Papers

The reading list behind the tools: 65 papers, 65 to-read. Status moves from to-read through reading, read, reproducing to reproduced or not-reproduced, and write-up once the note is published.

Citation rule: a title links to its source only after we fetched that source (arxiv.org, doi.org, a proceedings page or the standard's own page) and logged the fetch. Entries marked "citation to verify" have not been checked yet and carry no link.

PaperAreaClaimStatusNotes
Document binarization with automatic parameter tuning
Nicholas R. Howe · 2013 · International Journal on Document Analysis and Recognition (IJDAR)
document-imaging
pagetriage
Binarisation parameters can be tuned automatically per image, so a degraded-document binariser needs no hand-set thresholds. to-read citation verified 2026-11-09 against doi.org (CSL JSON); online 2012, IJDAR volume 2013
A Laplacian energy for document binarization
Nicholas R. Howe · 2011
document-imaging
pagetriage
A Laplacian energy minimised by graph cuts binarises degraded documents better than local thresholding. to-read citation to verify
A Threshold Selection Method from Gray-Level Histograms
Nobuyuki Otsu · 1979 · IEEE Transactions on Systems, Man, and Cybernetics
document-imaging
pagetriage
A single global threshold that maximises between-class variance separates foreground from background with no parameters. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
Adaptive document image binarization
J. Sauvola, M. Pietikäinen · 2000 · Pattern Recognition
document-imaging
pagetriage
A threshold adapted to the local mean and standard deviation copes with uneven illumination and stains on document images. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
A Computational Approach to Edge Detection
John Canny · 1986 · IEEE Transactions on Pattern Analysis and Machine Intelligence
document-imaging
pagetriage
Framing edge detection as an optimisation of detection, localisation and single response gives a principled detector. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
An experimental comparison of min-cut/max- flow algorithms for energy minimization in vision
Y. Boykov, V. Kolmogorov · 2004 · IEEE Transactions on Pattern Analysis and Machine Intelligence
document-imaging
pagetriage
A min-cut/max-flow algorithm designed for vision graphs runs several times faster than general-purpose ones. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
Skew angle detection of digitized Indian script documents
B.B. Chaudhuri, U. Pal · 1997 · IEEE Transactions on Pattern Analysis and Machine Intelligence
document-imaging
pagetriage
The headline stroke of Devanagari and Bangla lets skew be detected in a way specific to Indian scripts. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
UAX #29: Unicode Text Segmentation
Unicode Consortium · Unicode Standard Annex #29
text-segmentation
glyphsift
Default, language-independent rules determine grapheme cluster, word and sentence boundaries in Unicode text. to-read citation verified 2026-11-09 against the standard's own page
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
Penedo, Guilherme, Kydlíček, Hynek, allal, Loubna Ben, et al. · 2024 · arXiv:2406.17557
data-quality
glyphsift
Careful filtering and deduplication of Common Crawl yields a 15 trillion token dataset that trains better models than earlier open sets. to-read citation verified 2026-11-09 against arxiv.org abstract page
DataComp-LM: In search of the next generation of training sets for language models
Li, Jeffrey, Fang, Alex, Smyrnis, Georgios, et al. · 2024 · arXiv:2406.11794
data-quality
glyphsift
A standardised testbed shows that model-based data filtering matters more than raw scale for training-set quality. to-read citation verified 2026-11-09 against arxiv.org abstract page
ByT5: Towards a token-free future with pre-trained byte-to-byte models
Xue, Linting, Barua, Aditya, Constant, Noah, et al. · 2021 · arXiv:2105.13626
small-models Byte-level models with no tokenizer compete with token-based models and are more robust to noisy text. to-read citation verified 2026-11-09 against arxiv.org abstract page
Efficient Memory Management for Large Language Model Serving with PagedAttention
Kwon, Woosuk, Li, Zhuohan, Zhuang, Siyuan, et al. · 2023 · arXiv:2309.06180
inference Paging the KV cache like virtual memory removes fragmentation and raises serving throughput two to four times. to-read citation verified 2026-11-09 against arxiv.org abstract page
LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding
Elhoushi, Mostafa, Shrivastava, Akshat, Liskovich, Diana, et al. · 2024 · arXiv:2404.16710
inference Layer dropout and an early-exit loss during training let a model exit early and self-speculate for faster decoding. to-read citation verified 2026-11-09 against arxiv.org abstract page
Efficient Streaming Language Models with Attention Sinks
Xiao, Guangxuan, Tian, Yuandong, Chen, Beidi, et al. · 2023 · arXiv:2309.17453
kv-cache Keeping a few initial attention-sink tokens plus a sliding window lets a model stream over unbounded input. to-read citation verified 2026-11-09 against arxiv.org abstract page
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
Xiao, Guangxuan, Tang, Jiaming, Zuo, Jingwei, et al. · 2024 · arXiv:2410.10819
kv-cache Only retrieval heads need the full KV cache; streaming heads can use a short window, cutting memory sharply. to-read citation verified 2026-11-09 against arxiv.org abstract page
ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference
Liu, Xiang, Tang, Zhenheng, Dong, Peijie, et al. · 2025 · arXiv:2502.00299
kv-cache Compressing the KV cache in semantic chunks rather than per token preserves quality at lower memory. to-read citation verified 2026-11-09 against arxiv.org abstract page
The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
Ma, Shuming, Wang, Hongyu, Ma, Lingxiao, et al. · 2024 · arXiv:2402.17764
small-models Ternary weights match full-precision quality at 3 billion parameters with far less memory and energy. to-read citation verified 2026-11-09 against arxiv.org abstract page
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Gu, Albert, Dao, Tri · 2023 · arXiv:2312.00752
inference Selective state-space models match transformers with linear-time modelling and constant-memory inference. to-read citation verified 2026-11-09 against arxiv.org abstract page
RWKV-7 "Goose" with Expressive Dynamic State Evolution
Peng, Bo, Zhang, Ruichong, Goldstein, Daniel, et al. · 2025 · arXiv:2503.14456
inference A generalised delta rule gives an RNN with expressive state evolution and constant memory per token. to-read citation verified 2026-11-09 against arxiv.org abstract page
Improving language models by retrieving from trillions of tokens
Borgeaud, Sebastian, Mensch, Arthur, Hoffmann, Jordan, et al. · 2021 · arXiv:2112.04426
retrieval Retrieving from a trillion-token database lets a comparatively small model match far larger ones. to-read citation verified 2026-11-09 against arxiv.org abstract page
FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations
Wang, Ziyao, Shen, Zheyu, He, Yexiao, et al. · 2024 · arXiv:2409.05976
training Stacking heterogeneous LoRA adapters allows federated fine-tuning without the noise of naive aggregation. to-read citation verified 2026-11-09 against arxiv.org abstract page
Editing Models with Task Arithmetic
Ilharco, Gabriel, Ribeiro, Marco Tulio, Wortsman, Mitchell, et al. · 2022 · arXiv:2212.04089
model-merging Adding and subtracting task vectors (weight deltas) edits model behaviour without retraining. to-read citation verified 2026-11-09 against arxiv.org abstract page
TIES-Merging: Resolving Interference When Merging Models
Yadav, Prateek, Tam, Derek, Choshen, Leshem, et al. · 2023 · arXiv:2306.01708
model-merging Trimming small deltas, electing signs and merging only agreeing parameters reduces interference between merged models. to-read citation verified 2026-11-09 against arxiv.org abstract page
LLM as Runtime Error Handler
(authors to verify) · 2024
agents A language model can catch runtime errors in a running program and propose or apply a fix. to-read citation to verify
Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs
Malkov, Yu. A., Yashunin, D. A. · 2016 · arXiv:1603.09320
vector-search
localdata
A hierarchy of navigable small-world graphs gives logarithmic-time approximate nearest-neighbour search. to-read citation verified 2026-11-09 against arxiv.org abstract page
Product Quantization for Nearest Neighbor Search
H Jégou, M Douze, C Schmid · 2011 · IEEE Transactions on Pattern Analysis and Machine Intelligence
vector-search
localdata
Quantising sub-vectors separately gives compact codes with accurate asymmetric distance estimation. to-read citation verified 2026-11-09 against doi.org (CSL JSON)
Fast Inference from Transformers via Speculative Decoding
Leviathan, Yaniv, Kalman, Matan, Matias, Yossi · 2022 · arXiv:2211.17192
inference A small draft model proposes tokens that the large model verifies in parallel, speeding decoding two to three times with identical output. to-read citation verified 2026-11-09 against arxiv.org abstract page
LoRA: Low-Rank Adaptation of Large Language Models
Hu, Edward J., Shen, Yelong, Wallis, Phillip, et al. · 2021 · arXiv:2106.09685
training Freezing the weights and training low-rank update matrices fine-tunes large models with a small fraction of the parameters. to-read citation verified 2026-11-09 against arxiv.org abstract page
Supply chain security for ML
OpenSSF Model Signing project (Sigstore) · sigstore/model-transparency (GitHub)
supply-chain
runtime
Signing model artefacts with Sigstore gives verifiable provenance for machine-learning models. to-read citation verified 2026-11-09 against the standard's own page
LLM01:2025 Prompt Injection
OWASP Gen AI Security Project · 2025 · OWASP Top 10 for LLM Applications 2025
security
runtime
Prompt injection is the top risk for LLM applications; the defence is layered controls, not a single filter. to-read citation verified 2026-11-09 against the standard's own page
Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
Jon Saad-Falcon, Avanika Narayan, Hakki Orhun Akengin, et al. · 2025 · arXiv:2511.07885
local-inference
runtime
On 1M real single-turn chat and reasoning queries, local models of up to 20B active parameters answered 88.7% successfully, and intelligence per watt rose 5.3x between 2023 and 2025. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: Conference on Neural Information Processing Systems (NeurIPS) 2026
Minions: Cost-efficient Collaboration Between On-device and Cloud Language Models
Avanika Narayan, Dan Biderman, Sabri Eyuboglu, et al. · 2025 · arXiv:2502.15964
local-inference
dockit
A small on-device model that reads the whole document, working with a cloud model that only plans, sends far less to the cloud: 30.4x less remote spend at 87% of the cloud model's quality for a naive chat protocol, and 5.7x less at 97.9% for MinionS. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Small Language Models are the Future of Agentic AI
Peter Belcak, Greg Heinrich, Shizhe Diao, et al. · 2025 · arXiv:2506.02153
small-models
runtime
Position paper: small language models are powerful enough, a better fit and more economical for the repetitive, narrow calls inside agentic systems, with mixed-model systems where general conversation is needed. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation
Junyuan Zhang, Qintong Zhang, Bin Wang, et al. · 2024 · arXiv:2412.02592
ocr
pagetriage
OCR noise, both semantic and formatting, degrades retrieval-augmented answers, and none of the OCR solutions the benchmark tested builds a high-quality knowledge base for RAG. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: Accepted by ICCV 2025
Can OCR-VLMs Read Devanagari? A Stress-Test Benchmark and Post-Correction Study
Aditya Pratap Singh · 2026 · arXiv:2606.29213
ocr
glyphsift
On 300 real printed Hindi scans nine of the ten OCR systems collapse, synthetic renders badly overstate Devanagari quality, and strong English OCR does not predict Indic OCR. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems
Ali Faraz, Raja Kolla, Ashish Kulkarni, Shubham Agarwal · 2026 · arXiv:2602.16430
ocr
dockit
For Indic OCR, fine-tuning an existing OCR model gave better accuracy-latency trade-offs than training a new vision-language OCR end to end (the Chitrapathak series), and a separate model series reads key fields from nine kinds of Indian government document. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); the paper title differs from the short name "Chitrapathak" used in the research list; the series name is in the abstract
Seeing Straight: Document Orientation Detection for Efficient OCR
Suranjan Goswami, Abhinav Ravi, Raja Kolla, et al. · 2025 · arXiv:2511.04161
ocr
pagetriage
A lightweight rotation classifier and the OCR-Rotation-Bench benchmark (ORB-En and ORB-Indic, 11 Indic languages) show that correcting page orientation before OCR raises OCR accuracy, by up to 14% for closed models and up to 4x for open-weights models in the simulated setting. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); the paper title differs from the benchmark name "OCR-Rotation-Bench" used in the research list; the benchmark name is in the abstract
Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets
Julia Kreutzer, Isaac Caswell, Lisa Wang, et al. · 2021 · arXiv:2103.12028
data-quality
glyphsift
A manual audit of 205 language-specific corpora from five public web-crawled datasets found at least 15 with no usable text, a significant fraction with under 50% sentences of acceptable quality, and mislabeled or nonstandard language codes. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv journal_ref field: Transactions of the Association for Computational Linguistics (2022) 10: 50-72
FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language
Guilherme Penedo, Hynek Kydlíček, Vinko Sabolčec, et al. · 2025 · arXiv:2506.20920
data-quality
glyphsift
A pre-training data curation pipeline based on FineWeb that adapts automatically to any language, ablated on nine languages, yields non-English corpora that train better models than earlier datasets. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); the research list gave arXiv 2502.10361 for FineWeb2, which is a different paper (Messmer et al., model-based data selection) and is not added; this entry is the FineWeb2 paper found by its own title
GlotLID: Language Identification for Low-Resource Languages
Amir Hossein Kargaran, Ayyoob Imani, François Yvon, Hinrich Schütze · 2023 · arXiv:2310.16248
data-quality
glyphsift
GlotLID-M identifies 1665 languages and beats four baselines (CLD3, FT176, OpenLID, NLLB) when balancing F1 against false positive rate, with an analysis of why low-resource language identification is hard. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: EMNLP 2023
IndicLLMSuite: A Blueprint for Creating Pre-training and Fine-Tuning Datasets for Indian Languages
Mohammed Safi Ur Rahman Khan, Priyam Mehta, Ananth Sankar, et al. · 2024 · arXiv:2403.06350
data-quality
glyphsift
A blueprint and resource suite for Indic LLM data across 22 languages (251B tokens, 74.8M instruction-response pairs) with a clean open pipeline for crawling, cleaning, flagging and deduplicating web, PDF and video sources. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: ACL-2024 Outstanding Paper
MILU: A Multi-task Indic Language Understanding Benchmark
Sshubam Verma, Mohammed Safi Ur Rahman Khan, Vishwajeet Kumar, et al. · 2024 · arXiv:2411.02538
evaluation
runtime
MILU is a multi-task benchmark of 8 domains and 41 subjects in 11 Indic languages; of the 42+ models evaluated, the best (GPT-4o) reached 74% average accuracy. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
MicroNN: An On-device Disk-resident Updatable Vector Database
Jeffrey Pound, Floris Chabert, Arjun Bhushan, et al. · 2025 · arXiv:2504.05573
vector-search
localdata
An embedded, disk-resident vector search engine for low-resource devices that supports continuous inserts and deletes and queries that combine nearest-neighbour search with attribute filters. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
LEANN: A Low-Storage Vector Index
Yichuan Wang, Zhifei Li, Shu Liu, et al. · 2025 · arXiv:2506.08276
vector-search
localdata
A vector index that recomputes embeddings on the fly and compresses a proximity graph, using a small fraction of storage (for example 5% of the original data) and cutting index size by up to 50x against conventional indices. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
EdgeRAG: Online-Indexed RAG for Edge Devices
Korakit Seemakhupt, Sihang Liu, Samira Khan · 2024 · arXiv:2412.21023
vector-search
localdata
Pruning embeddings inside clusters and generating them on demand lets retrieval-augmented generation fit in memory on a constrained edge device, with lower latency than the IVF baseline and similar generation quality. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
RaBitQ: Quantizing High-Dimensional Vectors with a Theoretical Error Bound for Approximate Nearest Neighbor Search
Jianyang Gao, Cheng Long · 2024 · arXiv:2405.12497
vector-search
localdata
Quantises D-dimensional vectors into D-bit strings with a sharp theoretical error bound and SIMD-friendly distance estimation, beating product quantisation on the accuracy-efficiency trade-off. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: The paper has been accepted by SIGMOD 2024
Matryoshka Representation Learning
Aditya Kusupati, Gantavya Bhatt, Aniket Rege, et al. · 2022 · arXiv:2205.13147
vector-search
localdata
Trains one embedding whose leading dimensions are themselves usable representations, so a downstream system can pick a smaller size to fit its compute and storage limits. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node
Suhas Jayaram Subramanya, Devvrit, Harsha Vardhan Simhadri, et al. · 2019 · Advances in Neural Information Processing Systems 32 (NeurIPS 2019)
vector-search
localdata
A graph-based index kept on an inexpensive SSD instead of main memory can index, store and search a billion-point database on one workstation with 64GB of RAM at high recall and low latency. to-read citation verified 2026-11-14 against the NeurIPS 2019 proceedings page (papers.nips.cc); author order and list as shown on that page, abbreviated
The Probabilistic Relevance Framework: BM25 and Beyond
Stephen Robertson, Hugo Zaragoza · 2009 · Foundations and Trends® in Information Retrieval
retrieval
localdata
Presents the probabilistic relevance framework and the BM25 and BM25F ranking functions that grew out of it. to-read citation verified 2026-11-14 against doi.org (CSL JSON); informs lexical plus vector retrieval in localdata; the research list dates it 2009 and Crossref gives volume 4, issues 1-2
Reciprocal rank fusion outperforms condorcet and individual rank learning methods
Gordon V. Cormack, Charles L A Clarke, Stefan Buettcher · 2009 · Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval
retrieval
localdata
Fusing ranked lists by reciprocal rank outperforms Condorcet fusion and individual rank-learning methods. to-read citation verified 2026-11-14 against doi.org (CSL JSON); claim written from the title, abstract not read; SIGIR 2009
EmbeddingGemma: Powerful and Lightweight Text Representations
Henrique Schechter Vera, Sahil Dua, Biao Zhang, et al. · 2025 · arXiv:2509.20354
retrieval
localdata
A 300M-parameter open text embedding model that leads MTEB among models under 500M parameters and keeps its lead when the weights are quantised or the embeddings truncated. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
MMTEB: Massive Multilingual Text Embedding Benchmark
Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, et al. · 2025 · arXiv:2502.13595
retrieval
localdata
A community-built benchmark of over 500 tasks across 250+ languages for text embeddings; the best publicly available model on it is multilingual-e5-large-instruct at 560M parameters. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: Accepted for ICLR
Survivable key compromise in software update systems
Justin Samuel, Nick Mathewson, Justin Cappos, Roger Dingledine · 2010 · Proceedings of the 17th ACM conference on Computer and communications security
supply-chain
runtime
Describes a software update system designed so that the compromise of some signing keys does not by itself let an attacker push malicious updates. to-read citation verified 2026-11-14 against doi.org (CSL JSON); claim written from the title, abstract not read; the research list calls this paper "The Update Framework" and the title shown is the one on the record; informs the open key revocation, rollback and offline-update findings
in-toto: Providing farm-to-table guarantees for bits and bytes
Santiago Torres-Arias, Hammad Afzali, Trishank Karthik Kuppusamy, et al. · 2019 · 28th USENIX Security Symposium (USENIX Security 19)
supply-chain
control-plane
A framework that cryptographically ties together every step of a software supply chain, from source to deployment, so an end user can verify the whole chain and not only the final artifact. to-read citation verified 2026-11-14 against the USENIX Security 2019 presentation page
Sigstore: Software Signing for Everybody
Zachary Newman, John Speed Meyers, Santiago Torres-Arias · 2022 · Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security
supply-chain
control-plane
Describes Sigstore, a software signing system meant to be usable by everybody. to-read citation verified 2026-11-14 against doi.org (CSL JSON); claim written from the title, abstract not read
Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, et al. · 2023 · arXiv:2302.12173
security
dockit
LLM-integrated applications blur data and instructions, so an attacker can inject prompts into data likely to be retrieved and attack remotely; the paper gives a taxonomy of impacts including data theft and worming. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Defeating Prompt Injections by Design
Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan, et al. · 2025 · arXiv:2503.18813
security
control-plane
CaMeL wraps the model in a protective layer that takes control and data flow from the trusted query and uses capabilities to block exfiltration, solving 77% of AgentDojo tasks with provable security against 84% undefended. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Design Patterns for Securing LLM Agents against Prompt Injections
Luca Beurer-Kellner, Beat Buesser, Ana-Maria Creţu, et al. · 2025 · arXiv:2506.08837
security
control-plane
A set of principled design patterns for building agents with provable resistance to prompt injection, with the utility and security trade-offs of each and case studies. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
WebLLM: A High-Performance In-Browser LLM Inference Engine
Charlie F. Ruan, Yucheng Qin, Akaash R. Parthasarathy, et al. · 2024 · arXiv:2412.15803
inference
runtime
An open-source JavaScript framework for LLM inference inside the browser on WebGPU and WebAssembly, retaining up to 80% of native performance in its evaluation. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); reference only: the WebLLM rung in runtime is not yet (cut from Sprint 4)
XGrammar: Flexible and Efficient Structured Generation Engine for Large Language Models
Yixin Dong, Charlie F. Ruan, Yaxing Cai, et al. · 2024 · arXiv:2411.15100
inference
runtime
Speeds up grammar-constrained decoding by prechecking the context-independent tokens and interpreting only the context-dependent ones at run time. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: MLSys '25
SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, et al. · 2025 · arXiv:2502.02737
small-models
runtime
A 1.7B-parameter small language model overtrained on about 11 trillion tokens through a multi-stage data mix, outperforming Qwen2.5-1.5B and Llama3.2-1B. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Qwen3 Technical Report
An Yang, Anfeng Li, Baosong Yang, et al. · 2025 · arXiv:2505.09388
small-models
runtime
The Qwen3 family (0.6B to 235B parameters, dense and mixture-of-experts) unifies a thinking mode and a non-thinking mode, with a thinking budget that trades latency for quality. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Ragas: Automated Evaluation of Retrieval Augmented Generation
Shahul Es, Jithin James, Luis Espinosa-Anke, Steven Schockaert · 2023 · arXiv:2309.15217
retrieval
dockit
Reference-free metrics for evaluating retrieval-augmented generation along three dimensions (relevant retrieved context, faithful use of it, quality of the generation) without human-labelled ground truth. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API)
Enabling Large Language Models to Generate Text with Citations
Tianyu Gao, Howard Yen, Jiatong Yu, Danqi Chen · 2023 · arXiv:2305.14627
retrieval
dockit
ALCE is a benchmark and set of metrics for answers with citations (fluency, correctness, citation quality); on ELI5 even the best models lack complete citation support 50% of the time. to-read citation verified 2026-11-14 against the arxiv.org abstract page (record from the export.arxiv.org API); arXiv comment field: Accepted by EMNLP 2023; the paper title differs from the short name "ALCE" used in the research list
Local-first software: you own your data, in spite of the cloud
Martin Kleppmann, Adam Wiggins, Peter van Hardenberg, Mark McGranaghan · 2019 · Proceedings of the 2019 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software
local-first
localdata
Argues for local-first software, in which users own their data in spite of the cloud. to-read citation verified 2026-11-14 against doi.org (CSL JSON); claim written from the title, abstract not read; Onward! 2019