Topics — The twenty-eight kinds of claim this anthology files against — the primary axis of both the practice registry and the reading list, and the same twenty-eight for each (ADR-026). The filing rule is the point: take a domain topic when the claim is ABOUT the domain; a claim merely discovered in one still takes its kind. The axis organizes the record, it does not bound it: content the list cannot place is evidence the list is short (ADR-059).
Data pipeline — loading, quality assessment and selection, preprocessing, batch preparation.
60 of 737 LIT documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| LIT-050 v2 | Analyzing and Mitigating Data Stalls in DNN Training | Mohan et al. (2020), ARXIV-2007.06775. Training time is frequently dominated by data stalls — waiting on fetch and preprocessing — and the only way to know is to measure the input pipeline separately from the step. | Active |
| LIT-053 v2 | High Performance I/O For Large Scale Deep Learning | Aizman et al. (2020), ARXIV-2001.01858. Existing distributed filesystems were not built for the access pattern deep learning has; WebDataset stores training data as sharded POSIX tar archives so the pattern becomes sequential reads of large files. | Active |
| LIT-080 v2 | PaLI: A Jointly-Scaled Multilingual Language-Image Model | Chen et al. (2022), ARXIV-2209.06794. Joint vision-language scaling. | Active |
| LIT-082 v3 | Constitutional AI: Harmlessness from AI Feedback | Bai et al. (2022), ARXIV-2212.08073. Train harmlessness without human harm labels: the model critiques and revises its own responses against a written list of principles, and the preference model learns from those comparisons. | Active |
| LIT-087 v5 | Data Selection for Language Models via Importance Resampling | Xie et al. (2023), ARXIV-2302.03169. Data quality assessment. | Active |
| LIT-096 v2 | Segment Anything | Kirillov et al. (2023), ARXIV-2304.02643. Universal segmentation model. | Active |
| LIT-099 v2 | PaLM 2 Technical Report | Anil et al. (2023), ARXIV-2305.10403. Architecture improvements. | Active |
| LIT-117 v2 | Efficient Online Data Mixing For Language Model Pre-Training | Albalak et al. (2023), ARXIV-2312.02406. Online Data Mixing: an Exp3 bandit over data domains, rewarded by per-domain training loss, revising the mixture during training. Reaches the next best method’s final perplexity in 19% fewer iterations at negligible wall-clock cost. | Active |
| LIT-120 | Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance | Zuo et al. (2025), ARXIV-2507.22448. A parallel hybrid mixer — attention heads and Mamba heads side by side in one block — from 0.5B to 34B, with a µP recipe adapted to it and multi-epoch training on high-quality data; the architecture and the memorization-window measurement LIT-119 builds on. | Active |
| LIT-124 | Efficient Training of Language Models to Fill in the Middle | Bavarian et al. (2022), ARXIV-2207.14255. Moving a span from the middle of a document to its end teaches an autoregressive model to infill, and doing it to a large fraction of the data costs nothing in left-to-right quality — the FIM-for-free result — so the authors recommend training with FIM by default. | Active |
| LIT-125 | Qwen2.5-Coder Technical Report | Hui et al. (2024), ARXIV-2409.12186. Six code models from 0.5B to 32B, continued-pretrained from Qwen2.5 on over 5.5T tokens with cleaned, synthetic and balanced code/text/math data, and trained for fill-in-the-middle at file and repository level; state of the art on more than ten code benchmarks at release. | Active |
| LIT-126 | Structure-Aware Fill-in-the-Middle Pretraining for Code | Gong et al. (2025), ARXIV-2506.00204. Mask complete syntactic structures chosen from the AST instead of random character spans, so FIM examples match how code is actually edited; on 1B and 8B models AST-FIM beats random-character FIM by up to 5 points, and Real-FIM-Eval, built from 30k+ GitHub commits in 12 languages, measures the real task. | Active |
| LIT-127 v2 | DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models | Shao et al. (2024), ARXIV-2402.03300. A 7B model continued-pretrained on 120B math tokens mined from Common Crawl reaches 51.7% on MATH without tools or voting, and introduces Group Relative Policy Optimization — PPO without a critic, with the baseline taken from a group of sampled outputs — the RL algorithm most later reasoning recipes use. | Active |
| LIT-166 | Scaling Data-Constrained Language Models | Muennighoff et al. (2023), ARXIV-2305.16264. Up to four epochs of repeated data cost almost nothing against unique data at fixed compute; past that the value of added compute decays to zero. 400 runs, up to 900B tokens and 9B parameters. | Active |
| LIT-175 | Demystifying Training-Time Augmentation for Data-Constrained Language Model Pretraining | Chen et al. (2026), ARXIV-2606.16246. Autoregressive pretraining overfits badly under heavy repetition; three orthogonal families of augmentation — token noise, sequence permutation, offset targets — delay it and make hundreds of epochs on a fixed corpus productive. | Active |
| LIT-178 | Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset | Su et al. (2024), ARXIV-2412.02595. Aggressive model-based filtering throws away 90% of the web, which is fine at 1T tokens and ruinous at 15T; classifier ensembling plus synthetic rephrasing keeps four times the unique tokens at the same quality. | Active |
| LIT-184 | The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale | Penedo et al. (2024), ARXIV-2406.17557. 15T tokens from 96 Common Crawl snapshots, with every design choice documented and ablated — and FineWeb-Edu, the 1.3T educational filter that moves MMLU and ARC sharply. | Active |
| LIT-201 | Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance | Ye et al. (2024), ARXIV-2403.16952. Mixture proportions are tuned by heuristic because nobody had shown they were predictable. They are: loss is a fittable function of the proportions, so training a handful of small mixtures predicts performance on mixtures never run. Nested with the scaling laws for steps and model size, a 1B model over 100B RedPajama tokens matched the default mixture trained 48% longer. | Active |
| LIT-202 | Deduplicating Training Data Makes Language Models Better | Lee et al. (2021), ARXIV-2107.06499. Standard corpora are full of near-duplicates — C4 contains one 61-word English sentence repeated thousands of times. Deduplicating makes models emit memorised text ten times less often and reach the same or better accuracy in fewer steps, and it fixes evaluation too: train-test overlap affects a large fraction of standard validation sets. | Active |
| LIT-203 | Fewer Truncations Improve Language Modeling | Ding et al. (2024), ARXIV-2404.10830. Everyone concatenates documents and splits at a fixed length to avoid padding, which shreds documents into incomplete pieces. Best-fit Packing treats the grouping as bin packing instead and eliminates unnecessary truncation at the same training efficiency — no padding added. Across 22 tasks at 7B–13B: +4.7% reading comprehension, +16.8% context following, +9.2% program synthesis, and up to 58.3% less closed-domain hallucination. | Active |
| LIT-205 v2 | How to Synthesize Text Data without Model Collapse? | Zhu et al. (2024), ARXIV-2412.14689. Pre-training across proportions of synthetic data gives a negative correlation between that proportion and performance, and the statistical signature is distributional shift plus over-concentration of n-gram features. The proposed answer is not better generation but token-level editing of human text into semi-synthetic data, with a proof that the test error is then bounded. | Active |
| LIT-391 | DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining | Xie et al. (2023), ARXIV-2305.10429. Train a 280M proxy under group DRO to produce domain weights with no knowledge of downstream tasks, then resample and train an 8B model with them: +6.5 points average few-shot accuracy and the baseline reached in 2.6x fewer steps. The design point is the objective — worst-case EXCESS loss against a reference model, because worst-case raw loss would chase the noisiest domain. | Active |
| LIT-392 | Skill-it! A Data-Driven Skills Framework for Understanding and Training Language Models | Chen et al. (2023), ARXIV-2307.14430. Training data has a prerequisite structure: skills are learned in an order, and training on a prerequisite makes the dependent skill reachable with less data. Formalises a skill and an ordered skill set in terms of the data that induces them, and samples over skill mixtures online. | Active |
| LIT-397 | Deep Learning on a Data Diet: Finding Important Examples Early in Training | Paul et al. (2021), ARXIV-2107.07075. Two scores — GraNd and EL2N — that rank training examples by importance a few epochs into training rather than at the end of it, and transfer across architectures. Also the finding that the very highest-scoring examples should be excluded, more so under label noise. | Active |
| LIT-398 | DataComp-LM: In search of the next generation of training sets for language models | Li et al. (2024), ARXIV-2406.11794. A controlled testbed for data curation — a fixed 240T-token corpus, fixed training recipes at 412M–7B, and 53 evaluations — whose baseline finding is that model-based filtering is what makes a high-quality training set. The position two documents in this record already argue with. | Active |
| LIT-399 | Beyond neural scaling laws: beating power law scaling via data pruning | Sorscher et al. (2022), ARXIV-2206.14486. Power-law scaling in dataset size can be beaten — reduced toward exponential — if you have a good ranking of which examples to discard and you prune harder as the corpus grows. Which end of the ranking to discard inverts with data abundance. | Active |
| LIT-400 | Estimating Training Data Influence by Tracing Gradient Descent | Pruthi et al. (2020), ARXIV-2002.08484. Attributes a prediction to training data by tracing the actual optimization: how much did the loss on the test point move on the steps where this training example was used? Needs only gradients, saved checkpoints and a loss — no Hessian, no convexity. | Active |
| LIT-401 | Studying Large Language Model Generalization with Influence Functions | Grosse et al. (2023), ARXIV-2308.03296. Scales influence functions to language models up to 52B parameters with an EK-FAC approximation to the inverse-Hessian-vector product, then uses them to ask what training data a behaviour came from. Generalization gets more abstract with scale — and stays startlingly sensitive to word order. | Active |
| LIT-403 | Understanding Black-box Predictions via Influence Functions | Koh and Liang (2017), ARXIV-1703.04730. Traces a prediction back through the learning algorithm to the training points responsible for it, using a classical technique from robust statistics made tractable by needing only gradients and Hessian-vector products. The origin of training-data attribution in deep learning. | Active |
| LIT-425 v2 | Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer | Raffel et al. (2019), ARXIV-1910.10683. T5’s paper, and the origin of C4 — the heuristically filtered Common Crawl slice that forty-six documents here train on. Filed for the corpus: the record cites C4 constantly and had no document behind it. | Active |
| LIT-426 v2 | The Pile: An 800GB Dataset of Diverse Text for Language Modeling | Gao et al. (2020), ARXIV-2101.00027. 800GB assembled from 22 named sub-corpora with declared weights. The named domains and the default mixture are why the data-mixing line in this record is measured on it — and every “over The Pile’s default weights” number is relative to a choice its authors made, not to an optimum. | Active |
| LIT-427 | SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model | Ben Allal et al. (2025), ARXIV-2502.02737. The model report behind SmolLM2-135M-Instruct, which is the base LIT-129’s whole experiment runs on — and the source of the overtraining ratio that makes a 135M model worth fine-tuning at all. | Active |
| LIT-432 | Finetuned Language Models Are Zero-Shot Learners | Wei et al. (2021), ARXIV-2109.01652. FLAN, and the name “instruction tuning”: fine-tune on 60+ existing NLP tasks rewritten as natural-language instructions and zero-shot performance on unseen task types improves. The rival post-training answer to LIT-377’s, and the one InstructGPT measured itself against. | Active |
| LIT-441 | Pre-training under infinite compute | Kim et al. (2025), ARXIV-2509.14786. Under a fixed corpus and no compute limit, the standard recipe of more epochs and more parameters overfits — and the fix is regularization, at roughly 30x the customary weight decay. Ensembling independently trained models then beats parameter scaling outright, and distilling the ensemble keeps most of the gain. The paper also proposes judging such recipes by their scaling law’s asymptote rather than by loss at a budget. | Active |
| LIT-442 | Diffusion Beats Autoregressive in Data-Constrained Settings | Prabhudesai et al. (2025), ARXIV-2507.15857. Refits Muennighoff’s data-constrained scaling law for masked diffusion and gets a data-reuse half-life of 512 epochs against autoregressive training’s 32. Under a fixed corpus, AR wins at Chinchilla-optimal compute and then overfits; diffusion starts far worse and keeps improving to a lower final loss. A closed-form critical-compute threshold says which side of the crossover you are on. | Active |
| LIT-450 v3 | How do language models learn facts? Dynamics, curricula and hallucinations | Zucchet et al. (2025), ARXIV-2503.21676. Factual recall is learned in three phases with a long plateau between generic statistics and individual knowledge, and the plateau is the attention recall circuit forming — shown by patching a reference model’s attention patterns in, at which point the plateau disappears. Imbalanced data shortens the plateau, up to a point, and the paper argues it slows acquisition afterwards; a warm-up on a subset of individuals, then all of them, beat the fixed distributions tried (single seed). | Active |
| LIT-451 | Data Mixing Can Induce Phase Transitions in Knowledge Acquisition | Gu et al. (2025), ARXIV-2505.18091. Knowledge acquired from a knowledge-dense dataset scales linearly with model size when that dataset is trained on alone — and stops doing so once it is mixed into web text. Below a critical mixing ratio the model memorises almost nothing however long it trains; below a critical model size, likewise. The critical ratio follows a power law in model size, so the transitions are predictable. | Active |
| LIT-484 v2 | TinyStories: How Small Can Language Models Be and Still Speak Coherent English? | Eldan and Li (2023), ARXIV-2305.07759. Restrict the training corpus to what a three-year-old’s vocabulary can say and a model below 10M parameters — or one with a single transformer block — writes fluent, grammatical, multi-paragraph stories. The finding is about the corpus, not the architecture: what stops small models speaking is the breadth of what they are asked to learn, not the complexity of language. | Active |
| LIT-485 | Parameterized Synthetic Text Generation with SimpleStories | Finke et al. (2025), ARXIV-2504.09184. Sampling three words from a child’s vocabulary was not enough: 59.38% of TinyStories begins “once upon a time”. Parameterize the generating prompt by theme, topic, style, narrative feature, grammar feature and author persona, and constrain the opening part of speech and letter, and every diversity metric moves. The one clean ablation in the paper is about something else — a 4,096-token domain tokenizer beats GPT-2’s 50,257 by +26.7 coherence and +27.5 quality at fixed data and architecture. | Active |
| LIT-565 | FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark | Fang et al. (2025), ARXIV-2509.09680. 6 million FLUX-generated images with 20 million bilingual captions, organized by six characteristics (imagination, entity, text rendering, style, affection, composition), and with “Generation Chain-of-Thought” breakdowns. Curation cost 15,000 A100 GPU-days. Released with PRISM-Bench, seven tracks scored by vision-language models, used to evaluate 19 text-to-image models. A dataset and a benchmark, with no training ablation read here. | Active |
| LIT-583 | Highly accurate protein structure prediction with AlphaFold | Jumper et al. (Nature 2021), DOI 10.1038/s41586-021-03819-2. AlphaFold 2 predicts protein structure from sequence and a multiple sequence alignment. At CASP14 its median backbone error was 0.96 Å r.m.s.d.95, against 2.8 Å for the next best method. Its per-residue confidence, pLDDT, tracks true accuracy (r = 0.76 over 10,795 chains). Self-distillation on its own confident predictions for 350,000 unlabeled sequences improves it. Accuracy falls sharply when the alignment has fewer than about 30 sequences. | Active |
| LIT-591 | A Simple Framework for Contrastive Learning of Visual Representations | Chen et al. (2020), ARXIV-2002.05709. Strips contrastive learning to augment, encode, project, NT-Xent — no memory bank, no special architecture — and then ablates each piece. The two findings that outlived the method: the augmentation pair is the pretext task, and the projection head is a sacrificial layer you throw away. | Active |
| LIT-598 | Unsupervised Learning of Visual Features by Contrasting Cluster Assignments | Caron et al. (2020), ARXIV-2006.09882. Predict one view’s cluster code from another view’s representation instead of comparing features pairwise, with an equipartition constraint doing the anti-collapse work. 75.3% ImageNet linear — and multi-crop, which is the part that transplants into other methods for 2-4 points. | Active |
| LIT-599 v3 | DINOv2: Learning Robust Visual Features without Supervision | Oquab et al. (2023), ARXIV-2304.07193. The thesis is that existing self-supervised objectives already produce general-purpose features if the data is curated enough, so the contribution is a retrieval-based curation pipeline and the engineering to scale it — not a new objective. Frozen features, no finetuning, competitive with the best open weakly-supervised models. | Active |
| LIT-619 v2 | Wan: Open and Advanced Large-Scale Video Generative Models | Wan Team, Alibaba (2025), ARXIV-2503.20314. The open Wan2.1 video models, 1.3B and 14B, and the full recipe behind them. The recipe covers data curation, a 127M causal 3D VAE compressing 4×8×8, a cross-attention DiT trained by rectified flow on an image-then-video resolution curriculum, the 2D context parallelism it needs, and a mask-conditioned image-to-video variant. The abstract claims video scaling laws in data and model size, but the report plots no scaling curve. | Active |
| LIT-625 v2 | Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets | Blattmann, Dockhorn, Kulal et al., Stability AI (2023), ARXIV-2311.15127. Video LDM’s architecture is held fixed, and the paper studies what the data does. Curating 4× fewer clips beats the uncurated set, and the advantage survives high-quality fine-tuning. It is the only controlled video-curation study in this line. One of its own filters lost the ablation and was kept anyway. | Active |
| LIT-648 | Improving Image Generation with Better Captions | Betker et al., OpenAI and Microsoft (2023), no arXiv id. Identical T5-conditioned latent diffusion models are trained on the same images with alt-text, short synthetic or descriptive synthetic captions. The descriptive ones give higher CLIP score, and more synthetic in the blend is better up to the 95% tested. The controlled evidence is CLIP score alone, at 256px, and on ground-truth prompts the gain is inside the checkpoint noise. The paper does not describe the DALL-E 3 model, and it says its DALL-E 3 comparisons are not a measurement of recaptioning. | Active |
| LIT-652 v2 | Video generation models as world simulators | Brooks, Peebles et al., OpenAI (2024), no arXiv id. Sora is a latent diffusion transformer over spacetime patches of a video autoencoder’s latents. It is trained jointly on images and videos at their native durations, resolutions and aspect ratios, with captions from a DALL-E 3-style recaptioner. The report withholds model and implementation details, and it has no table, number, metric or baseline. Every claim, including the scaling comparison and the square-crop ablation, is shown by example videos only. The video reports filed after it cite it and benchmark against it, and it supplies no measurement of its own. | Active |
| LIT-667 | Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization | Wang, Yue, Su and Sun (2024), ARXIV-2405.15071. Grokking outside algorithmic data, on reasoning over stored facts — and a named correction to the critical-data-size hypothesis the record holds as THEORY-071. With the inferred/atomic ratio φ held at 9.0, varying the training-set size changes nothing; varying φ at fixed size moves grokking speed monotonically and at φ = 18.0 all but removes it. Also traces the generalizing circuit for two tasks and explains from its layout why composition never generalizes out of distribution while comparison does. | Active |
| LIT-671 | RoBERTa: A Robustly Optimized BERT Pretraining Approach | Liu, Ott, Goyal, Du, Joshi, Chen, Levy, Lewis, Zettlemoyer and Stoyanov (2019), ARXIV-1907.11692. A replication study that finds BERT was significantly undertrained and, more usefully here, that BERT’s own NSP ablation could not have shown what it claimed: it removed the loss while keeping the paired-segment input format. Separate the two and removing NSP matches or slightly improves. The paper is careful about its own margins — dynamic masking is “comparable or slightly better”, and byte-level BPE is adopted despite being slightly worse. | Active |
| LIT-681 | Cramming: Training a Language Model on a Single GPU in One Day | Geiping and Goldstein (ICML 2023), ARXIV-2212.14034. A BERT-style MLM trained from scratch for 24 hours on one consumer GPU, which reaches 79.2–80.4 GLUE against BERT-base’s 80.9. The transferable finding is negative: model shape does not beat the scaling law at this budget. Every reshape it tried lands near the same final loss, because a smaller model’s extra throughput cancels its slower learning. So the only architectural lever is to cut step time at constant parameter count. Its data result is the cheapest filter in the record: drop text the tokenizer compresses badly. | Active |
| LIT-685 | Sub-Scaling Laws: On the Role of Data Density and Training Strategies in LLMs | Chen, Wang, Xiao, Wang, Chen, Cai, He and Wang (ACL 2025), ARXIV-2507.10613, published at ACL as “Revisiting Scaling Laws for Language Models: The Role of Data Quality and Training Strategies”. It claims that gains decelerate faster than a power law when data is “dense” or when tokens per parameter run high, and fits a modified Chinchilla law over more than 400 models from 20M to 7B. None of it should be leaned on. The density metric’s embedding, dimension and cluster count are never stated. The fitted law has no density term, only two logistic multipliers on D/N bounded in [1.5, 2). And the paper’s own Table 4 contradicts its “across all model sizes”. | Active |
| LIT-691 | Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation | Ke, Obukhov, Huang, Metzger, Daudt and Schindler (CVPR 2024), ARXIV-2312.02145. Marigold fine-tunes Stable Diffusion v2’s U-Net on 74K synthetic images to denoise a depth latent conditioned on the image latent, with the VAE frozen. It samples 50 DDIM steps and ensembles 10 runs, and ranks first on average across five zero-shot depth benchmarks. Its step-count curve, one step at ≈34% NYUv2 AbsRel, was measured under diffusers’ default leading spacing, which the paper never names. That is the scheduler bug LIT-687 corrects. The claim that a generative prior is what helps is not tested. | Active |
| LIT-739 | TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems | Abadi et al. (2016), ARXIV-1603.04467 — the TensorFlow white paper (dated November 2015). Describes TensorFlow’s dataflow-graph interface and its execution from a single device to clusters; among its graph features are input operations that read from storage directly and queues that let input be prefetched from disk while the previous batch is still being computed. | Active |
| LIT-743 | Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation | Bahmani et al. (2025), ARXIV-2509.19296 — Lyra. Train a 3D Gaussian Splatting decoder on the latents of a frozen camera-controlled video diffusion model (GEN3C), supervised only by that model’s own RGB decodings along six camera trajectories per image, with no real multi-view data. Decoding latents rather than pixels is what makes 726 views per scene fit in memory. Single image to 3D beats the reported numbers of ZeroNVS, ViewCrafter, Wonderland and Bolt3D on RealEstate10K, DL3DV and Tanks-and-Temples, and adding real multi-view data to the synthetic supervision does not help. | Active |
| LIT-744 | Text Embeddings by Weakly-Supervised Contrastive Pre-training | Wang et al. (2022), ARXIV-2212.03533 — E5. A BERT-initialised bi-encoder trained with InfoNCE and in-batch negatives (batch 32,768) on ~270M web text pairs, cut from ~1.3B by a consistency filter that keeps a pair only if a model trained on the noisy set ranks its passage in the top 2 of 1M. With no labelled data the base model averages 42.9 nDCG@10 on 15 BEIR datasets against BM25’s 41.7, the first unsupervised model reported above BM25; after fine-tuning on MS-MARCO, NQ and NLI the 330M large model averages 61.4 on 56 MTEB English datasets, above the 4.8B GTR-xxl and Sentence-T5-xxl. | Active |
| LIT-745 | SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features | Tschannen et al. (2025), ARXIV-2502.14786. Keep SigLIP’s architecture and sigmoid loss, and add a captioning-and-grounding decoder loss (LocCa) throughout, self-distillation and masked prediction for the last 20% of training, a 90/10 English/multilingual mix with de-biasing filters, and distillation through active data selection for the smallest models. It beats SigLIP at every size and resolution compared on zero-shot classification, retrieval, VLM transfer, dense probing and localization (RefCOCO val, L/256: 67.3 → 86.0). None of the ingredients is ablated, so the gains cannot be credited to any one of them. | Active |
| LIT-755 | Wonderland: Navigating 3D Scenes from a Single Image | Liang et al. (2024), ARXIV-2412.12091 — Wonderland (CVPR 2025). Fine-tune CogVideoX-5B-I2V for camera control with a ControlNet branch plus LoRA, then train a reconstructor (LaLRM) that regresses 3D Gaussians directly from the video latents rather than decoded frames. Reading latents beat reading RGB at matched cost (RealEstate10K PSNR 27.10 against 25.06). The reconstructor is trained on captured RealEstate10K, ACID and DL3DV first, then fine-tuned on those plus 20K videos the video model generated, supervised by its own decoded frames; that addition moved benchmark PSNR by at most 0.09. | Active |
| LIT-757 | Promptagator: Few-shot Dense Retrieval From 8 Examples | Dai et al. (2022), ARXIV-2209.11755 — Promptagator. Prompt a 137B FLAN with at most 8 examples from the target task to write queries for its corpus, keep a generated pair only if a retriever trained on the generated pairs ranks its source passage top-1, and train a 110M dual encoder on what is left. On 11 BEIR datasets it averages 47.8 nDCG@10 against 46.6 for SPLADE v2 and 46.2 for ColBERT v2, both trained on MS MARCO. The filter is worth 2.5 points on average and lowers three of the eleven datasets. | Active |
| LIT-782 | Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation | Kirstain, Polyak, Singer, Matiana, Penna and Levy, Tel Aviv University and Stability AI (2023), ARXIV-2305.01569. A web app that generates two images per prompt and logs which one the user keeps, or a tie. The paper’s version holds 583,747 training pairs over 37,523 prompts from 4,375 users. PickScore is CLIP-H fine-tuned on those pairs with a reward-model loss. It predicts held-out user choices at 70.5% under a tie-aware accuracy where chance is 56.8%, against 68.0% for the authors’ expert annotators. Its ranking of 45 model configurations correlates 0.790 with real users’ Elo. On MS-COCO captions FID ranks nine configurations at −0.900 against human experts, because people prefer higher guidance. Every validation is in-distribution or uses colleague annotators. Diffusion-DPO trains on this data and Flow-GRPO optimizes PickScore as a reward. | Active |