Routing · adu2021Align VLA efficiency with human cognition through 3-stage progressive routing: instruction-aware aggregation, instruction-irrelevant pruning, and coupled attention for 2.8x inference speedup
- ▌ Computer Using World Model · adu2021Enable AI agents to safely explore action outcomes before execution by predicting UI state changes in desktop applications. Two-stage approach: first predict textual description of what changes, then synthesize visual representation of resulting screen. Allows agents to compare multiple candidate actions without risky trial-and-error, trained on Microsoft Office interactions (Word, Excel, PowerPoint).
- ▌ Cosinegate Dynamic Routing · adu2021Achieve efficient neural networks via self-supervised dynamic routing using Cosine Incompatibility Ratio (CIR). Ground gating decisions in geometric novelty rather than learned heuristics, enable per-sample/per-block binary routing via Gumbel-softmax, constrain with progressive FLOPs regularization—maintaining accuracy while reducing computation 28.5% on CIFAR-10.
- ▌ Deep Search Research Agent · adu2021Build research agents that systematically search for comprehensive answers to complex questions by maintaining search state, iterating on queries, and validating answer completeness. Implement strategies for identifying knowledge gaps and conducting follow-up searches to ensure thorough coverage of topics.
- ▌ Direct Denoising Diffusion · adu2021Improve diffusion model capacity efficiency by directly predicting clean data instead of noise—leverage the manifold assumption that natural data occupies low-dimensional space while noise spans full dimensionality.
- ▌ Discreteness Diffusion LLM · adu2021Understand fundamental limitations of applying diffusion to discrete text: position-agnostic corruption ignores linguistic structure, and token-wise training misses multi-token dependencies. Design text diffusion systems satisfying five essential properties: position-aware corruption, dependency-aware training, parallel consistency, linguistic structure respecting, and robust handling of token boundaries.
- ▌ Distilling Feedback Memory · adu2021Convert inference-time feedback into persistent, retrievable guidelines stored as agent memory. Framework enables LLMs to improve performance over time by systematically accumulating and applying learned critiques. Augmented models rapidly match test-time refinement performance while drastically reducing inference cost. Memory-as-tool pattern enables agents to learn from feedback without expensive retraining.
- ▌ Dynamic Fine Tuning Sft Rl · adu2021Minimal modification to SFT that dynamically rescales objectives by token probability. Rectifies implicit reward structure to improve generalization comparable to RL while maintaining SFT simplicity.
- ▌ E2rank Embedding Reranking · adu2021Extends text embedding models to perform listwise reranking through continued training on ranking objectives. Constructs listwise prompts from queries and top-K candidates, leveraging pseudo-relevance feedback while maintaining embedding model efficiency. Unifies retrieval and reranking in single model.
- ▌ Efficient Reasoning Models · adu2021Comprehensive survey of techniques for optimizing large reasoning models. Covers single-model optimization and multi-model collaboration approaches to reduce reasoning path length without sacrificing capability.
- ▌ Entropy Guided Exploration · adu2021Stabilize RL training on reasoning tasks by performing entropy-guided rollouts from uncertain decision points, avoiding policy collapse and premature convergence. Increases fully correct trajectories on math reasoning while maintaining stable entropy throughout training.
- ▌ Farmer Flow Autoregressive · adu2021Generates high-quality images directly from pixels using flow-matching-based latent sequences. Transforms images via invertible flows into manageable latent sequences, applies autoregressive modeling, and uses classifier-free guidance. Provides exact likelihood estimates and one-step distillation capabilities.
- ▌ General Agentic Memory Jit · adu2021Build persistent, lossless agent memory using just-in-time compilation: store complete history in a universal page-store while performing dynamic deep research at query time, enabling test-time scalability through iterative information synthesis and reflection.
- ▌ Grove Moe Adjugate Experts · adu2021Enables efficient MoE architectures through heterogeneous expert sizing and dynamic activation mechanisms that adjust parameter count based on input complexity.
- ▌ Hierarchical Lvm Reasoning · adu2021Reason about long-horizon dynamics by disentangling structure and motion in video VAE latents. Learn continuous latent motion chains that preserve temporal coherence while predicting terminal keyframes, enabling efficient reasoning about multi-step scenarios.
- ▌ Kinetics Test Time Scaling · adu2021Reveals that memory bandwidth—not computation—dominates test-time costs; proposes eFLOPs metric incorporating both computation and memory, showing 14B+ minimum threshold for reasoning value.
- ▌ Kv Embedding Training Free · adu2021Extract high-quality embeddings from frozen decoder-only LLMs by re-routing internal key-value states without training—outperforming training-free baselines by 10% on MTEB while maintaining robustness across sequences up to 4,096 tokens.
- ▌ Learning Reason Factuality · adu2021Multi-dimensional reward function combining factual precision, response detail, and answer relevance for online RL. Reduces hallucinations 23.1% while maintaining helpfulness and detail.
- ▌ LLM Forecasting Evaluation · adu2021Benchmark LLMs on real-world forecasting questions from Metaculus, comparing against human crowds and expert forecasters. Identifies which domains LLMs handle well and where they fall short relative to human intelligence.
- ▌ LLM Memorization Landscape · adu2021Understand and mitigate unintended memorization of training data in LLMs by systematizing mechanisms, detection methods, and mitigation strategies across the model lifecycle.
- ▌ Mano Manifold Optimization · adu2021Improve LLM training efficiency through manifold-based optimization that projects momentum onto tangent spaces and constrains updates on rotational Oblique manifolds. Achieves 1.75× faster convergence than Muon with reduced memory.
- ▌ Mobe Mixture Basis Experts · adu2021Compresses MoE language models through shared basis factorization of expert weight matrices, achieving 24-30% parameter reduction with minimal accuracy loss.
- ▌ Moca Multimodal Embeddings · adu2021Transform pre-trained vision-language models into powerful bidirectional multimodal embeddings through modality-aware continual pre-training and heterogeneous contrastive fine-tuning. 3B model matches 7B baselines.
- ▌ Model Merging Dual Anchors · adu2021Merges multiple fine-tuned models by operating in input-representation space rather than parameter space. Creates synthetic inputs whose gradients align with task vectors, bridging joint training and post-hoc merging for robust multi-task model combination.
- ▌ Multiagent Process Rewards · adu2021Train specialized agents in pipelines using dense per-action process rewards from AI coaching. Solves credit assignment in sequential workflows, enabling better generalization and faster convergence than outcome-only training.
- ▌ Nbdiff Block Diffusion LLM · adu2021Convert auto-regressive language models to efficient diffusion-based generators through gradual block size increments. NBDiff-7B inherits long-context capabilities from AR predecessors while achieving state-of-the-art parallel generation—ideal when you need efficiency without sacrificing reasoning.
- ▌ Nemotron Math Long Context · adu2021Create 7.5M long-form mathematical solution traces with multi-mode supervision (high/medium/low reasoning depths, with/without Python). Integrate 85K competition problems (AoPS) and 262K community questions (StackExchange). Implement sequential bucketing training achieving 2-3× speedup while maintaining accuracy.
- ▌ Overclocking LLM Reasoning · adu2021Extract and manipulate internal progress vectors in reasoning models to accelerate thinking phases while maintaining answer quality, achieving 80%+ improvements in token efficiency.
- ▌ Ovis U1 Unified Multimodal · adu2021A 3B unified model combining image understanding, text-to-image generation, and image editing end-to-end rather than as separate frozen components. Use when you need a single efficient model for multiple vision-language tasks without the overhead of separate specialized systems.
- ▌ Parallelizable Linear Rnns · adu2021Linear RNN architecture enabling efficient parallel computation over directed acyclic graphs with stabilization modes for long-range dependencies without sequential traversal.
- ▌ Plenoptic Video Generation · adu2021Generate 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.
- ▌ Ramp Adaptive Quantization · adu2021Learn optimal per-layer bit-width assignments for LLM quantization via RL, generalizing across models without retraining. Achieves superior compression under fixed bit budgets.
- ▌ Rectified Sparse Attention · adu2021Enable efficient long-sequence generation by combining block-sparse attention with periodic dense rectification to bound error accumulation.
- ▌ Reinforcement Mid Training · adu2021Add an intermediate RL stage between pretraining and post-training using dynamic token budgeting, curriculum sampling, and dual training. Trigger: reduce reasoning steps while maintaining or improving performance in post-training.
- ▌ Repair Robust Editing Llms · adu2021Perform robust, low-cost model updates via REPAIR: closed-loop feedback, dynamic memory management, and frequent knowledge fusion. Preserve non-target knowledge while achieving 10-30% improvements in editing accuracy. Enable sequential edits without catastrophic forgetting through strong locality guards and stable integration mechanisms.
- ▌ Repo Context Repositioning · adu2021Enable language models to dynamically assign continuous position values to tokens based on contextual relevance instead of fixed linear positions. Implement learnable SwiGLU module extracting position information, integrate with RoPE for end-to-end optimization. Improves performance on noisy contexts and long-sequence tasks.
- ▌ Resa Transparent Reasoning · adu2021Extract and transfer reasoning abilities using sparse autoencoders (SAE-Tuning) on CoT-free data, achieving RL-equivalent performance at 2000x lower cost and 450x faster training.
- ▌ Sail Rl Adaptive Reasoning · adu2021Teach multimodal models to determine when deep reasoning is necessary and how to reason effectively through dual-reward reinforcement learning, preventing both overthinking on simple tasks and underthinking on complex ones while reducing hallucinations.
- ▌ Sentinel Prompt Protection · adu2021Deploy a state-of-the-art binary classifier using ModernBERT to detect prompt injection attacks and protect LLMs from adversarial input manipulation.
- ▌ Signround V2 Low Bit Quant · adu2021DeltaLoss sensitivity metric combining gradient and quantization-induced parameter deviation for adaptive bit-width allocation, with lightweight pre-tuning search for scale initialization, enabling competitive accuracy at 4-5 bits in 2.5-6 hours.
- ▌ Sparse Attention Reasoning · adu2021Learn sparse attention patterns for reasoning model decoding via self-distilled gating, achieving 9x speedup at 90% sparsity while maintaining reasoning quality.
- ▌ Sq Format Sparse Quantized · adu2021Unify sparse and quantized representations in a single hardware-friendly format for efficient LLM inference. Exploit complementary acceleration properties—high precision for sparse operations, low precision for dense—when W4A8 bottlenecks GPU throughput.
- ▌ Streaming Video Generation · adu2021Generate videos in real-time (16 FPS) by streaming frames continuously via modified flow matching with moving buffer mechanism and adaptive time embeddings.
- ▌ Swe Agents Long Context Rl · adu2021Train LLM-based agents for multi-turn SWE tasks via rejection fine-tuning and DAPO RL, scaling to 131k context length achieving 39% Pass@1.
- ▌ Tcandon Multi Agent Router · adu2021Route queries to multiple specialized agents dynamically using reasoning-aware routing that generates natural-language justification before predicting candidate agents. Enables seamless addition of new agents without system redesign. Routes aggregate responses from multiple specialists into coherent final answers, supporting enterprise-scale multi-agent systems with overlapping capabilities.
- ▌ Temporal Self Rewarding Lm · adu2021Improves self-rewarding language models through temporal framework that prevents representational collapse by anchoring rejected responses and guiding chosen responses across training phases.
- ▌ Thinksafe Safety Alignment · adu2021Align reasoning models to be safe without external supervision by prepending refusal instructions to harmful queries and training on self-generated safe responses. Unlocks latent safety knowledge within the model's native distribution.
- ▌ Tower Mind Agent Benchmark · adu2021Evaluate LLM agent capabilities using tower defense game environment with multimodal observations (pixel, text, structured state). Benchmark reveals critical agent limitations: inadequate planning validation, inflexible decision-making, and inefficient action use. Demonstrates significant performance gap between current LLMs and human experts, providing structured framework for measuring agent planning, adaptation, and hallucination tendencies.
- ▌ Treegrpo Tree Advantage Rl · adu2021Achieve 2.4× faster RL training for diffusion models by restructuring denoising as tree search with shared computation. TreeGRPO computes fine-grained step advantages instead of trajectory-level rewards—crucial for efficient diffusion model optimization.
- ▌ Ultramemv2 Memory Networks · adu2021Scale memory networks to 120B parameters with improved long-context learning through integrated memory layers, simplified value projection, and optimized parameter ratios for superior memory-intensive tasks.
- ▌ Uniql Quantization Pruning · adu2021Single cloud-side weight-sorting and fine-tuning supporting multiple on-device pruning rates via efficient SVD and MLP decomposition, achieving 4-5.7× memory reduction and 2.7-3.4× throughput across Transformers, SSMs, and hybrid architectures.
- ▌ Visplay Self Evolving Vlms · adu2021Enable VLMs to self-improve from unlabeled images via dual-role framework—questioner generates challenging visual questions while reasoner answers them, trained jointly with GRPO using difficulty and diversity rewards.
- ▌ Webwatcher Vision Research · adu2021Develops multimodal research agents that combine visual and textual reasoning for complex information-seeking tasks, using synthetic training trajectories and reinforcement learning.
- ▌ Agent Lightning Rl Training · adu2021Train RL on diverse agent frameworks (LangChain, AutoGen, custom) via unified data interface and transition-based RL decomposition.
- ▌ Agentic Context Engineering · adu2021Evolve agent behavior through iterative context refinement using delta updates rather than full rewrites, accumulating strategies and insights across execution traces.
- ▌ Aha Moment Vlm Verification · adu2021Reveals that inference-time scaling techniques for LLMs don't transfer to VLMs: majority voting beats verification, self-correction happens in <10% of cases, and models verify better without images. Use insights to design VLM evaluation methods that work rather than assuming LLM techniques apply directly.
- ▌ Arag Hierarchical Retrieval · adu2021Enable LLM agents to autonomously retrieve information across multiple granularities using keyword search, semantic search, and chunk read tools. Simple ReAct-based loop with hierarchical interfaces outperforms dense retrieval by allowing adaptive information seeking without complex graph construction.
- ▌ Arm Thinker Reward Modeling · adu2021Agentic reward model framework enabling active tool invocation (cropping, retrieval, validation) to ground judgments in verifiable evidence, using multi-stage GRPO with adaptive reward shaping for systematic evidence-based evaluation.
- ▌ Audio Roll Video Generation · adu2021Generate realistic video footage of people from audio input using a unified self-attention framework, producing convincing speaker performances without domain-specific restrictions.
- ▌ Avey Recurrence Free Ranker · adu2021Avey architecture pairs a ranker with autoregressive processor to select relevant tokens, decoupling context window from sequence length for efficient long-range processing.
- ▌ Canvas To Image Composition · adu2021Generate images with unified control over identity, spatial position, pose, and layout: encode diverse control modalities (spatial canvas, pose canvas, box canvas) into single RGB image, train diffusion model jointly across all control types, and enable flexible multi-modal composition at inference without task-specific fine-tuning.
- ▌ Code A1 Adversarial Rl Code · adu2021Train code and test generators through adversarial co-evolution where test LLM generates adversarial test cases to expose code defects. Prevent self-collusion by separating models and enabling white-box test generation.
- ▌ Computer Use Hybrid Actions · adu2021Enable computer-use agents to flexibly choose between GUI primitives (click, type) and high-level tool calls, reducing cascading errors by 22% and improving execution speed by 11%.
- ▌ Cooper Spatial Intelligence · adu2021Enhance spatial reasoning in multimodal LLMs by integrating depth and segmentation as auxiliary modalities with adaptive reasoning strategies. COOPER achieves 6.91% improvement in spatial understanding—when you need 3D-aware vision-language capabilities.
- ▌ Covrl Variational Reasoning · adu2021Enhance language model reasoning through coupled sampling from prior (question-only) and posterior (answer-conditioned) distributions. Construct composite distribution mixing both at token level using hybrid sampling. Combine reconstruction term, selective NLL loss, and KL regularization. Achieve 12.4% improvement over base model and 2.3% over comparable baselines.
- ▌ Critic Guided Formalization · adu2021Improve formal theorem proofs by treating criticism—evaluation of semantic correctness—as a learning signal. Train critic models to distinguish correct from incorrect formalizations, then use their feedback to guide RL-based proof generation.
- ▌ Cuda L2 Kernel Optimization · adu2021Uses LLMs with RL to automatically optimize HGEMM CUDA kernels across 1,000 configurations, systematically outperforming NVIDIA's cuBLAS and cuBLASLt through continued pretraining, general RL, and specialized HGEMM RL stages.
- ▌ Cure Coevolving LLM Testing · adu2021Improve code and test generation through co-evolution where LLMs generate both solutions and tests, optimizing each based on mutual evaluation and discriminative testing performance.
- ▌ Data Recipe Rl Optimization · adu2021Automatically synthesize and optimize training data using GRPO to generate data recipes (specifications for dataset creation). Use a Data Verifier to efficiently evaluate sample quality without full model training. Achieve performance comparable to human expert curation across diverse domains.