Datasets:

JensCS
/

top-ml-conference-papers-2024

paper_title stringlengths 7 156	affiliation stringlengths 0 292	research_domain stringclasses 580 values	task_type stringlengths 0 75	core_problem stringlengths 0 445	key_innovation stringlengths 0 474	code_repository stringlengths 0 127	key_results stringlengths 0 1.03k	method_name stringlengths 0 140	method_summary stringlengths 0 858	agent_framework_name stringclasses 161 values	environment_or_tools stringclasses 132 values	planning_mechanism stringclasses 158 values	eval_benchmark stringclasses 159 values	authors listlengths 1 100	baseline_models listlengths 0 54	paper_type stringclasses 4 values	type_name stringclasses 4 values	extraction_quality float64 0.07 1	total_chars int64 0 800k	conference stringclasses 4 values	chunks listlengths 0 1.12k	hierarchy listlengths 0 372
Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models	Carnegie Mellon University	Natural Language Processing	Parameter-Efficient Fine-Tuning	Existing finetuning methods cannot simultaneously reduce the memory footprint of model weights, optimizer states, and intermediate activations, leading to high memory requirements.	A dual-stage framework combining 4-bit quantization of the LLM with a separate side network to eliminate backpropagation through the LLM and reduce all three memory contributors.		QST achieves the lowest memory footprint among all methods while attaining competent accuracy. On GLUE, QST reduces memory footprint by ~2× compared with QLoRA, LoRA, and Adapter, and reduces trainable parameters by ~10× and ~5× compared with QLoRA and other baselines. On MMLU, QST improves average accuracy by 0.1% whi...	Quantized Side-Tuning (QST)	QST quantizes the LLM to 4-bit to reduce memory of weights, then introduces a side network that takes downsampled hidden states from the LLM layers and combines them with previous layer outputs via a gated mechanism. The side network’s parameters (including low-rank adaptors and pooling-based downsample modules) are up...					[ "Zhengxin Zhang", "Dan Zhao", "Xupeng Miao", "Gabriele Oliaro", "Zhihao Zhang", "Qing Li", "Yong Jiang", "Zhihao Jia" ]	[ "QLoRA", "LoRA", "Adapter", "LST" ]	method_and_pipeline	Method and Pipeline	0.923	82,487	ACL	[ { "content": "Zhengxin Zhang $^{\\ddagger\\S}$ , Dan Zhao $^{b}$ , Xupeng Miao $^{\\ddagger}$ , Gabriele Oliaro $^{\\ddagger}$ Zhihao Zhang $^{\\ddagger}$ , Qing Li $^{b}$ , Yong Jiang $^{\\ddagger b}$ , Zhihao Jia $^{\\ddagger}$\n$^{\\ddagger}$ Carnegie Mellon University, $^{\\S}$ Tsinghua University,\n$^{b}$ ...	[ { "title": "Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models", "path": "2024.acl-long.1.pdf/Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models", "level": 1, "summary": "Zhengxin Zhang $^{\\ddagger\\S}$ , Dan Zhao $^{b}...
UNICODER : Scaling Code Large Language Model via Universal Code	State Key Laboratory of Complex & Critical Software Environment, Beihang University	Code Generation	Code Translation and Generation	"Standard chain-of-thought prompting uses natural language intermediate steps which are not well-ali(...TRUNCATED)	"Introduces UniCode, an intermediate representation using programming language conventions (assignme(...TRUNCATED)		"UNICODER significantly outperforms previous open-source baselines, especially Magicoder and WaveCod(...TRUNCATED)	UNICODER	"UNICODER is a code LLM fine-tuned on the UNICODER-INSTRUCT dataset. It uses Universal Code (UniCode(...TRUNCATED)					["Tao Sun","Linzheng Chai","Jian Yang","Yuwei Yin","Hongcheng Guo","Jiaheng Liu","Bing Wang","Liqun (...TRUNCATED)	["GPT-3.5","GPT-4","StarCoder","WizardCoder","OctoCoder","WaveCoder-SC","Code-Llama","Code-Llama-Ins(...TRUNCATED)	method_and_pipeline	Method and Pipeline	0.923	40,093	ACL	[{"content":"Tao Sun $^{1}$ , Linzheng Chai $^{1}$ , Jian Yang $^{1*†}$ , Yuwei Yin $^{2}$ , Hon(...TRUNCATED)	[{"title":"UNICODER : Scaling Code Large Language Model via Universal Code","path":"2024.acl-long.10(...TRUNCATED)
AoE: Angle-optimized Embeddings for Semantic Textual Similarity	Department of Computing, The Hong Kong Polytechnic University	Natural Language Processing	Semantic Textual Similarity	"The cosine similarity function has saturation zones that cause vanishing gradients, hindering the m(...TRUNCATED)	"AoE decomposes embeddings into real and imaginary components via complex division, optimizing angle(...TRUNCATED)	https://github.com/SeanLee97/Angle	"AoE consistently outperforms all baselines on STS benchmarks, achieving average score improvements (...TRUNCATED)	Angle-optimized Embedding (AoE)	"AoE decomposes text embeddings into real and imaginary components via complex division, computes th(...TRUNCATED)					[ "Xianming Li", "Jing Li" ]	["openai-ada-002","openai-text-embedding-3","InferSent-GloVe","USE","ConSERT","CoSENT","SBERT","SimC(...TRUNCATED)	method_and_pipeline	Method and Pipeline	1	70,203	ACL	[{"content":"Xianming Li $^{1}$ , Jing Li $^{1,2\\dagger}$\n$^{1}$ Department of Computing\n$^{2}$ R(...TRUNCATED)	[{"title":"AoE: Angle-optimized Embeddings for Semantic Textual Similarity \\*","path":"2024.acl-lon(...TRUNCATED)
"Does DETECTGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned Contrastiv(...TRUNCATED)	Faculty of Electronic and Information Engineering, Xi'an Jiaotong University	Natural Language Processing	Machine-generated text detection	"DetectGPT's random perturbation introduces noise, its logit regression relies on a threshold limiti(...TRUNCATED)	"Proposes PECOLA, a fine-tuned detector that bridges metric-based and fine-tuned methods via selecti(...TRUNCATED)		"PECOLA surpasses all baselines on all datasets. Compared to the best competitor, it achieves accura(...TRUNCATED)	PECOLA	"PECOLA operates in two stages: first, it uses YAKE-based token importance assessment to selectively(...TRUNCATED)					["Shengchao Liu","Xiaoming Liu","Yichen Wang","Zehua Cheng","Chengzhengxu Li","Zhaohan Zhang","Yu La(...TRUNCATED)	[ "RoBERTa", "GLTR", "CE+SCL", "CE+Margin", "IT:Clust", "CoCo", "DetectGPT", "Fast-Detect." ]	method_and_pipeline	Method and Pipeline	0.923	68,824	ACL	[{"content":"Shengchao Liu $^{1}$ , Xiaoming Liu $^{1,*}$ , Yichen Wang $^{1}$ , Zehua Cheng $^{1}$ (...TRUNCATED)	[{"title":"Does DETECTGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned (...TRUNCATED)
AFaCTA: Assisting the Annotation of Factual Claim Detection with Reliable LLM Annotators	ETH Zürich	Natural Language Processing	Factual Claim Detection	Inconsistency in definitions of factual claims and the high cost of manual annotation.	"AFaCTA framework that leverages LLMs with multi-step reasoning and consistency calibration to assis(...TRUNCATED)		"AFaCTA with GPT-4 achieves 98.49% accuracy and 0.833 Cohen's Kappa agreement with experts on perfec(...TRUNCATED)	AFaCTA	"AFaCTA assists in factual claim annotation via three prompting steps: (1) Direct Classification, (2(...TRUNCATED)					[ "Jingwei Ni", "Minjing Shi", "Dominik Stammbach", "Mrinmaya Sachan", "Elliott Ash", "Markus Leippold" ]	[ "Random", "Human experts" ]	method_and_pipeline	Method and Pipeline	0.923	110,860	ACL	[{"content":"Jingwei Ni $^{1}$ , Minjing Shi $^{1}$ , Dominik Stammbach $^{1}$ , Mrinmaya Sachan $^{(...TRUNCATED)	[{"title":"AFaCTA: Assisting the Annotation of Factual Claim Detection with Reliable LLM Annotators"(...TRUNCATED)
Towards Faithful and Robust LLM Specialists for Evidence-Based Question-Answering	University of Zürich	Natural Language Processing	Evidence-Based Question Answering	"LLMs struggle with source quality and answer attributability in evidence-based question answering, (...TRUNCATED)	"A synthetic data generation pipeline with automated quality filters for Evidence-Based QA, along wi(...TRUNCATED)		"Fine-tuning with high-quality synthetic data (SynSciQA++) achieves source quality scores of 80 on S(...TRUNCATED)		"The pipeline generates synthetic evidence-based QA data by: (1) generating diverse scientific topic(...TRUNCATED)					[ "Tobias Schimanski", "Jingwei Ni", "Mathias Kraus", "Elliott Ash", "Markus Leippold" ]	[ "GPT-3.5", "GPT-4", "Llama-2-13b-chat", "Zephyr-7b-beta" ]	method_and_pipeline	Method and Pipeline	0.846	99,550	ACL	[{"content":"Tobias Schimanski $^{1}$ , Jingwei Ni $^{1,2}$ , Mathias Kraus $^{3}$ , Elliott Ash $(...TRUNCATED)	[{"title":"Towards Faithful and Robust LLM Specialists for Evidence-Based Question-Answering","path"(...TRUNCATED)
"M-RAG: Reinforcing Large Language Model Performance through Retrieval-Augmented Generation with Mul(...TRUNCATED)	Huawei Technologies, Co., Ltd.	Natural Language Processing	Retrieval-Augmented Generation	"Existing RAG methods retrieve from a whole database, which limits focus on crucial memories and int(...TRUNCATED)	"Proposes M-RAG, a multiple partition paradigm for RAG with Multi-Agent Reinforcement Learning to op(...TRUNCATED)		"M-RAG achieves state-of-the-art results on text summarization, machine translation, and dialogue ge(...TRUNCATED)			M-RAG	"Database partitions, frozen language model (LLM), and CPT-Text embedded model for similarity assess(...TRUNCATED)	"Multi-agent reinforcement learning with Deep Q-Networks (DQN); Agent-S treats partition selection a(...TRUNCATED)	XSum, BigPatent, Es→En, En→Es, De→En, En→De, DailyDialog	[ "Zheng Wang", "Shu Xian Teo", "Jieer Ouyang", "Yongjun Xu", "Wei Shi" ]	[]	agent_system	Agent System	0.929	54,829	ACL	[{"content":"Zheng Wang $^{1}$ , Shu Xian Teo $^{1}$ , Jieer Ouyang $^{1}$ , Yongjun Xu $^{1}$ , Wei(...TRUNCATED)	[{"title":"M-RAG: Reinforcing Large Language Model Performance through Retrieval-Augmented Generatio(...TRUNCATED)
Open-Set Semi-Supervised Text Classification via Adversarial Disagreement Maximization	CCSE, School of Computer Science and Engineering, Beihang University, Beijing, China	Natural Language Processing	Open-Set Semi-Supervised Text Classification	"The main challenge in Open-Set Semi-Supervised Text Classification (OSTC) is the false positive inf(...TRUNCATED)	"The paper proposes an Adversarial Disagreement Maximization (ADM) model that directly maximizes mea(...TRUNCATED)		"ADM achieves the best performance in all settings of AGNews and most settings on Yahoo and DBPedia,(...TRUNCATED)	Adversarial Disagreement Maximization (ADM)	"The ADM model specifies cross-entropy loss and outlier detection confidence as two measurements. It(...TRUNCATED)					[ "Junfan Chen", "Richong Zhang", "Junchi Chen", "Chunming Hu" ]	["UDA+MSP","UDA+LSoftmax","UDA+DOC","UDA+LMCL","MixText+MSP","MixText+LSoftmax","MixText+DOC","MixTe(...TRUNCATED)	method_and_pipeline	Method and Pipeline	0.923	55,359	ACL	[{"content":"Junfan Chen $^{1,2}$ , Richong Zhang $^{1,3*}$ , Junchi Chen $^{1}$ , Chunming Hu $^{1,(...TRUNCATED)	[{"title":"Open-Set Semi-Supervised Text Classification via Adversarial Disagreement Maximization","(...TRUNCATED)
ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages	School of Computer Science, Fudan University	Natural Language Processing	Tool Learning Safety Assessment	"Existing research on LLM tool learning emphasizes leveraging tools to augment LLMs but neglects eme(...TRUNCATED)	"ToolSword is a comprehensive framework that delineates six safety scenarios for LLMs in tool learni(...TRUNCATED)	https://github.com/Junjie-Ye/ToolSword	"In the MQ scenario, most LLMs struggle with ASR over 60%, with Qwen-chat-72B achieving the lowest A(...TRUNCATED)			ToolSword				["Junjie Ye","Sixian Li","Guanyu Li","Caishuang Huang","Songyang Gao","Yilong Wu","Qi Zhang","Tao Gu(...TRUNCATED)	[]	agent_system	Agent System	0.786	84,217	ACL	[{"content":"Junjie Ye $^{1}$ , Sixian Li $^{1}$ , Guanyu Li $^{1}$ , Caishuang Huang $^{1}$ ,\nSong(...TRUNCATED)	[{"title":"ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three(...TRUNCATED)
How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Composition	Alibaba Group	Natural Language Processing	Supervised Fine-Tuning Data Composition	"Limited understanding of how data composition during supervised fine-tuning affects multiple abilit(...TRUNCATED)	"Systematic investigation of data composition effects on multiple LLM abilities and proposal of Dual(...TRUNCATED)		"DMT achieves competitive or superior performance on HumanEval and MT-Bench across LLaMA-7B, 13B, an(...TRUNCATED)	Dual-stage Mixed Fine-tuning (DMT)	"DMT first applies supervised fine-tuning on specialized datasets (code and math), then performs a s(...TRUNCATED)					["Guanting Dong","Hongyi Yuan","Keming Lu","Chengpeng Li","Mingfeng Xue","Dayiheng Liu","Wei Wang","(...TRUNCATED)	["General only","Math only","Code only","Multi-task learning","Sequential Training","Mixed Sequentia(...TRUNCATED)	method_and_pipeline	Method and Pipeline	0.923	104,321	ACL	[{"content":"Guanting Dong $^{}$ , Hongyi Yuan $^{}$ , Keming Lu, Chengpeng Li $^{*}$ , Mingfeng X(...TRUNCATED)	[{"title":"How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Compos(...TRUNCATED)

End of preview. Expand in Data Studio

Top ML Conference Papers 2024

A dataset of 9,817 papers from ACL 2024, CVPR 2024, ICLR 2024, and NeurIPS 2024, processed with KNOWHERE — a document understanding pipeline that prepares unstructured data for AI agents.

GitHub Repo: https://github.com/Ontos-AI/knowhere

Conference	Papers	Domain
ACL 2024	914	Natural Language Processing
CVPR 2024	2,692	Computer Vision
ICLR 2024	2,251	Machine Learning
NeurIPS 2024	3,960	Machine Learning

Construction

Built with KNOWHERE:

PDF Parsing — KNOWHERE parses each PDF into structured chunks (chunks.json): text blocks with section paths. The dataset retains each chunk's content and path.
Hierarchy Extraction — KNOWHERE outputs a doc_nav.json section tree per paper, with title, path, level, summary, chunk_count, and recursive children.
Agentic Extraction — An LLM pipeline classifies papers into 4 types and extracts fields from KNOWHERE's parsed chunks and section hierarchy.
Assembly — Extraction fields + KNOWHERE chunks + hierarchy merge into unified records.

Applications

Literature analysis and survey support — structured extraction fields (method, results, baselines) enable systematic literature review, cross-conference trend tracking, and automated survey generation
Benchmark for PDF parsing & scientific understanding — evaluate document parsing pipelines and information extraction models on real academic papers with ground-truth hierarchy and typed fields
Foundation for AI for Science experiments — structured paper representations serve as input for downstream tasks such as paper generation, research idea proposal, and citation-aware knowledge construction

Field Schema

All 24 extraction fields appear in every record. Fields not applicable to a paper's type are empty strings (or empty lists).

Common (all papers)

Field	Type	Description
`paper_title`	string	Verbatim title
`authors`	list[string]	Author names
`affiliation`	string	First author's institution
`research_domain`	string	AI/CS subfield
`task_type`	string	Specific technical task
`core_problem`	string	Unresolved limitation
`key_innovation`	string	Primary contribution
`code_repository`	string	Code URL (if stated)

model_architecture (1,282 papers)

Field	Description
`model_name`	Proposed model name
`architecture_type`	Design paradigm (e.g., Transformer)
`model_size_parameters`	Parameter count
`training_evaluation_dataset`	Datasets used
`key_results`	Quantitative results

method_and_pipeline (7,257 papers)

Field	Description
`method_name`	Proposed method name
`method_summary`	Core mechanism
`baseline_models`	list[string] — baselines compared
`key_results`	Advantage over baselines

theory_and_analysis (1,087 papers)

Field	Description
`analysis_target`	Object/phenomenon studied
`theoretical_tools`	Proof techniques
`key_findings`	Theorems or conclusions
`prior_work_comparison`	Comparison to prior results
`limitations`	Assumptions or scope

agent_system (191 papers)

Field	Description
`agent_framework_name`	Proposed framework name
`environment_or_tools`	Environments/tools used
`planning_mechanism`	Reasoning strategy
`eval_benchmark`	Evaluation benchmarks
`key_results`	Benchmark performance

Structural & Metadata

Field	Type	Description
`paper_type`	string	Classification key
`type_name`	string	Human-readable type name
`extraction_quality`	float	Non-null field fraction
`total_chars`	int	Paper character count
`conference`	string	Source conference
`chunks`	list[{content, path}]	KNOWHERE text chunks
`hierarchy`	list[{title, path, level, summary, chunk_count}]	KNOWHERE section tree

Formats

JSONL — Nested hierarchy tree with recursive children, human-readable.
Parquet — Columnar format, Zstd compression; hierarchy flattened to depth-first list.

Loading

from datasets import load_dataset

ds = load_dataset("JensCS/top-ml-conference-papers-2024", split="train")

print(ds[0]["paper_title"])
print(ds[0]["conference"], ds[0]["paper_type"])

# Type-specific fields
agent = ds.filter(lambda x: x["paper_type"] == "agent_system")
print(agent[0]["agent_framework_name"])

# KNOWHERE hierarchy
for sec in ds[0]["hierarchy"]:
    depth = "  " * (sec["level"] - 1)
    print(f"{depth}├─ {sec['title']}")

Pipeline

KNOWHERE → Parse → Structure → Build Memory → Agentic extraction → Dataset assembly.

License

CC BY 4.0. Original papers retain their respective copyrights.

Downloads last month: 46