Representative papers Choose a field, then a paper: what question did it ask, and how did it test its idea? Each card includes an explanation, a reading starting point, and the original.
FieldAll fields Agents Interdisciplinary AI Embodied AI Generative models Language models Multimodal learning Reinforcement learning Efficiency & systems Computer vision World models Find a paper
49 papers
01 2012 · AlexNet · features ImageNet Classification with Deep Convolutional Neural Networks Can a network learn recognition features from a large image collection? Open card ↗ 2012 · NIPS 2012 ×
ImageNet Classification with Deep Convolutional Neural Networks Figure excerpt from the paper · Krizhevsky et al. · Fig. 2 · Original paper The problem Can a network learn recognition features from a large image collection?
The key idea Stack convolutional layers and combine GPU training with ReLU activations and regularization.
Why this paper It connects advances in data, compute, and training practice.
Where to start Start with the architecture and training choices, then inspect classification results and errors.
A question to keep asking It classifies predefined categories; it does not directly locate objects.
02 2014 · GAN · adversarial learning Generative Adversarial Nets How can a model learn to generate without directly specifying a complex image probability? Open card ↗ 2014 · NIPS 2014 ×
Generative Adversarial Nets Figure excerpt from the paper · Goodfellow et al., 2014 · Figure 1 · Original paper The problem How can a model learn to generate without directly specifying a complex image probability?
The key idea A generator turns random inputs into samples; a discriminator learns to distinguish them from training examples and supplies feedback.
Why this paper The two-model game gives a clear alternative training principle.
Where to start Start with both objectives and alternating updates, then the samples.
A question to keep asking Training can be unstable or cover only some patterns in the data.
03 2015 · DQN · Value learning Human-level control through deep reinforcement learning How can pixels guide discrete actions? Open card ↗ 2015 · Nature ×
Human-level control through deep reinforcement learning Figure excerpt from the paper · Mnih et al. · Fig. 1 · Original paper The problem How can pixels guide discrete actions?
The key idea Predict action values; replay experience and use a delayed target network to stabilize learning.
Why this paper A clear entry to observations, values, and deep RL training targets.
Where to start Trace the interaction loop, Q updates, and Atari evaluation.
A question to keep asking The original method enumerates discrete actions; continuous control needs other machinery.
04 2016 · ResNet · residuals Deep Residual Learning for Image Recognition Why can a deeper network fit even its training data worse? Open card ↗ 2016 · CVPR 2016 ×
Deep Residual Learning for Image Recognition Figure excerpt from the paper · He et al. · Fig. 5 · Original paper The problem Why can a deeper network fit even its training data worse?
The key idea Learn a change F(x) and add the input back, producing x + F(x).
Why this paper A clear architectural idea is tested against plain networks at different depths.
Where to start Read the degradation problem and residual block, then compare training errors.
A question to keep asking Easier optimization does not make arbitrary depth beneficial; data and compute still matter.
05 2016 · YOLO · detection You Only Look Once: Unified, Real-Time Object Detection Can one network pass predict both object categories and locations? Open card ↗ 2016 · CVPR 2016 ×
You Only Look Once: Unified, Real-Time Object Detection Figure excerpt from the paper · Redmon et al., YOLO, Figure 2 · Original paper The problem Can one network pass predict both object categories and locations?
The key idea Predict boxes and scores from the full image, training recognition and localization together.
Why this paper It connects directly to a detection demo and its misses, false positives, and runtime.
Where to start Start with the detection pipeline and grid formulation, then localization errors.
A question to keep asking The original struggles with small or crowded objects and precise localization. Modern YOLO tools have changed substantially.
06 2017 · MPNN · Molecular graphs Neural Message Passing for Quantum Chemistry How can atoms and their relationships predict molecular properties? Open card ↗ 2017 · ICML ×
Neural Message Passing for Quantum Chemistry Figure excerpt from the paper · Gilmer et al. · Fig. 1 · Original paper The problem How can atoms and their relationships predict molecular properties?
The key idea Update node representations through messages, then aggregate them for a molecular prediction.
Why this paper It grounds graph learning in a concrete chemistry task.
Where to start Start with message passing, readout, targets, and data.
07 2017 · Transformer Attention Is All You Need Can a sequence model handle dependencies without recurrent computation? Open card ↗ 2017 · NIPS 2017 ×
Figure excerpt from the paper · Vaswani et al. · Fig. 1 · Original paper The problem Can a sequence model handle dependencies without recurrent computation?
The key idea Each position reads other positions through attention, combined with feed-forward layers and positional information.
Why this paper Later models reuse these components. It is a useful first map of information flow.
Where to start Start with the architecture and attention explanation, then the translation experiments.
A question to keep asking The original is an encoder–decoder translation model.
08 2017 · PPO · Policy optimization Proximal Policy Optimization Algorithms How can simple policy updates avoid overly large changes? Open card ↗ 2017 · arXiv ×
Proximal Policy Optimization Algorithms Figure excerpt from the paper · Schulman et al. (2017), Fig. 1 · Original paper The problem How can simple policy updates avoid overly large changes?
The key idea The clipped objective removes incentives for excessive changes in action-probability ratios.
Why this paper A practical entry to policy gradients and many RL implementations.
Where to start Inspect the clipped objective and the sampling/minibatch-update loop.
A question to keep asking Clipping does not guarantee monotonic improvement; collection and implementation choices still matter.
09 2018 · Deep Potential · Faster simulation Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics How can accurate energy and force calculations become cheaper? Open card ↗ 2018 · Physical Review Letters ×
Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics Figure excerpt from the paper · Zhang et al., Deep Potential Molecular Dynamics, Figure 1 · Original paper The problem How can accurate energy and force calculations become cheaper?
The key idea Learn local atomic contributions from first-principles data while preserving problem symmetries.
Why this paper An example of replacing an expensive step in a scientific computation.
Where to start Trace energies and forces into simulation, then inspect physical comparisons.
A question to keep asking Accuracy depends on covered configurations and reference calculations; new conditions need testing.
10 2018 · GPT · pretraining Improving Language Understanding by Generative Pre-Training How can unlabeled text help supervised language-understanding tasks? Open card ↗ 2018 · OpenAI technical report · 2018 ×
Improving Language Understanding by Generative Pre-Training Figure excerpt from the paper · Radford et al., 2018 · Figure 1 · Original paper The problem How can unlabeled text help supervised language-understanding tasks?
The key idea Pretrain by predicting text, then fine-tune on task data arranged as input sequences.
Why this paper It makes the transfer from general representations to particular tasks concrete.
Where to start Figure 1 shows the training stages and task-specific input transformations.
A question to keep asking Downstream task fine-tuning is still required.
11 2018 · SAC · Continuous control Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor How can experience be reused without committing too early to one action? Open card ↗ 2018 · ICML ×
Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor Figure excerpt from the paper · Haarnoja et al. (2018), Fig. 1 · Original paper The problem How can experience be reused without committing too early to one action?
The key idea Optimize reward and policy entropy, using a critic and replayed experience.
Why this paper Compare with PPO to understand data reuse and exploration objectives.
Where to start Connect the maximum-entropy objective, actor, critic, and replay buffer.
A question to keep asking This paper focuses on continuous control; reward and entropy scales affect behavior.
12 2018 · Representation · Memory · Control World Models Can a learned representation support a small controller? Open card ↗ 2018 · arXiv ×
World Models Figure excerpt from the paper · Ha & Schmidhuber · Fig. 4 · Original paper The problem Can a learned representation support a small controller?
The key idea Compress images, learn dynamics with a recurrent model, and feed their representations to a controller.
Why this paper The interactive article makes the components tangible.
Where to start Read Agent Model and Car Racing, then the imagined VizDoom experiment.
A question to keep asking Controllers can exploit model errors; test model-trained behavior in the original environment.
13 2019 · Optuna · Hyperparameter search Optuna: A Next-generation Hyperparameter Optimization Framework How can a limited training budget find a good configuration? Open card ↗ 2019 · KDD ×
Optuna: A Next-generation Hyperparameter Optimization Framework Figure excerpt from the paper · Akiba et al. · Figure 6 · Original paper The problem How can a limited training budget find a good configuration?
The key idea Define the search space in code, then combine sampling and pruning to allocate the budget across trials.
Where to start Start with the search space and trial loop, then examine how early stopping saves computation.
14 2020 · Qlib · Quantitative workflow Qlib: An AI-oriented Quantitative Investment Platform How do data, models, portfolios, and backtesting form a research workflow? Open card ↗ 2020 · arXiv ×
Qlib: An AI-oriented Quantitative Investment Platform Figure excerpt from the paper · Figure 1 · Original paper The problem How do data, models, portfolios, and backtesting form a research workflow?
The key idea Connect financial data processing and machine learning pipelines to strategy evaluation so that different models can be compared within a common workflow.
Where to start Trace the data flow through the framework diagram, then identify the time split, labels, and backtesting configuration in an example.
15 2020 · DDPM · diffusion Denoising Diffusion Probabilistic Models Can learning denoising tasks produce clear images? Open card ↗ 2020 · NeurIPS 2020 ×
Denoising Diffusion Probabilistic Models Figure excerpt from the paper · Ho et al. · Fig. 1 · Original paper The problem Can learning denoising tasks produce clear images?
The key idea Train on known noise added to real images; repeatedly use the learned predictions to sample from an initial random state.
Why this paper Its training and sampling algorithms provide a useful foundation for later diffusion work.
Where to start Compare the two algorithms, tracing model inputs and predictions.
A question to keep asking Original sampling uses many sequential model calls.
16 2020 · GPT-3 · in-context learning Language Models are Few-Shot Learners Can a few examples specify a new task without weight updates? Open card ↗ 2020 · NeurIPS 2020 ×
Language Models are Few-Shot Learners Figure excerpt from the paper · Brown et al., 2020 · Figure 2.1 · Original paper The problem Can a few examples specify a new task without weight updates?
The key idea Scale an autoregressive model and evaluate instructions and demonstrations under zero-, one-, and few-shot settings.
Why this paper It distinguishes in-context learning from fine-tuning and connects to questions about scale.
Where to start Start with the three evaluation settings, then inspect a familiar task.
A question to keep asking Results depend on the task and prompt; overlap between web training data and tests complicates evaluation.
17 2020 · Distributed training ZeRO: Memory Optimizations Toward Training Trillion Parameter Models How much memory is wasted by replicating training state on every GPU? Open card ↗ 2020 · SC 2020 ×
ZeRO: Memory Optimizations Toward Training Trillion Parameter Models Figure excerpt from the paper · Rajbhandari et al. (2020), Fig. 1 · Original paper The problem How much memory is wasted by replicating training state on every GPU?
The key idea Progressively partition optimizer states, gradients, and parameters, communicating required pieces when needed.
Why this paper It explains training memory beyond parameter count.
Where to start Identify what each of the three stages partitions, then compare memory and communication costs.
A question to keep asking Sharding changes communication. Benefits depend on bandwidth and device count.
18 2020 · MuZero · Learning and search Mastering Atari, Go, chess and shogi by planning with a learned model Can a learned model support search without supplied dynamics? Open card ↗ 2020 · Nature ×
Mastering Atari, Go, chess and shogi by planning with a learned model Figure excerpt from the paper · Schrittwieser et al. (2020), Fig. 1 · Original paper The problem Can a learned model support search without supplied dynamics?
The key idea Predict reward, value, and policy in internal states, then compare choices with tree search.
Why this paper Useful models need not reconstruct every pixel.
Where to start Follow planning, acting, and training before game results.
A question to keep asking The paper evaluates specific games and action interfaces.
19 2021 · AlphaFold2 · Protein structure Highly accurate protein structure prediction with AlphaFold How can sequence information support three-dimensional structure prediction? Open card ↗ 2021 · Nature ×
Highly accurate protein structure prediction with AlphaFold Figure excerpt from the paper · Jumper et al., AlphaFold, Figure 1(e) · CC BY 4.0 · Original paper The problem How can sequence information support three-dimensional structure prediction?
The key idea Combine evolutionary sequence information and residue relationships, iteratively refining representations and structure.
Why this paper It connects data, geometry, and system design around a disciplinary problem.
Where to start Start with the architecture, CASP14 evaluation, and confidence estimates.
20 2021 · FNO · Neural operators Fourier Neural Operator for Parametric Partial Differential Equations How can a set of initial conditions directly predict continuous physical fields? Open card ↗ 2021 · ICLR ×
Fourier Neural Operator for Parametric Partial Differential Equations Figure excerpt from the paper · Figure 1 · Original paper The problem How can a set of initial conditions directly predict continuous physical fields?
The key idea Parameterize integral kernels in the frequency domain to learn mappings from functions to functions for a family of partial differential equation problems.
Where to start Start with the operator layer and the inputs and outputs, then compare errors across resolutions and equation settings.
21 2021 · CLIP · alignment Learning Transferable Visual Models From Natural Language Supervision Can captions teach visual concepts beyond fixed category labels? Open card ↗ 2021 · ICML 2021 ×
Learning Transferable Visual Models From Natural Language Supervision Figure excerpt from the paper · Radford et al. · Fig. 1 · Original paper The problem Can captions teach visual concepts beyond fixed category labels?
The key idea Train image and text encoders so matched pairs are more similar than mismatches. Compare an image with candidate descriptions at inference.
Why this paper Retrieval and zero-shot classification become an observable matching problem.
Where to start Follow pretraining and zero-shot prediction, then change the official example's candidate text.
A question to keep asking CLIP produces representations or scores, not conversation. Correct matching need not imply relational understanding.
22 2021 · ViT · image patches An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale Does image classification require convolutional structure? Open card ↗ 2021 · ICLR 2021 ×
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale Figure excerpt from the paper · Dosovitskiy et al. · Fig. 1 · Original paper The problem Does image classification require convolutional structure?
The key idea Turn image patches into vectors and exchange information using a Transformer.
Why this paper It connects vision to sequence processing and highlights pretraining scale.
Where to start Follow patches into the model, then compare pretraining dataset sizes.
A question to keep asking The original strong results rely on large-scale pretraining; they do not imply Transformers always beat CNNs on small datasets.
23 2022 · RT-1 · Multitask control RT-1: Robotics Transformer for Real-World Control at Scale Can one model learn instruction-following from multitask robot data? Open card ↗ 2022 · arXiv ×
RT-1: Robotics Transformer for Real-World Control at Scale Figure excerpt from the paper · Brohan et al., RT-1, Figure 1(a) · Original paper The problem Can one model learn instruction-following from multitask robot data?
The key idea Condition a Transformer on images and language to predict discretized actions.
Why this paper It makes data scale and diversity concrete in robotics.
Where to start Inspect inputs, outputs, data collection, and generalization evaluations.
A question to keep asking Its results depend on a particular robot body and action interface.
24 2022 · LDM · latent space High-Resolution Image Synthesis with Latent Diffusion Models Can diffusion process less information while retaining useful image detail? Open card ↗ 2022 · CVPR 2022 ×
High-Resolution Image Synthesis with Latent Diffusion Models Figure excerpt from the paper · Rombach et al. · Fig. 3 · Original paper The problem Can diffusion process less information while retaining useful image detail?
The key idea Compress images with an autoencoder, diffuse in latent space, and decode. Cross-attention introduces conditions such as text.
Why this paper It links compression, quality, and compute, and underpins Stable Diffusion.
Where to start Trace encoding, diffusion, and decoding; then compare compression levels.
A question to keep asking Compression loses information. Text conditioning does not guarantee accurate counts or spatial relations.
25 2022 · InstructGPT · human feedback Training language models to follow instructions with human feedback How can a text-completion model better follow a user's request? Open card ↗ 2022 · NeurIPS 2022 ×
Training language models to follow instructions with human feedback Figure excerpt from the paper · Ouyang et al. · Fig. 2 · Original paper The problem How can a text-completion model better follow a user's request?
The key idea Fine-tune on demonstrations, learn a reward model from ranked answers, and use reinforcement learning to adjust behavior.
Why this paper It separates pretrained capability from the way answers are produced.
Where to start Follow the training pipeline and identify the data needed for demonstrations, ranking, and policy updates.
A question to keep asking Preferences depend on labelers and prompt distributions; preferred answers can still be factually wrong.
26 2022 · Attention kernel FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness Can attention be limited by data movement rather than multiplication count? Open card ↗ 2022 · NeurIPS 2022 ×
FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness Figure excerpt from the paper · Dao et al. · Fig. 1 · Original paper The problem Can attention be limited by data movement rather than multiplication count?
The key idea Tile work into fast on-chip memory instead of repeatedly materializing the full attention matrix in device memory.
Why this paper It shows how the same mathematical operation can have very different execution costs.
Where to start Start with GPU memory levels and tiling, then IO analysis and runtime.
A question to keep asking Exact dense attention still has quadratic arithmetic complexity. Speedup depends on shapes, hardware, and implementation.
27 2023 · Action loop ReAct: Synergizing Reasoning and Acting in Language Models Can a model gather external information while adjusting its approach? Open card ↗ 2023 · ICLR 2023 ×
ReAct: Synergizing Reasoning and Acting in Language Models Figure excerpt from the paper · Yao et al. · Fig. 1 · Original paper The problem Can a model gather external information while adjusting its approach?
The key idea Interleave reasoning text and actions, then add environmental observations to context.
Why this paper It makes task execution visible as a sequence you can inspect.
Where to start Start with the project's HotpotQA example and then its ALFWorld failure.
A question to keep asking Models can ignore feedback or repeat mistakes. Written reasoning is not a complete account of internal computation.
28 2023 · Learning tool use Toolformer: Language Models Can Teach Themselves to Use Tools How can tool-use training data be obtained without labeling every call? Open card ↗ 2023 · NeurIPS 2023 ×
Figure excerpt from the paper · Schick et al., 2023 · Figure 2 · Original paper The problem How can tool-use training data be obtained without labeling every call?
The key idea Generate candidate calls, execute them, retain those that improve subsequent text prediction, and fine-tune on the resulting data.
Why this paper Contrast it with ReAct to separate execution design from trained capability.
Where to start Figure 2 shows sampling, execution, and filtering; then inspect tool examples.
A question to keep asking The original method has limitations with chained calls and interactive tools. Better text prediction is not the same as long-task success.
29 2023 · Feedback and memory Reflexion: Language Agents with Verbal Reinforcement Learning How can another attempt use lessons from a failure? Open card ↗ 2023 · NeurIPS 2023 ×
Reflexion: Language Agents with Verbal Reinforcement Learning Figure excerpt from the paper · Shinn et al., 2023 · Figure 2(a) · Original paper The problem How can another attempt use lessons from a failure?
The key idea Turn task feedback into written reflections, store them, and supply them to subsequent attempts.
Why this paper It provides a concrete contrast between changing context and changing parameters.
Where to start Follow execution, evaluation, reflection, and the next attempt through the method.
A question to keep asking Reflections need useful feedback. This is not traditional reinforcement-learning training through gradient updates.
30 2023 · Memory and social interaction Generative Agents: Interactive Simulacra of Human Behavior How can 25 town residents remember experiences, plan their days, and interact? Open card ↗ 2023 · UIST 2023 ×
Generative Agents: Interactive Simulacra of Human Behavior Figure excerpt from the paper · Park et al. · Smallville · Original paper The problem How can 25 town residents remember experiences, plan their days, and interact?
The key idea Record observations in language, retrieve memories by relevance, importance, and recency, then reflect and plan.
Why this paper Follow how a party invitation spreads through individual memories, conversations, and actions.
Where to start Start with Smallville and the Valentine’s Day party, then study memory retrieval, reflection, and planning.
31 2023 · Minecraft and reusable skills Voyager: An Open-Ended Embodied Agent with Large Language Models How can an agent build on what it knows, from chopping wood to mining diamonds? Open card ↗ 2023 · arXiv 2023 ×
Voyager: An Open-Ended Embodied Agent with Large Language Models Figure excerpt from the paper · Wang et al. · Fig. 1 · Original paper The problem How can an agent build on what it knows, from chopping wood to mining diamonds?
The key idea An automatic curriculum proposes tasks. GPT-4 writes and revises executable code using feedback; successful skills are stored for retrieval and reuse.
Why this paper The tech tree makes cumulative progress tangible: gathering, crafting, and tool use support later skills.
Where to start Start with the tech-tree exploration figure, then study the curriculum, skill library, and code-refinement loop.
32 2023 · DreamDiffusion · EEG decoding DreamDiffusion: Generating High-Quality Images from Brain EEG Signals How can EEG representations connect to an image generation model? Open card ↗ 2023 · arXiv ×
DreamDiffusion: Generating High-Quality Images from Brain EEG Signals Figure excerpt from the paper · Figure 1 · Original paper The problem How can EEG representations connect to an image generation model?
The key idea Learn EEG representations through temporally masked signal modeling, combine them with CLIP visual supervision, and use the signals for conditional generation.
Where to start Trace the role of each data source and supervision signal through the training stages, then examine participant splits and controls.
33 2023 · Diffusion Policy · Action generation Diffusion Policy: Visuomotor Policy Learning via Action Diffusion How can a policy represent several valid, coherent action sequences? Open card ↗ 2023 · Robotics: Science and Systems ×
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion Figure excerpt from the paper · Chi et al. · Fig. 3 · Original paper The problem How can a policy represent several valid, coherent action sequences?
The key idea Denoise an observation-conditioned action sequence, execute part, then observe and predict again.
Why this paper A tangible bridge from generative models to control.
Where to start Follow denoising and receding-horizon execution, then Push-T videos.
A question to keep asking Iterative denoising costs time; demonstration coverage and control frequency constrain deployment.
34 2023 · RT-2 · Vision-language-action RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control Can web-trained knowledge help robots interpret new instructions? Open card ↗ 2023 · CoRL ×
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control Figure excerpt from the paper · Zitkovich et al. · Fig. 1 · Original paper The problem Can web-trained knowledge help robots interpret new instructions?
The key idea Encode actions as tokens and co-fine-tune with vision-language data.
Why this paper It connects semantic knowledge to existing action capabilities.
Where to start Read action encoding, then semantic-generalization tasks.
A question to keep asking Semantic transfer does not automatically create new low-level motor skills; robot demonstrations remain important.
35 2023 · DiT · diffusion Transformer Scalable Diffusion Models with Transformers Can a scalable Transformer serve as the diffusion network? Open card ↗ 2023 · ICCV 2023 ×
Scalable Diffusion Models with Transformers Figure excerpt from the paper · Peebles & Xie · Fig. 3 · Original paper The problem Can a scalable Transformer serve as the diffusion network?
The key idea Process latent patches with a Transformer and compare depth, width, and patch size.
Why this paper It separates the generation objective from architecture and examines scaling through compute.
Where to start Read the architecture, then compare compute and generation metrics across sizes.
A question to keep asking The original studies class-conditioned ImageNet images, not text-to-video generation.
36 2023 · BLIP-2 · connector BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models Can existing vision and language models be reused with less retraining? Open card ↗ 2023 · ICML 2023 ×
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models Figure excerpt from the paper · Li et al., BLIP-2, Figure 1 · Original paper The problem Can existing vision and language models be reused with less retraining?
The key idea Freeze both models and train a Q-Former to extract a compact set of visual features and connect them to the language model.
Why this paper It turns the idea of giving a language model vision into modules and training stages.
Where to start Mark frozen and trainable parts in the two-stage framework, then trace visual features into language inputs.
A question to keep asking Frozen components retain their limitations, and visual errors can accompany fluent but inaccurate answers.
37 2023 · LLaVA · visual instructions Visual Instruction Tuning How can a model answer natural-language instructions about an image? Open card ↗ 2023 · NeurIPS 2023 ×
Visual Instruction Tuning Figure excerpt from the paper · Liu et al., Visual Instruction Tuning, Figure 1 · Original paper The problem How can a model answer natural-language instructions about an image?
The key idea Connect an image encoder to a language model with a projection and train on visual instructions. The original uses captions and box information with text-only GPT-4 to create part of its dialogue data.
Why this paper Data organization can change interaction as much as the connection itself.
Where to start Start with instruction-data construction, then the connector and staged training.
A question to keep asking Answers can invent visual details.
38 2023 · Quantization SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models Why do a few very large activations make INT8 quantization difficult? Open card ↗ 2023 · ICML 2023 ×
SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models Figure excerpt from the paper · Xiao et al. (2023), Fig. 2 · Original paper The problem Why do a few very large activations make INT8 quantization difficult?
The key idea Redistribute scale between weights and activations with an equivalent transformation, then quantize.
Why this paper It connects outliers, numerical error, and hardware execution in one example.
Where to start Start with activation outliers and scale migration, then accuracy and speed comparisons.
A question to keep asking The pre-quantization transformation is equivalent; quantization still introduces error. Representative calibration data and suitable INT8 kernels matter.
39 2023 · vLLM · serving cache Efficient Memory Management for Large Language Model Serving with PagedAttention How can changing request lengths avoid wasting KV-cache memory? Open card ↗ 2023 · SOSP 2023 ×
Efficient Memory Management for Large Language Model Serving with PagedAttention Figure excerpt from the paper · Kwon et al. (2023), Fig. 6 · Original paper The problem How can changing request lengths avoid wasting KV-cache memory?
The key idea Split the cache into blocks, map logical sequences to noncontiguous storage, and support cache sharing.
Why this paper It connects operating-system paging to the capacity of a model-serving system.
Where to start Start with fragmentation and block mapping, then throughput under matched latency conditions.
A question to keep asking It optimizes cache management. Concurrency gains must be assessed with latency targets and request lengths.
40 2024 · Repository-level evaluation SWE-bench: Can Language Models Resolve Real-World GitHub Issues? Can a model turn a repository and issue into a working fix? Open card ↗ 2024 · ICLR 2024 ×
SWE-bench: Can Language Models Resolve Real-World GitHub Issues? Figure excerpt from the paper · Jimenez et al., 2024 · Figure 1 · Original paper The problem Can a model turn a repository and issue into a working fix?
The key idea Collect real issues and fixes, provide the repository and problem, and evaluate proposed patches with tests.
Why this paper It joins reading, localization, editing, and verification in one task.
Where to start Start with task construction and evaluation, then a failure case.
A question to keep asking The original data comes from Python repositories. Passing tests does not cover all software quality or establish transfer to other languages.
41 2024 · AlphaGeometry · Geometry proofs Solving olympiad geometry without human demonstrations How can auxiliary constructions and symbolic reasoning form a proof process? Open card ↗ 2024 · Nature ×
Solving olympiad geometry without human demonstrations Figure excerpt from the paper · Google DeepMind · AlphaGeometry 官方方法示意图 · Original paper The problem How can auxiliary constructions and symbolic reasoning form a proof process?
The key idea A symbolic engine first derives consequences from the given conditions. When needed, a language model proposes auxiliary constructions, and proof search continues.
Where to start Trace the added constructions and the deductions they enable in a simple geometry problem, then examine how the training data are synthesized.
42 2024 · π0 · Generalist robot policy π0: A Vision-Language-Action Flow Model for General Robot Control How can a vision-language model produce continuous actions across robot tasks and platforms? Open card ↗ 2024 · arXiv ×
π0: A Vision-Language-Action Flow Model for General Robot Control Figure excerpt from the paper · Physical Intelligence · Fig. 1 · Original paper The problem How can a vision-language model produce continuous actions across robot tasks and platforms?
The key idea An action expert uses flow matching to generate continuous action sequences alongside a pretrained VLM. The policy is pretrained on demonstrations from multiple robots, then fine-tuned on task data.
Why this paper It brings VLM pretraining, action generation, and cross-robot data together, connecting RT-2 and Diffusion Policy.
Where to start Start with the first-page overview, then the action-expert architecture. Compare direct use after pretraining with task-specific fine-tuning.
43 2025 · DGM · Self-improvement Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents Can a coding agent improve its abilities by modifying its own code? Open card ↗ 2025 · arXiv ×
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents Figure excerpt from the paper · Figure 1 · Original paper The problem Can a coding agent improve its abilities by modifying its own code?
The key idea Maintain multiple agent versions, generate self-modifications, and select candidates using coding tasks, creating an expanding version tree.
Where to start Start with one specific code change, then examine version selection, evaluation tasks, and the computation budget.
44 2025 · AI Scientist-v2 · Automated research The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search How can experimental results guide the next round of research? Open card ↗ 2025 · arXiv ×
The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search Figure excerpt from the paper · Yamada et al. · Figure 1 · Original paper The problem How can experimental results guide the next round of research?
The key idea Use agentic tree search to organize experiments, connecting plans, execution, analysis, and paper writing.
Where to start Start with one complete experimental branch and check whether the final claims correspond to actual results.
45 2025 · AMix-1 · Protein generation AMix-1: A Pathway to Test-Time Scalable Protein Foundation Model How can family information and verification feedback support protein design? Open card ↗ 2025 · arXiv ×
AMix-1: A Pathway to Test-Time Scalable Protein Foundation Model Figure excerpt from the paper · Figure 1 · Original paper The problem How can family information and verification feedback support protein design?
The key idea Model proteins with Bayesian flow networks, use MSA context, and explore design workflows with larger inference-time search budgets.
Where to start Trace generation conditions, verifiers, and experimental measurements separately, comparing their roles in the iterative process.
46 2025 · CMRINet · Cardiac imaging CMRINet: Joint Groupwise Registration and Segmentation for Cardiac Function Quantification from Cine-MRI How can dynamic cardiac MRI provide both structure and motion information? Open card ↗ 2025 · arXiv ×
CMRINet: Joint Groupwise Registration and Segmentation for Cardiac Function Quantification from Cine-MRI Figure excerpt from the paper · Figure 1 · Original paper The problem How can dynamic cardiac MRI provide both structure and motion information?
The key idea Jointly handle groupwise registration and segmentation, connecting temporal image analysis to cardiac function quantification.
Where to start Start with the temporal dimension of cine MRI, then examine how the two tasks work together and how functional measures are calculated.
47 2025 · π0.5 · New-environment generalization π0.5: a Vision-Language-Action Model with Open-World Generalization Can a robot carry out long tasks in homes it did not visit during training? Open card ↗ 2025 · arXiv ×
π0.5: a Vision-Language-Action Model with Open-World Generalization Figure excerpt from the paper · Physical Intelligence · Fig. 1 · Original paper The problem Can a robot carry out long tasks in homes it did not visit during training?
The key idea Building on π0, the model co-trains on robot actions, subtask prediction, object detection, and web vision-language data, connecting high-level subtasks to continuous actions.
Why this paper It studies multi-stage manipulation in new homes and how the mix of training data affects generalization.
Where to start Start with the first-page data and deployment overview, then inspect the new-home evaluation and data-source ablations.
48 2025 · DreamerV3 · Learning in imagination Mastering diverse control tasks through world models How can one algorithm handle different task signal scales? Open card ↗ 2025 · Nature ×
Mastering diverse control tasks through world models Figure excerpt from the paper · Hafner et al. (2025), Fig. 1 (CC BY 4.0) · Original paper The problem How can one algorithm handle different task signal scales?
The key idea Learn a world model from experience, train actor and critic on imagined trajectories, and stabilize losses and signal scales.
Why this paper It connects predictive learning with behavior learning and cross-task evaluation.
Where to start Start with the training diagram, learning algorithm, and ablations.
A question to keep asking A shared configuration does not mean one pretrained agent solves every task without training.
49 2026 · FLAG · Spatial gene expression FLAG: Foundation model representation with Latent diffusion Alignment via Graph for spatial gene expression prediction How can tissue images predict spatial gene expression while preserving structural relationships? Open card ↗ 2026 · ICML ×
FLAG: Foundation model representation with Latent diffusion Alignment via Graph for spatial gene expression prediction Figure excerpt from the paper · Figure 1 · Original paper The problem How can tissue images predict spatial gene expression while preserving structural relationships?
The key idea Combine spatial graph encoding with gene foundation model representations, and use a diffusion process to model expression distributions.
Where to start Start with input images, spatial adjacency, and expression outputs, then compare errors at individual locations with structural evaluations.
Here are papers selected by question, each with an angle to begin from. For the reading process, see finding and reading papers .
A first encounter with a paper
ResNet Mu Li's readings , then return to vision .
Attention Is All You Need The illustrated explanation helps a first pass. Continue to foundations and architectures .
ReAct agents .
Follow your interests
BERT : how objectives shape representations and how those representations serve tasks. See LLMs .LoRA : what reducing trainable parameters saves and how effectiveness is evaluated. See LLMs and systems .CLIP : image–text matching as training and how the trained model is used. See multimodal models .DDPM : connections between noisy data, objectives, and generation. Pair with generation .World Models : combining compressed observations, dynamics prediction, and action selection. See world models .DreamerV3 : learned models for behavior learning and the tasks covered by evaluation. See RL .ACT / ALOHA : connecting demonstrations and action prediction to real robot tasks. See embodied AI .Diffusion Policy : generative modeling of actions and the reason for that representation. See generation and embodied AI .
Longer lists once you have a question
Agents · Multimodal models · World Models · Embodied guide · AI for Science
For new work, see Hugging Face Daily Papers .
Read with someone experienced
Compare your own reading notes with Mu Li's close readings . For more projects and paper lists, use the direction catalog .