Dataset Preview

Split (3)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed

Error code:   DatasetGenerationError
Exception:    ArrowInvalid
Message:      JSON parse error: Missing a closing quotation mark in string. in row 721
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 324, in _generate_tables
                  df = pandas_read_json(f)
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
                  return pd.read_json(path_or_buf, **kwargs)
                         ~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 815, in read_json
                  return json_reader.read()
                         ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1014, in read
                  obj = self._get_object_parser(self.data)
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1040, in _get_object_parser
                  obj = FrameParser(json, **kwargs).parse()
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1176, in parse
                  self._parse()
                  ~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1392, in _parse
                  ujson_loads(json, precise_float=self.precise_float), dtype=None
                  ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
              ValueError: Trailing data
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 327, in _generate_tables
                  raise e
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
                  pa_table = paj.read_json(
                      io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
                  )
                File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Missing a closing quotation mark in string. in row 721
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

arxiv_id string	arxiv_url string	title string	authors list	abstract string	categories list	date timestamp[s]	problem_definition string	intuition string	proposal_masked string	split string	generated_at timestamp[s]
2605.03871	https://arxiv.org/abs/2605.03871	EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics	[ "Li, Shuyue Stella", "Xin, Rui", "Xiao, Teng", "Wang, Yike", "Shao, Rulin", "Hao, Zoey", "Sclar, Melanie", "Oh, Sewoong", "Brahman, Faeze", "Koh, Pang Wei", "Tsvetkov, Yulia" ]	Language models encode substantial evaluative knowledge from pretraining, yet current post-training methods rely on external supervision (human annotations, proprietary models, or scalar reward models) to produce reward signals. Each imposes a ceiling. Human judgment cannot supervise capabilities beyond its own, propri...	[ "cs.AI" ]	2026-05-10T00:00:00	Formal setting: The problem involves generating evaluation criteria (rubrics) that enable a judge model to distinguish between preferred and dispreferred responses for a given question. The goal is to optimize these criteria so that they provide discriminative utility, allowing the judge to correctly recover preference...	Main idea: Co-evolving discriminative rubrics. Explanation: Structures implicit model knowledge into explicit, instance-specific natural language evaluation criteria.; Trains a rubric generator and response policy in alternation, using the policy's own historical outputs to create preference pairs.; Optimizes rubrics s...	# EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics ## Thesis A language model can continuously improve itself by co-evolving explicit evaluation rubrics with its response generation policy, eliminating the need for external human or proprietary model supervision. ## Topic Sub topic: Larg...	train	2026-07-14T07:10:18
2605.01386	https://arxiv.org/abs/2605.01386	MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents	[ "Van, Hung Pham", "Hieu, Nguyen Manh", "Tuan, Khang Pham Tran", "Hai, Nam Le", "Van, Linh Ngo", "Diep, Nguyen Thi Ngoc", "Le, Trung" ]	Large Language Models (LLMs) lack persistent memory for long-term personalized conversations. Existing graph-based memory systems suffer from information dilution, absent provenance tracking, and uniform retrieval that ignores query context. We introduce MemORAI (Memory Organization and Retrieval via Adaptive Graph Int...	[ "cs.CL" ]	2026-05-09T00:00:00	Formal setting: Given a multi-session dialogue history and a current query q, retrieve a ranked subset of turns and triplets to augment generation.. Inputs: Raw multi-turn conversation history, current user query q.. Outputs: Top-ranked conversation turns with supporting entity-relation triplets. Problem statement: LL...	Main idea: MemORAI. Explanation: Filters dialogue to keep only user-persona details while summarizing the rest to save space.; Builds a graph that explicitly links facts to the exact conversation turn they came from.; Ranks memories by dynamically weighting graph connections based on how well they match the current que...	# MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents ## Thesis A framework that filters user-relevant dialogue, structures it into a provenance-tracked graph, and retrieves context-aware memories using adaptive edge weighting. ## Topic Sub topic: Long-term memory ...	train	2026-07-14T07:10:17
2605.04217	https://arxiv.org/abs/2605.04217	Jordan-RoPE: Non-Semisimple Relative Positional Encoding via Complex Jordan Blocks	[ "Yaobo Zhang" ]	Relative positional encodings determine which functions of query-key lag can enter the primitive attention logit. RoPE supplies a rotary phase, while ALiBi supplies an additive distance bias. Motivated by group-theoretic views of linear translation-invariant positional encodings, we study a non-semisimple case in which...	[ "cs.LG", "cs.CL" ]	2026-05-05T00:00:00	Formal setting: Attention computes logits via inner products. Relative positional encoding applies a one-parameter matrix group G(d) to queries and keys so the score depends on causal lag d = i - j.. Inputs: Query vectors, key vectors, and positional indices.. Outputs: Transformed query and key vectors that yield relat...	Main idea: Jordan-RoPE. Explanation: Couples a complex rotary eigenvalue with a polynomial distance response inside a single defective Jordan block.; Generates a finite frequency-jet basis where higher-order blocks supply polynomially weighted phase modes.; Uses a contragredient query transformation to recover pure rel...	# Jordan-RoPE: Non-Semisimple Relative Positional Encoding via Complex Jordan Blocks ## Thesis The paper proposes a positional encoding that combines rotational and distance features into a single mathematical structure to help models better understand sequences where timing and rhythm interact. ## Topic Sub topic: P...	train	2026-07-14T07:10:16
2605.04215	https://arxiv.org/abs/2605.04215	Predict-then-Diffuse: Adaptive Response Length for Compute-Budgeted Inference in Diffusion LLMs	[ "Michael Rottoli", "Subhankar Roy", "Stefano Paraboschi" ]	Diffusion-based Large Language Models (D-LLMs) represent a promising frontier in generative AI, offering fully parallel token generation that can lead to significant throughput advantages and superior GPU utilization over traditional autoregressive paradigm. However, this parallelism is constrained by the requirement o...	[ "cs.LG", "cs.AI" ]	2026-05-05T00:00:00	Formal setting: Given prompt s, determine the optimal response length k to minimize computational cost (FLOPs), which scales with the fixed canvas size L (dominated by quadratic terms for long sequences).. Inputs: Raw text input prompt. Outputs: The optimal response length for the given prompt Problem statement: Diffu...	Main idea: Predict-then-Diffuse framework. Explanation: Predicts optimal response length from the input prompt before generation begins.; Uses a lightweight gradient boosting model for fast, accurate length estimation.; Adds a data-driven safety margin to prevent output truncation without costly re-runs.	# Predict-then-Diffuse: Adaptive Response Length for Compute-Budgeted Inference in Diffusion LLMs ## Thesis Predicting optimal output length before diffusion generation minimizes wasted computation on padding tokens while preventing truncation through a data-driven safety margin. ## Topic Sub topic: Efficient Inferen...	train	2026-07-14T07:10:16
2605.04263	https://arxiv.org/abs/2605.04263	Parallel Prefix Verification for Speculative Generation	[ "Yuncheng Yao", "Yuxuan Xia", "Shengjie Wang", "Danyang Zhuo" ]	We introduce PARSE (PArallel pRefix Speculative Engine), a speculative generation framework that accelerates large language model (LLM) inference by parallelizing prefix verification on a semantic level. Existing speculative decoding methods are fundamentally limited by token-level equivalence: the target model must ve...	[ "cs.AI" ]	2026-05-05T00:00:00	Formal setting: Given a prompt q, draft answer y_{1:T}, and target model M_b, find the maximal boundary t^* such that prefix y_{1:t^} is semantically correct.. Inputs: User prompt q; Full draft answer y_{1:T} from a small model. Outputs: Longest valid prefix length t^; Final completed answer Problem statement: Token...	Main idea: Parallel Prefix Verification (PPV). Explanation: Treats every partial draft output as a prefix to be evaluated simultaneously.; Uses a custom attention mask to isolate multiple verification queries in one forward pass.; Identifies the longest correct prefix instantly, enabling the target model to resume gene...	# Parallel Prefix Verification for Speculative Generation ## Thesis A speculative generation framework that accelerates LLM inference by verifying multiple draft prefixes simultaneously in a single forward pass, bypassing sequential bottlenecks. ## Topic Sub topic: LLM Inference Optimization; Speculative Decoding; Mo...	train	2026-07-14T07:10:18
2605.04295	https://arxiv.org/abs/2605.04295	LLMs Uncertainty Quantification via Adaptive Conformal Semantic Entropy	[ "Hamed Karimi", "Vaishali Meyappan", "Reza Samavi" ]	LLMs' overconfidence, particularly when hallucinating, poses a significant challenge for the deployment of the models in safety-critical settings and makes a reliable estimation of uncertainty necessary. Existing approaches for uncertainty quantification typically prioritize lexical or probabilistic measures; however, ...	[ "cs.LG", "cs.AI" ]	2026-05-05T00:00:00	Formal setting: Given a prompt x, the task is to estimate the model's uncertainty regarding the correctness of its generated responses. This involves quantifying semantic dispersion among multiple sampled responses to distinguish between genuine semantic ambiguity and lexical variation, ultimately producing a calibrate...	Main idea: Adaptive Conformal Semantic Entropy (ACSE). Explanation: Measures uncertainty by clustering semantically equivalent responses and computing entropy over soft cluster assignments; Adaptively inflates uncertainty scores using geometric features like cluster dispersion, centroid distance, and support size to pe...	# LLMs Uncertainty Quantification via Adaptive Conformal Semantic Entropy ## Thesis A framework that quantifies language model uncertainty by measuring semantic dispersion across multiple generated responses and adaptively adjusting scores based on cluster stability to enable reliable prompt abstention. ## Topic Sub ...	train	2026-07-14T07:10:14
2605.04308	https://arxiv.org/abs/2605.04308	Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping	[ "Kaustubh Pethkar", "Ziyang Xiong", "Zuofeng Shang", "Yingcong Li" ]	Continual incorporation of new knowledge is essential for the long-term evolution of large language models (LLMs). Existing approaches typically rely on parameter-update algorithms to mitigate catastrophic forgetting, yet they suffer from fundamental limitations: 1) forgetting is unavoidable as the amount of newly inje...	[ "cs.LG", "cs.AI" ]	2026-05-05T00:00:00	Formal setting: Autoregressive generation is modeled as a Markov chain over a token vocabulary V. Transition probabilities p(x_{t+1}\|x_t) define the model's memory. Expanding vocabulary V to V ∪ U requires learning new transition distributions for u ∈ U while strictly preserving p(v'\|v) for all v, v' ∈ V.. Inputs: A pr...	Main idea: Token-to-Dictionary Mapping via Embedding Tuning. Explanation: Models next-token prediction as transitions in a Markov chain where vocabulary items function as states.; Adds new knowledge by mapping new tokens to sparse combinations of existing tokens without altering original transition rules.; Updates only...	# Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping ## Thesis Modeling language generation as state transitions enables adding new vocabulary by updating only token representations, which prevents the loss of previously learned knowledge. ## Topic Sub topic: Continual Lea...	train	2026-07-14T07:10:15
2605.04313	https://arxiv.org/abs/2605.04313	NoisyCausal: A Benchmark for Evaluating Causal Reasoning Under Structured Noise	[ "Zhi Xu", "Yun Fu" ]	Causal reasoning in natural language requires identifying relevant variables, understanding their interactions, and reasoning about effects and interventions, often under noisy or ambiguous conditions. While large language models (LLMs) exhibit strong general reasoning abilities, they struggle to disentangle correlatio...	[ "cs.CL", "cs.AI" ]	2026-05-05T00:00:00	Formal setting: Given a directed acyclic causal graph G=(V,E) and a structural causal model defining conditional probabilities, the system receives a natural language scenario with perturbed observations or injected noise and must answer interventional, counterfactual, or attributional queries.. Inputs: Natural languag...	Main idea: NoisyCausal Benchmark and Graph-Guided Reasoning Framework. Explanation: Introduces a synthetic benchmark that injects six controlled types of structured noise into causal scenarios to test model robustness.; Decomposes reasoning into variable extraction, causal graph construction, and graph-guided inference...	# NoisyCausal: A Benchmark for Evaluating Causal Reasoning Under Structured Noise ## Thesis This paper introduces a benchmark with controlled noise and a modular framework that guides language models using explicit causal graphs to improve robust reasoning. ## Topic Sub topic: Causal Reasoning; Large Language Models;...	train	2026-07-14T07:10:13
2605.04357	https://arxiv.org/abs/2605.04357	Coral: Cost-Efficient Multi-LLM Serving over Heterogeneous Cloud GPUs	[ "Yixuan Mei", "Zikun Li", "Zixuan Chen", "Shiqi Pan", "Mengdi Wu", "Xupeng Miao", "Zhihao Jia", "K. V. Rashmi" ]	The usage of large language models (LLMs) has grown increasingly fragmented, with no single model dominating. Meanwhile, cloud providers offer a wide range of mid-tier and older-generation GPUs that enjoy better availability and deliver comparable performance per dollar to top-tier hardware. To efficiently harness thes...	[ "cs.DC", "cs.AI", "cs.CL", "cs.LG" ]	2026-05-05T00:00:00	Formal setting: Minimize total provisioning cost sum(p(g)) subject to per-model throughput demands sum(T(Psi(G_m^i))) >= T_m, optimizing over resource allocation Phi and model placement Psi across heterogeneous nodes G and models M.. Inputs: Set of models with latency SLOs, current per-model throughput demand, real-tim...	Main idea: Serving Templates with Lossless Two-Stage Decomposition. Explanation: Precomputes optimal model placements for specific GPU combinations offline, decoupling placement search from real-time allocation.; Caches these reusable templates to transform an intractable joint optimization into a fast online selection...	# Coral: Cost-Efficient Multi-LLM Serving over Heterogeneous Cloud GPUs ## Thesis Coral jointly optimizes cloud GPU allocation and model placement for multiple language models using precomputed templates to minimize serving costs while meeting latency targets. ## Topic Sub topic: LLM Serving Systems; Cloud Resource M...	train	2026-07-14T07:10:15
2605.04341	https://arxiv.org/abs/2605.04341	Budgeted LoRA: Distillation as Structured Compute Allocation for Efficient Inference	[ "Mohammed Sabry", "Anya Belz" ]	We study distillation for large language models under explicit compute constraints, with the goal of producing student models that are not only cheaper to train, but structurally efficient at inference time. While prior approaches to parameter-efficient distillation, such as LoRA, reduce adaptation cost, they leave the...	[ "cs.LG", "cs.AI", "cs.CL" ]	2026-05-05T00:00:00	Formal setting: Distillation from a teacher model to a student under a global dense-compute budget fraction F in [0,1], optimizing a KL-divergence and cross-entropy objective while enforcing a scheduled dense-cost constraint.. Inputs: Teacher logits, ground-truth labels, input sequences, global budget fraction F.. Outp...	Main idea: Budgeted LoRA. Explanation: Treats distillation as a structured compute allocation problem under a global budget constraint.; Dynamically adjusts module-level dense retention and low-rank capacity during training.; Applies post-training compression to selectively drop, approximate, or preserve dense pathways...	# Budgeted LoRA: Distillation as Structured Compute Allocation for Efficient Inference ## Thesis A distillation framework that allocates computation between dense and low-rank pathways under a global budget to produce structurally efficient student models. ## Topic Sub topic: Model Compression; Knowledge Distillation...	train	2026-07-14T07:10:16
2605.04363	https://arxiv.org/abs/2605.04363	Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment	[ "Seunghan Lee", "Jaehoon Lee", "Jun Seo", "Sungdong Yoo", "Minjae Kim", "Tae Yoon Lim", "Dongwan Kang", "Hwanil Choi", "SoonYoung Lee", "Wonbin Ahn" ]	TabPFN has recently gained attention as a foundation model for tabular datasets, achieving strong performance by leveraging in-context learning on synthetic data. However, we find that TabPFN is vulnerable to label shift, often overfitting to the majority class in the training dataset. To address this limitation, we pr...	[ "cs.LG", "cs.AI" ]	2026-05-06T00:00:00	Formal setting: Given training data D_train = {(x_i, y_i)} and test data D_test = {x_j}, label shift occurs when p_train(y) != p_test(y) while the conditional distribution p(x\|y) remains invariant.. Inputs: Test instance(s) x_j, Training dataset D_train, Training class prior p_train(y), Base model f. Outputs: Adjusted ...	Main idea: DistPFN / DistPFN-T. Explanation: Reweights model predictions at inference time using the ratio between the model's own predicted distribution and the known training prior.; Avoids iterative estimation of the test prior, operating in a single forward pass without architectural changes or parameter updates.; ...	# Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment ## Thesis Adjusts predicted probabilities of tabular foundation models at test time to correct for mismatched class distributions without retraining. ## Topic Sub topic: Tabular Machine Learning; In-Context Learning; Distributi...	train	2026-07-14T07:10:16
2605.04453	https://arxiv.org/abs/2605.04453	StableI2I: Spotting Unintended Changes in Image-to-Image Transition	[ "Jiayang Li", "Shuo Cao", "Xiaohui Li", "Zhizhen Zhang", "Kaiwen Zhu", "Yule Duan", "Yu Qiao", "Jian Zhang", "Yihao Liu" ]	In most real-world image-to-image (I2I) scenarios, existing evaluations primarily focus on instruction following and the perceptual quality or aesthetics of the generated images. However, they largely fail to assess whether the output image preserves the semantic correspondence and spatial structure of the input image....	[ "cs.CV", "cs.AI" ]	2026-05-06T00:00:00	Formal setting: Given input image I_in, output image I_out, and instruction x, evaluate transition fidelity across Semantic, Structure, and Low-level Appearance dimensions without ground truth references.. Inputs: Input image, output image, I2I control instruction. Outputs: Evaluation of transition fidelity indicating ...	Main idea: StableI2I. Explanation: Integrates semantic, structural, and low-level appearance dimensions for comprehensive fidelity assessment.; Uses an error-amplification data pipeline to generate diverse training samples with subtle consistency violations.; Employs a multi-stage training scheme combining supervised f...	# StableI2I: Spotting Unintended Changes in Image-to-Image Transition ## Thesis A unified evaluation framework that measures content fidelity and pre-post consistency across image-to-image tasks without requiring reference images. ## Topic Sub topic: Image-to-Image Generation Evaluation; Multimodal Large Language Mod...	train	2026-07-14T07:10:15
2605.04449	https://arxiv.org/abs/2605.04449	GEM: Graph-Enhanced Mixture-of-Experts with ReAct Agents for Dialogue State Tracking	[ "Ziqi Zhu", "Adithya Suresh", "Tomal Deb", "Iman Abbasnejad" ]	Dialogue State Tracking (DST) requires precise extraction of structured information from multi-domain conversations, a task where Large Language Models (LLMs) struggle despite their impressive general capabilities. We present GEM (Graph-Enhanced Mixture-of-Experts), a novel framework that combines language models and g...	[ "cs.CL", "cs.AI" ]	2026-05-06T00:00:00	Formal setting: The task is to predict intents, slots, and slot-values across turns in a multi-turn dialogue.. Inputs: Dialogue history, current user utterance, and system response.. Outputs: Predicted intents, identified slots, and extracted slot-value pairs. Problem statement: Large language models struggle to accur...	Main idea: Graph-Enhanced Mixture-of-Experts (GEM) with ReAct Agents. Explanation: Dynamically routes each dialogue turn to either a graph neural network or a sequence model based on domain-specific structural needs.; Integrates a ReAct agent that performs explicit stepwise reasoning over retrieved few-shot examples to...	# GEM: Graph-Enhanced Mixture-of-Experts with ReAct Agents for Dialogue State Tracking ## Thesis A hybrid framework dynamically routes between graph and sequence models for dialogue understanding and uses agentic reasoning to accurately extract structured conversation states. ## Topic Sub topic: Dialogue State Tracki...	train	2026-07-14T07:09:56
2605.04468	https://arxiv.org/abs/2605.04468	Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control	[ "Xinyu Wang", "Changzhi Sun", "Yuanbin Wu", "Xiaoling Wang" ]	Post-training large language models (LLMs) often suffers from catastrophic forgetting, where improvements on a target objective degrade previously acquired capabilities. Recent evidence suggests that this phenomenon is primarily driven by excessive distributional drift during optimization. Motivated by this perspective...	[ "cs.LG", "cs.AI", "cs.CL" ]	2026-05-06T00:00:00	Formal setting: Given a base model p_base, a fixed SFT reference p_sft, and target dataset D, adapt p_theta to D while minimizing distributional drift relative to p_base without access to pretraining data.. Inputs: Labeled target dataset D = {(x_i, y_i)} Base model p_base Fixed SFT reference model p_sft. Outputs: Adapt...	Main idea: Anchored Learning. Explanation: Constructs a dynamic target by interpolating between the current model state and a fixed reference model.; Transforms global fine-tuning into a sequence of local, trust-region-like updates in distribution space.; Explicitly bounds per-iteration KL divergence to prevent excessi...	# Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control ## Thesis A dynamically interpolated moving anchor stabilizes LLM fine-tuning by constraining distributional updates to small, controlled steps. ## Topic Sub topic: Large Language Model Post-Training; Continual Learning; Supervised Fine-Tuni...	train	2026-07-14T07:10:17
2605.04458	https://arxiv.org/abs/2605.04458	DoGMaTiQ: Automated Generation of Question-and-Answer Nuggets for Report Evaluation	[ "Bryan Li", "William Walden", "Yu Hou", "Gabrielle Kaili-May Liu", "Dawn Lawrie", "Jame Mayfield", "Eugene Yang", "Chris Callison-Burch", "Laura Dietz" ]	Evaluation of long-form, citation-backed reports has lately received significant attention due to the wide-scale adoption of retrieval-augmented generation (RAG) systems. Core to many evaluation frameworks is the use of atomic facts, or nuggets, to assess a report's coverage of query-relevant information attested in th...	[ "cs.CL", "cs.IR" ]	2026-05-06T00:00:00	Formal setting: Given a user query Q and a set of retrieved multilingual documents D, generate a set S of question-answer pairs that capture the critical information needs for report evaluation.. Inputs: User query, retrieved document collection (potentially multilingual). Outputs: A set of question-answer nuggets capt...	Main idea: DoGMaTiQ pipeline. Explanation: Represents nuggets as question-answer pairs to separate user information needs from multilingual supporting evidence.; Clusters paraphrased questions and validates answers to consolidate redundant facts and ensure factual precision.; Ranks and selects final nuggets using a lea...	# DoGMaTiQ: Automated Generation of Question-and-Answer Nuggets for Report Evaluation ## Thesis An automated pipeline generates question-answer nuggets to enable scalable, cross-lingual evaluation of long-form AI reports without manual curation. ## Topic Sub topic: Information Retrieval; Natural Language Processing; ...	train	2026-07-14T07:10:14

End of preview.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Svalbard Idea Vault

One record per arXiv paper: the paper's identity plus structured, machine-generated abstractions of the research idea — the problem it defines, the intuition behind its approach, and a results-masked statement of its proposal.

2,669,631 records. Built from 2,722,845 candidate papers; 52,803 were dropped by the eligibility filter and 411 withheld by the verbatim gate described below.

Fields

field	description
`arxiv_id`, `arxiv_url`	canonical arXiv identifier and abs page
`title`, `authors`, `abstract`	arXiv metadata
`categories`, `date`	arXiv primary/cross-list categories, submission date
`problem_definition`	generated: formal setting (inputs/outputs) + what fails today
`intuition`	generated: the core idea at analogy level, no notation
`proposal_masked`	generated: the full proposal with all measured results removed

Licensing

The papers themselves are not redistributed here, and copyright in them remains with their authors. This dataset contains two separable things:

arXiv metadata (title, abstract, authors, arxiv_url, categories, date) — redistributed under arXiv's CC0 1.0 metadata terms.
Our generated annotations (problem_definition, intuition, proposal_masked) — released under CC BY 4.0.

How the annotations were produced, and their limits

The annotation fields are generated by a large language model from the paper's full text. They are abstractive paraphrases, not excerpts: every record was scored for word 8-gram overlap against its source and records above a 0.15 threshold were withheld.

Measured on a random sample of 10,000 records scored against the paper's full text: median 0.0051, p90 0.0292, p99 0.0904, max 0.2110. A copied sentence scores near 1.0; a genuine paraphrase near 0.

Papers excluded from the source corpus:

44,407 because no structured summary had been generated for them
8,396 because they did not resolve to an arXiv identifier

These fields may be inaccurate, incomplete, or misleading, and they do not represent the views or claims of the original authors. Do not cite them as statements of what a paper says; go to the paper.

Takedown

If you are an author and want your paper's record removed, open an issue on the dataset repository or contact the maintainers; we will remove it.

Downloads last month: 21

Total file size:

7.19 GB