[
  {
    "category": "core",
    "title": "Omnilingual SONAR / OmniSONAR",
    "url": "https://arxiv.org/abs/2603.16606",
    "date": "2026-03",
    "description": "The direct SONAR successor. Retains a 1,024-dimensional sentence vector and a decoder while extending multilingual and multimodal coverage. Combines reconstruction/translation and contrastive objectives.",
    "note": "Revised June 2026. Strong multilingual results do not establish universally best exact reconstruction.",
    "priority": "yes",
    "id": "work-omnilingual-sonar-omnisonar"
  },
  {
    "category": "core",
    "title": "LatentSeal: Fast, Secure, and High-Capacity Image Watermarking with Autoencoded Text Vectors",
    "url": "https://arxiv.org/abs/2510.00799",
    "date": "2025-10",
    "description": "Contains an independently useful text autoencoder: sentences → 256-dimensional vectors → sentences. Trained for robustness to latent noise.",
    "note": "A watermarking application with a reusable text autoencoder. Released model: https://huggingface.co/Gevennou/lseal",
    "priority": "yes",
    "id": "work-latentseal-fast-secure-and-high-capacity-image-watermarking-with-autoencoded-text-vectors"
  },
  {
    "category": "core",
    "title": "QwenAR: a sentence autoencoder built from open checkpoints",
    "url": "https://github.com/IvanDrokin/QwenAR",
    "date": "2026-08",
    "description": "Combines an embedding encoder and Qwen3 decoder through a learned bridge; reconstructs text from one 1,024-dimensional vector.",
    "note": "Software/model release. Advertised exact reconstruction uses only 32 probe sentences; broader evaluation is needed. Weights: https://huggingface.co/brivangl/qwenar-0.6b",
    "priority": "yes",
    "id": "work-qwenar-a-sentence-autoencoder-built-from-open-checkpoints"
  },
  {
    "category": "core",
    "title": "Quasi-symbolic Semantic Geometry over Transformer-based Variational AutoEncoder",
    "url": "https://aclanthology.org/2025.conll-1.2/",
    "date": "2025-07",
    "description": "Structures a sentence VAE’s latent space using semantic roles and word content. Relevant to meaningful latent geometry and controlled manipulation.",
    "note": "Date is the CoNLL proceedings date; an earlier preprint date was not established.",
    "priority": "no",
    "id": "work-quasi-symbolic-semantic-geometry-over-transformer-based-variational-autoencoder"
  },
  {
    "category": "core",
    "title": "GSPT-CVAE: A New Controlled Long Text Generation Method Based on T-CVAE",
    "url": "https://doi.org/10.32604/cmc.2025.063209",
    "date": "2025-06",
    "description": "Uses graph-based text representations merged into a single latent vector for conditional long-text generation.",
    "note": "Evaluates controlled generation rather than general-purpose exact reconstruction.",
    "priority": "no",
    "id": "work-gspt-cvae-a-new-controlled-long-text-generation-method-based-on-t-cvae"
  },
  {
    "category": "interpretability",
    "title": "ParaScopes: What do Language Models Activations Encode About Future Text?",
    "url": "https://arxiv.org/abs/2511.00180",
    "date": "2025-02",
    "description": "Maps LLM activations to the SONAR embedding of an upcoming paragraph, then decodes it. A direct application by the original review’s author.",
    "note": "Initial post February 2025; paper October 2025. Decodable future information does not establish that models fully plan paragraphs. Initial post: https://www.lesswrong.com/posts/9NqgYesCutErskdmu/parascopes-do-language-models-plan-the-upcoming-paragraph",
    "priority": "yes",
    "id": "work-parascopes-what-do-language-models-activations-encode-about-future-text"
  },
  {
    "category": "interpretability",
    "title": "Latent Reasoning via Sentence Embedding Prediction",
    "url": "https://arxiv.org/abs/2505.22202",
    "date": "2025-05",
    "description": "Builds single-vector sentence autoencoders and compares reconstruction-trained representations with representations trained to predict subsequent reasoning steps. SentenceLens decodes intermediate states.",
    "note": "Originally titled Let’s Predict Sentence by Sentence; renamed October 2025. These are one work, not two.",
    "priority": "yes",
    "id": "work-latent-reasoning-via-sentence-embedding-prediction"
  },
  {
    "category": "interpretability",
    "title": "Investigating Internal Representations of Correctness in SONAR Text Autoencoders",
    "url": "https://www.lesswrong.com/posts/NNGB77BipWpHn4vRW/investigating-internal-representations-of-correctness-in",
    "date": "2025-08",
    "description": "Probes whether SONAR embeddings distinguish correct code, grammatical sentences, arithmetic, and chess states.",
    "note": "Small exploratory research post with positive results on some tasks; not a general correctness detector.",
    "priority": "no",
    "id": "work-investigating-internal-representations-of-correctness-in-sonar-text-autoencoders"
  },
  {
    "category": "interpretability",
    "title": "Investigating Representations in the Embedding in SONAR Text Autoencoders",
    "url": "https://www.lesswrong.com/posts/hNBzHC7RADt3ERuqo/investigating-representations-in-the-embedding-in-sonar-text",
    "date": "2025-09",
    "description": "Examines language identity, token position, and replacement directions in SONAR embeddings.",
    "note": "Exploratory research post using narrow synthetic settings; does not establish universal semantic arithmetic.",
    "priority": "no",
    "id": "work-investigating-representations-in-the-embedding-in-sonar-text-autoencoders"
  },
  {
    "category": "interpretability",
    "title": "Forewarned is Forearmed: When Non-Sequential Embedding Turns Into an Anomaly Detector",
    "url": "https://arxiv.org/abs/2606.30196",
    "date": "2026-06",
    "description": "Studies SONAR embedding perturbations, decoding anomalies, and consistency across repeated encoding/decoding.",
    "note": "Relevant to the reliability of manipulating vectors before decoding.",
    "priority": "no",
    "id": "work-forewarned-is-forearmed-when-non-sequential-embedding-turns-into-an-anomaly-detector"
  },
  {
    "category": "interpretability",
    "title": "Interpretability of Text Auto-Encoders using Sparse Auto-Encoders: A Sandbox for Interpreting Neuralese — artifacts",
    "url": "https://huggingface.co/nickypro/sonar-saes-comparison",
    "date": "Uncertain",
    "description": "Pochinkov and Darmawan’s released SAE comparisons and large SONAR SAE checkpoints.",
    "note": "Artifacts found; first-release date, accessible paper, and acceptance status were not verified. Model card describes an EACL 2026 submission. Checkpoints: https://huggingface.co/nickypro/sonar-saes-large",
    "priority": "no",
    "id": "work-interpretability-of-text-auto-encoders-using-sparse-auto-encoders-a-sandbox-for-interpreting-ne"
  },
  {
    "category": "generation",
    "title": "Cosmos: Compressed and Smooth Latent Space for Text Diffusion Modeling",
    "url": "https://arxiv.org/abs/2506.21170",
    "date": "2025-06",
    "description": "Learns compressed text latents aligned with pretrained LM representations, emphasizing both reconstruction and a smooth space for diffusion.",
    "note": "Uses latent sequences rather than one fixed-size vector for an entire paragraph.",
    "priority": "no",
    "id": "work-cosmos-compressed-and-smooth-latent-space-for-text-diffusion-modeling"
  },
  {
    "category": "generation",
    "title": "LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning",
    "url": "https://arxiv.org/abs/2510.04573",
    "date": "2025-10",
    "description": "Encodes reasoning into latent blocks and uses diffusion to generate or revise them before decoding.",
    "note": "Blockwise latent reasoning; not a general-purpose single-vector text autoencoder.",
    "priority": "no",
    "id": "work-ladir-latent-diffusion-enhances-llms-for-text-reasoning"
  },
  {
    "category": "generation",
    "title": "Continuous Autoregressive Language Models — CALM",
    "url": "https://arxiv.org/abs/2510.27688",
    "date": "2025-10",
    "description": "Compresses small token chunks into individual vectors, then predicts vectors autoregressively. Explores faithful reconstruction versus robustness to latent noise.",
    "note": "A useful autoencoder design reference. A vector represents a small chunk, not a whole paragraph.",
    "priority": "yes",
    "id": "work-continuous-autoregressive-language-models-calm"
  },
  {
    "category": "generation",
    "title": "LaDi-RL: Latent Diffusion Reasoning Prevents Entropy Collapse in Reinforcement Learning",
    "url": "https://arxiv.org/abs/2602.01705",
    "date": "2026-02",
    "description": "Extends latent diffusion reasoning to reinforcement learning, assigning credit across latent trajectories and their decoded completions.",
    "note": "Reported gains are tied to the paper’s reasoning and RL settings.",
    "priority": "no",
    "id": "work-ladi-rl-latent-diffusion-reasoning-prevents-entropy-collapse-in-reinforcement-learning"
  },
  {
    "category": "generation",
    "title": "Stop-Think-AutoRegress: Language Modeling with Latent Diffusion Planning",
    "url": "https://arxiv.org/abs/2602.20528",
    "date": "2026-02",
    "description": "Alternates autoregressive text generation with latent semantic planning.",
    "note": "An adjacent planning architecture rather than a standalone sentence autoencoder.",
    "priority": "no",
    "id": "work-stop-think-autoregress-language-modeling-with-latent-diffusion-planning"
  },
  {
    "category": "generation",
    "title": "CoDAR: Continuous Diffusion Language Models are More Powerful Than You Think",
    "url": "https://arxiv.org/abs/2603.02547",
    "date": "2026-03",
    "description": "Uses an autoregressive decoder to turn denoised continuous representations into text, addressing limitations of simple token-by-token rounding.",
    "note": "Continuous embedding sequences, not one pooled sentence vector.",
    "priority": "no",
    "id": "work-codar-continuous-diffusion-language-models-are-more-powerful-than-you-think"
  },
  {
    "category": "generation",
    "title": "Continuous Latent Diffusion Language Model — Cola DLM",
    "url": "https://arxiv.org/abs/2605.06548",
    "date": "2026-05",
    "description": "Separates a text VAE from a diffusion model over its latent representations, with larger-scale language modelling experiments.",
    "note": "Latent-sequence generation.",
    "priority": "no",
    "id": "work-continuous-latent-diffusion-language-model-cola-dlm"
  },
  {
    "category": "generation",
    "title": "TextLDM: Language Modeling with Continuous Latent Diffusion",
    "url": "https://arxiv.org/abs/2605.07748",
    "date": "2026-05",
    "description": "Combines a transformer VAE with representation alignment and latent flow matching. Examines why reconstruction alone does not produce good generative latents.",
    "note": "Latent-sequence generation.",
    "priority": "no",
    "id": "work-textldm-language-modeling-with-continuous-latent-diffusion"
  },
  {
    "category": "generation",
    "title": "How to Train Your Latent Diffusion Language Model Jointly With the Latent Space",
    "url": "https://arxiv.org/abs/2605.07933",
    "date": "2026-05",
    "description": "Jointly trains the encoder, latent generative model, and decoder rather than treating autoencoder training as a completely separate stage.",
    "note": "Reported speedups are specific to its benchmarks and comparisons.",
    "priority": "no",
    "id": "work-how-to-train-your-latent-diffusion-language-model-jointly-with-the-latent-space"
  },
  {
    "category": "generation",
    "title": "Why Do Few-Step Text Latents Fail When Image Latents Work? Non-Commitment at Sharp Categorical Readouts",
    "url": "https://arxiv.org/abs/2606.30705",
    "date": "2026-06",
    "description": "Theoretical and diagnostic work on converting smooth continuous trajectories into sharply discrete text outputs.",
    "note": "Theoretical conclusions depend on the paper’s assumptions; not an unrestricted impossibility result.",
    "priority": "no",
    "id": "work-why-do-few-step-text-latents-fail-when-image-latents-work-non-commitment-at-sharp-categorical-r"
  },
  {
    "category": "generation",
    "title": "AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling",
    "url": "https://arxiv.org/abs/2608.02602",
    "date": "2026-08",
    "description": "Constructs a decodable latent sequence for blockwise diffusion generation; explores retaining sufficient representational capacity for text fidelity.",
    "note": "Uses a latent sequence rather than a single sentence bottleneck.",
    "priority": "no",
    "id": "work-aurora-lm-autoencoding-unified-representation-for-continuous-latent-diffusion-language-modeling"
  },
  {
    "category": "inversion",
    "title": "Universal Zero-shot Embedding Inversion — ZSInvert",
    "url": "https://arxiv.org/abs/2504.00147",
    "date": "2025-03",
    "description": "Reconstructs text without encoder-specific inversion training, using black-box queries to the target encoder.",
    "note": "Zero-shot training does not mean no encoder access, or guaranteed exact reconstruction.",
    "priority": "no",
    "id": "work-universal-zero-shot-embedding-inversion-zsinvert"
  },
  {
    "category": "inversion",
    "title": "Harnessing the Universal Geometry of Embeddings — vec2vec",
    "url": "https://arxiv.org/abs/2505.12540",
    "date": "2025-05",
    "description": "Learns translations between embedding spaces without paired examples. Relevant to connecting encoders with existing decoders.",
    "note": "Its results should not be described as universal exact text recovery.",
    "priority": "no",
    "id": "work-harnessing-the-universal-geometry-of-embeddings-vec2vec"
  },
  {
    "category": "inversion",
    "title": "LAGO: Few-shot Crosslingual Embedding Inversion Attacks via Language Similarity-Aware Graph Optimization",
    "url": "https://arxiv.org/abs/2505.16008",
    "date": "2025-05",
    "description": "Uses relationships between languages to improve crosslingual inversion with very little alignment data.",
    "note": "A few-shot inversion method, not a jointly trained text autoencoder.",
    "priority": "no",
    "id": "work-lago-few-shot-crosslingual-embedding-inversion-attacks-via-language-similarity-aware-graph-opti"
  },
  {
    "category": "inversion",
    "title": "Rethinking the Privacy of Text Embeddings: A Reproducibility Study of “Text Embeddings Reveal (Almost) As Much As Text”",
    "url": "https://arxiv.org/abs/2507.07700",
    "date": "2025-07",
    "description": "Re-examines vec2text, including sequence-length effects and noise/quantization defences.",
    "note": "Corrective evidence alongside new inversion methods.",
    "priority": "no",
    "id": "work-rethinking-the-privacy-of-text-embeddings-a-reproducibility-study-of-text-embeddings-reveal-alm"
  },
  {
    "category": "inversion",
    "title": "Zero2Text: Zero-Training Cross-Domain Inversion Attacks on Textual Embeddings",
    "url": "https://arxiv.org/abs/2602.01757",
    "date": "2026-02",
    "description": "Uses an LLM prior and online alignment to reconstruct across domains without leaked training pairs.",
    "note": "Still requires target-encoder queries.",
    "priority": "no",
    "id": "work-zero2text-zero-training-cross-domain-inversion-attacks-on-textual-embeddings"
  },
  {
    "category": "inversion",
    "title": "Embedding Inversion via Conditional Masked Diffusion Language Models",
    "url": "https://arxiv.org/abs/2602.11047",
    "date": "2026-02",
    "description": "Uses a small conditional masked diffusion model for inversion.",
    "note": "No target-encoder access at inference; this differs from requiring no paired training data.",
    "priority": "no",
    "id": "work-embedding-inversion-via-conditional-masked-diffusion-language-models"
  },
  {
    "category": "inversion",
    "title": "Denoising-Aware Inversion: Revealing Privacy Risks in Noise-Protected Text Embeddings",
    "url": "https://arxiv.org/abs/2608.18610",
    "date": "2026-08",
    "description": "Places a denoising autoencoder before text reconstruction to recover information from noisy embeddings.",
    "note": "Performance under tested noise mechanisms does not invalidate differential privacy guarantees.",
    "priority": "no",
    "id": "work-denoising-aware-inversion-revealing-privacy-risks-in-noise-protected-text-embeddings"
  },
  {
    "category": "inversion",
    "title": "How Much Do Embeddings Reveal? Inversion Attacks on Commercial Embedding Models",
    "url": "https://escholarship.org/uc/item/92k8b7p9",
    "date": "2026",
    "description": "Extends inversion evaluation across commercial embedding models and text domains, showing variation by model and content type.",
    "note": "Thesis. Exact public release date unverified.",
    "priority": "no",
    "id": "work-how-much-do-embeddings-reveal-inversion-attacks-on-commercial-embedding-models"
  },
  {
    "category": "defences",
    "title": "TextCrafter: Optimization-Calibrated Noise for Defending Against Text Embedding Inversion",
    "url": "https://arxiv.org/abs/2509.17302",
    "date": "2025-09",
    "description": "Designs perturbations intended to obstruct reconstruction while retaining useful embeddings.",
    "note": "An inversion defence rather than a new text autoencoder.",
    "priority": "no",
    "id": "work-textcrafter-optimization-calibrated-noise-for-defending-against-text-embedding-inversion"
  },
  {
    "category": "defences",
    "title": "Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacks — SPARSE",
    "url": "https://arxiv.org/abs/2602.07090",
    "date": "2026-02",
    "description": "Uses concept-sensitive masking and structured noise; relevant to which directions retain sensitive semantic information.",
    "note": "A defence and latent-structure application.",
    "priority": "no",
    "id": "work-concept-aware-privacy-mechanisms-for-defending-embedding-inversion-attacks-sparse"
  },
  {
    "category": "defences",
    "title": "Shadow Queries for Private Retrieval in Vector Databases",
    "url": "https://arxiv.org/abs/2609.04767",
    "date": "2026-09",
    "description": "Stores embeddings of generated retrieval queries in place of raw document embeddings, changing what inversion can expose.",
    "note": "First released 4 September 2026, within the search window.",
    "priority": "no",
    "id": "work-shadow-queries-for-private-retrieval-in-vector-databases"
  },
  {
    "category": "adjacent",
    "title": "SPARTA: Evaluating Reasoning Segmentation Robustness through Black-Box Adversarial Paraphrasing in Text Autoencoder Latent Space",
    "url": "https://arxiv.org/abs/2510.24446",
    "date": "2025-10",
    "description": "Optimizes SONAR vectors to generate meaning-preserving adversarial paraphrases.",
    "note": "A direct application of editable text latents; later appeared at EACL 2026.",
    "priority": "no",
    "id": "work-sparta-evaluating-reasoning-segmentation-robustness-through-black-box-adversarial-paraphrasing-"
  },
  {
    "category": "adjacent",
    "title": "Unified Vision-Language Modeling via Concept Space Alignment — V-SONAR / V-LCM",
    "url": "https://arxiv.org/abs/2603.01096",
    "date": "2026-03",
    "description": "Aligns visual representations with the SONAR concept space for decoding and multimodal concept modelling.",
    "note": "A multimodal extension.",
    "priority": "no",
    "id": "work-unified-vision-language-modeling-via-concept-space-alignment-v-sonar-v-lcm"
  },
  {
    "category": "adjacent",
    "title": "Omnilingual MT: Machine Translation for 1,600 Languages",
    "url": "https://arxiv.org/abs/2603.16309",
    "date": "2026-03",
    "description": "Associated multilingual translation work, including an encoder–decoder approach built with OmniSONAR components.",
    "note": "The paper also contains a decoder-only approach; not every model has a sentence-vector bottleneck.",
    "priority": "no",
    "id": "work-omnilingual-mt-machine-translation-for-1-600-languages"
  },
  {
    "category": "adjacent",
    "title": "SAKI-RAG: Mitigating Context Fragmentation in Long-Document RAG via Sentence-level Attention Knowledge Integration",
    "url": "https://aclanthology.org/2025.emnlp-main.63/",
    "date": "2025-11",
    "description": "Uses SentenceVAE representations and sentence-level modelling for document retrieval.",
    "note": "EMNLP proceedings date. An application rather than a new general-purpose text autoencoder.",
    "priority": "no",
    "id": "work-saki-rag-mitigating-context-fragmentation-in-long-document-rag-via-sentence-level-attention-kno"
  },
  {
    "category": "adjacent",
    "title": "TextEconomizer: Enhancing Lossy Text Compression with Denoising Transformers and Entropy Coding",
    "url": "https://arxiv.org/abs/2606.08184",
    "date": "2026",
    "description": "Lossy text compression combining autoencoding, denoising, and entropy coding.",
    "note": "2026 version/extension. An earlier TextEconomizer submission existed before the cutoff; not a wholly new first appearance.",
    "priority": "no",
    "id": "work-texteconomizer-enhancing-lossy-text-compression-with-denoising-transformers-and-entropy-coding"
  },
  {
    "category": "adjacent",
    "title": "Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations",
    "url": "https://www.transformer-circuits.pub/2026/nla/index.html",
    "date": "2026-05",
    "description": "Trains an activation verbalizer and reconstructor to explain LLM activations through natural language.",
    "note": "Reverses the bottleneck: activation → natural-language explanation → activation. Not text → vector → text.",
    "priority": "no",
    "id": "work-natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations"
  },
  {
    "category": "adjacent",
    "title": "Improving local interpretable classifier explanations exploiting self-generated semantic features",
    "url": "https://link.springer.com/article/10.1007/s00521-025-11227-0",
    "date": "2025-05",
    "description": "Uses a denoising adversarial autoencoder’s semantic neighbourhoods for classifier explanations.",
    "note": "Extends an earlier research line with a 2021 precursor.",
    "priority": "no",
    "id": "work-improving-local-interpretable-classifier-explanations-exploiting-self-generated-semantic-featur"
  },
  {
    "category": "adjacent",
    "title": "Decoder-Only LLMs can be Masked Auto-Encoders — UniMAE",
    "url": "https://aclanthology.org/2025.acl-short.57/",
    "date": "2025-07",
    "description": "Uses reconstruction-based embedding training.",
    "note": "ACL proceedings date. The decoder also receives partially visible input and is discarded after training: not a drop-in single-vector text decoder.",
    "priority": "no",
    "id": "work-decoder-only-llms-can-be-masked-auto-encoders-unimae"
  },
  {
    "category": "adjacent",
    "title": "Synergy: End-to-end Concept Model",
    "url": "https://arxiv.org/abs/2507.12769",
    "date": "2025-07",
    "description": "Learns a byte-to-concept-to-text architecture jointly.",
    "note": "Relevant to concept modelling, but not a separately trained single-vector sentence autoencoder.",
    "priority": "no",
    "id": "work-synergy-end-to-end-concept-model"
  },
  {
    "category": "adjacent",
    "title": "LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling",
    "url": "https://arxiv.org/abs/2604.11748",
    "date": "2026-04",
    "description": "Continuous-space language generation using embedding-space flow modelling.",
    "note": "Distinguish token-embedding methods from compressed sentence autoencoders.",
    "priority": "no",
    "id": "work-langflow-continuous-diffusion-rivals-discrete-in-language-modeling"
  },
  {
    "category": "adjacent",
    "title": "ELF: Embedded Language Flows",
    "url": "https://arxiv.org/abs/2605.10938",
    "date": "2026-05",
    "description": "Continuous embedding-space flow matching with a shared network for final discretization.",
    "note": "Not a learned compressed sentence autoencoder.",
    "priority": "no",
    "id": "work-elf-embedded-language-flows"
  },
  {
    "category": "adjacent",
    "title": "Continuous Diffusion Scales Competitively with Discrete Diffusion for Language",
    "url": "https://arxiv.org/abs/2605.18530",
    "date": "2026-05",
    "description": "Studies the scaling of continuous diffusion language models.",
    "note": "Continuous token embeddings rather than a single sentence bottleneck.",
    "priority": "no",
    "id": "work-continuous-diffusion-scales-competitively-with-discrete-diffusion-for-language"
  },
  {
    "category": "adjacent",
    "title": "BeamClean: Language Aware Embedding Reconstruction",
    "url": "https://arxiv.org/abs/2505.13758",
    "date": "2025-05",
    "description": "Reconstructs from obfuscated token-level embeddings using language information.",
    "note": "Token-level embeddings, not one pooled sentence vector.",
    "priority": "no",
    "id": "work-beamclean-language-aware-embedding-reconstruction"
  },
  {
    "category": "adjacent",
    "title": "DeepInvert: Semi-Supervised Embedding Inversion Against Obfuscated Language Models",
    "url": "https://arxiv.org/abs/2608.04477",
    "date": "2026-08",
    "description": "Inversion research involving obfuscated language-model embeddings.",
    "note": "Adjacent to the pooled sentence-vector setting.",
    "priority": "no",
    "id": "work-deepinvert-semi-supervised-embedding-inversion-against-obfuscated-language-models"
  },
  {
    "category": "adjacent",
    "title": "Transforming market whitespaces into new product concepts: a text embedding and inversion approach",
    "url": "https://doi.org/10.1080/09544828.2025.2568820",
    "date": "2025",
    "description": "Combines embedding compression and inversion to turn unoccupied regions of a product-description space into new textual concepts.",
    "note": "A domain-specific application; exact first-release date not established.",
    "priority": "no",
    "id": "work-transforming-market-whitespaces-into-new-product-concepts-a-text-embedding-and-inversion-approa"
  },
  {
    "category": "adjacent",
    "title": "Autoencoders in Natural Language Processing: A Comprehensive Review",
    "url": "https://doi.org/10.3390/computers15040232",
    "date": "2026-04",
    "description": "A newer survey spanning classical, variational, and masked autoencoders.",
    "note": "Broader in scope than NickyP’s single-vector review; a survey, not a novel model.",
    "priority": "no",
    "id": "work-autoencoders-in-natural-language-processing-a-comprehensive-review"
  },
  {
    "category": "candidate",
    "title": "NAT-CVAE: A Novel Method for Controlled Long Text Generation",
    "url": "https://doi.org/10.1109/AINIT65432.2025.11035169",
    "date": "2025",
    "description": "Identified in AINIT 2025 proceedings.",
    "note": "Bibliographic lead only. Insufficient primary-source content was accessible to assess its architecture or results confidently.",
    "priority": "no",
    "id": "work-nat-cvae-a-novel-method-for-controlled-long-text-generation"
  },
  {
    "category": "older",
    "title": "Cramming 1568 Tokens into a Single Vector and Back Again: Exploring the Limits of Embedding Space Capacity",
    "url": "https://arxiv.org/abs/2502.13063",
    "date": "2025-02-18",
    "description": "Demonstrates high-capacity reconstructable vectors using per-example optimization rather than a fast learned encoder.",
    "note": "One day before the original review; revised June 2025. Useful context, not a post-review first release.",
    "priority": "no",
    "id": "work-cramming-1568-tokens-into-a-single-vector-and-back-again-exploring-the-limits-of-embedding-spac"
  },
  {
    "category": "older",
    "title": "ALGEN",
    "url": "https://arxiv.org/abs/2502.11308",
    "date": "2025-02-16",
    "description": "Few-shot embedding alignment and inversion.",
    "note": "Predates the review. Later ACL publication does not make it a post-review first release.",
    "priority": "no",
    "id": "work-algen"
  },
  {
    "category": "older",
    "title": "LLM Pretraining with Continuous Concepts — CoCoMix",
    "url": "https://arxiv.org/abs/2502.08524",
    "date": "2025-02-12",
    "description": "Concept-based pretraining using sparse-autoencoder features.",
    "note": "Predates the review; adjacent rather than a sentence autoencoder.",
    "priority": "no",
    "id": "work-llm-pretraining-with-continuous-concepts-cocomix"
  },
  {
    "category": "older",
    "title": "SentenceVAE",
    "url": "https://arxiv.org/abs/2408.00655",
    "date": "2024-08",
    "description": "Sentence encoding, decoding, and next-sentence modelling.",
    "note": "A useful omission to add, but not new since the review.",
    "priority": "no",
    "id": "work-sentencevae"
  },
  {
    "category": "older",
    "title": "Large Concept Models",
    "url": "https://arxiv.org/abs/2412.08821",
    "date": "2024-12",
    "description": "Language modelling in a sentence representation space.",
    "note": "Already discussed in comments on the original post; not new since it.",
    "priority": "no",
    "id": "work-large-concept-models"
  },
  {
    "category": "older",
    "title": "MEXMA",
    "url": "https://arxiv.org/abs/2409.12737",
    "date": "2024-09",
    "description": "Multilingual sentence representation learning through crosslingual masked reconstruction.",
    "note": "A 2024 preprint with later publication activity.",
    "priority": "no",
    "id": "work-mexma"
  },
  {
    "category": "older",
    "title": "Against All Odds: Overcoming Typology, Script, and Language Confusion in Multilingual Embedding Inversion Attacks",
    "url": "https://arxiv.org/abs/2408.11749",
    "date": "2024-08",
    "description": "Crosslingual embedding inversion.",
    "note": "A 2024 preprint with later publication activity.",
    "priority": "no",
    "id": "work-against-all-odds-overcoming-typology-script-and-language-confusion-in-multilingual-embedding-in"
  }
]