Datasets:

ScholarCatalyst
/

ScholarCatalyst

query_id stringlengths 25 25	positive_docs listlengths 1 10	hard_negatives listlengths 4 13
q_2507.06542_core_query_0	[ { "id": "arxiv_2310.14423", "author_rationale": "Gu et al. introduced the Quadratic Synchronization Rule for data-parallel Local SGD, dynamically setting the number of local steps as H∝η−2, and demonstrated that nonuniform synchronization can improve both communication efficiency and test accuracy. This res...	[ "arxiv_2002.09692", "arxiv_1803.06443", "arxiv_2206.03093", "arxiv_2202.02580", "arxiv_1910.12308", "arxiv_1602.05629", "arxiv_2211.00533", "arxiv_2008.09246", "arxiv_2204.04452", "arxiv_2412.16968" ]
q_2509.24282_core_query_0	[ { "id": "arxiv_2505.19628", "author_rationale": "HomeBench evaluated smart-home LLMs on valid and invalid instructions across single- and multi-device tasks and found that models struggled more on the harder cases. Its evaluation compared model outputs with predefined device operations for each instruction....	[ "arxiv_2402.05733", "arxiv_2206.10498", "arxiv_2412.13178", "arxiv_2409.20222", "arxiv_2312.03009", "arxiv_2407.18550", "arxiv_2303.08268", "arxiv_2012.05672", "arxiv_2106.00808", "arxiv_2410.14803" ]
q_2509.01082_core_query_0	[ { "id": "arxiv_2402.17879", "author_rationale": "This paper showed that language models can be used for statistical model discovery by repeatedly proposing probabilistic programs and critiquing them in an iterative loop. The main idea we took from this work was that model generation does not need to be a on...	[ "arxiv_2309.17446", "arxiv_2204.11454", "arxiv_2111.08267", "s2_671d9f59d65ff10f63b94c48fd600542a8c18c08", "arxiv_2308.02944", "arxiv_2306.12672", "oa_W2048027175", "arxiv_2205.09735", "arxiv_2410.14865", "arxiv_2304.03287" ]
q_2603.08104_core_query_0	[ { "id": "oa_W2124890704", "author_rationale": "The paper provides a systematic survey of information-hiding techniques. It clarifies the fundamental goal of steganography: concealing the existence of the hidden message by embedding it inside innocuous cover information. This inspired me to introduce stegano...	[ "arxiv_2310.01581", "arxiv_2402.14020", "oa_W4402683797", "arxiv_2405.14023", "arxiv_2407.00869", "arxiv_2402.12343", "arxiv_2409.18169", "arxiv_2310.10077", "arxiv_2407.10264", "arxiv_2410.18469" ]
q_2502.01481_core_query_0	[ { "id": "arxiv_2309.16039", "author_rationale": "This paper basically trains extended context length (in its time) Llama models and measures input context scaling laws. So this is where the experimental scaling law (i.e. longer context could improve performance) is well shown. This paper provides experiment...	[ "arxiv_2207.04901", "arxiv_2310.04680", "arxiv_2308.10882", "arxiv_2503.04725", "arxiv_2411.03538", "arxiv_2310.04418", "arxiv_2212.10947", "arxiv_2309.10400", "arxiv_2405.18009", "arxiv_2402.10685" ]
q_2503.07982_core_query_0	[ { "id": "arxiv_2406.02842", "author_rationale": "DiffCut produces zero-shot semantic segmentation by running recursive Normalized Cut (NCut) over features taken from a frozen diffusion model at one fixed denoising timestep, without training or annotation. The diffusion prior itself, not a mask-trained head,...	[ "arxiv_2308.12469", "arxiv_2212.02773", "arxiv_2305.12410", "arxiv_2411.10495", "arxiv_2303.09813", "arxiv_2412.14580", "arxiv_2303.04803", "arxiv_2002.07705", "arxiv_2401.09709", "arxiv_2210.06366", "arxiv_2308.06739" ]
q_2509.25044_core_query_0	[ { "id": "arxiv_2205.14135", "author_rationale": "The paper used fused kernels that fuse operations in kernel code whose intermediates are not required. My idea extended this to cross correlation, mutual information, compositive updates, etc. I expected this to work because writing the kernels was straightfo...	[ "arxiv_2010.01465", "oa_W2133287637", "arxiv_2303.10211", "arxiv_2101.09639", "s2_187a4690bc6433c045425c01cb73aef8ceb150c6", "arxiv_2006.02338", "arxiv_2409.01068", "oa_W2001666669", "arxiv_2111.15509", "oa_W2344147939" ]
q_2505.12565_core_query_0	[ { "id": "s2_f86af9ec71e99613c9af500eb23dbc08fe3a6809", "author_rationale": "How to perform a chemical reaction used in automated synthesis on robots. We expected this to work because we did it." }, { "id": "s2_b44fb6bbe799780734c8c0362951d8f2a1210cf3", "author_rationale": "How to perform a chemi...	[ "arxiv_2110.06389", "arxiv_2301.12586", "arxiv_2410.02718", "arxiv_2106.03394", "arxiv_2202.00658", "oa_W1965815220", "arxiv_2012.11522", "arxiv_2004.12485", "arxiv_2410.03494", "arxiv_2303.06965", "arxiv_2410.20182" ]
q_2603.13300_core_query_0	[ { "id": "arxiv_2504.08661", "author_rationale": "This paper inspired my work by introducing Control Barrier Functions as a principled way to enforce safety constraints during flow-based generation. Motivated by this idea, we explored whether safe-region constraints and repulsive guidance could be extended f...	[ "arxiv_2211.05105", "arxiv_2307.16463", "arxiv_2411.13982", "arxiv_2412.16039", "arxiv_2410.14398", "arxiv_2412.01339", "arxiv_2412.03876", "arxiv_2502.08006", "arxiv_2410.12761", "arxiv_2404.07724" ]
q_2505.22785_core_query_0	[ { "id": "arxiv_1211.4246", "author_rationale": "Connects denoising autoencoder objectives with contractive objectives showing that locally, a denoising autoencoder and a contractive autoencoder both approximate the score of the data distribution. Validates the theory with synthetic examples in 2d. This pape...	[ "arxiv_2201.11902", "arxiv_2310.02897", "arxiv_2009.09525", "arxiv_2005.10516", "arxiv_2005.04321", "arxiv_2102.08373", "arxiv_2108.13910", "arxiv_2310.02250", "arxiv_2001.06116", "arxiv_2304.03376" ]
q_2510.19811_core_query_0	[ { "id": "arxiv_2309.14316", "author_rationale": "This paper trained many smaller language models with carefully crafted, synthetic training data. The training data was designed to answer questions about knowledge storage in LLMs. This is an interesting paper but the language models were not trained on natur...	[ "arxiv_2402.09363", "arxiv_2407.19262", "arxiv_2407.17817", "arxiv_2212.08619", "arxiv_2407.08707", "arxiv_2410.02159", "arxiv_2507.09937", "arxiv_2112.12938", "arxiv_2505.24832", "oa_W4411549888" ]
q_2510.16196_core_query_0	[ { "id": "s2_139e90776bde3e2657497d5bb759f2514da1b53f", "author_rationale": "This provides the fMRI dataset that can be used to train our model. I use the data in my paper." }, { "id": "arxiv_2112.09305", "author_rationale": "This provides a method to compute the similarity of two spaces. The new...	[ "arxiv_2408.06788", "arxiv_2303.05334", "arxiv_2409.05279", "arxiv_2406.02659", "arxiv_2202.12692", "arxiv_2310.02265", "arxiv_2401.11708", "arxiv_2305.11560", "arxiv_2410.20981", "arxiv_2308.07428", "arxiv_2212.02409" ]
q_2511.01266_core_query_0	[ { "id": "arxiv_2412.02700", "author_rationale": "Motion Prompting showed that trajectories can be used as a flexible control signal for objects, cameras, and general scene motion. We took the idea of trajectory-based control, but wanted users to provide it continuously during generation. Its different appli...	[ "arxiv_2403.12706", "arxiv_2412.09828", "arxiv_2503.19325", "arxiv_2406.10981", "arxiv_2406.06890", "arxiv_2508.13009", "arxiv_2411.16375", "arxiv_2410.08151", "oa_W4402904066", "arxiv_2411.18447" ]
q_2507.01037_core_query_0	[ { "id": "arxiv_2107.04139", "author_rationale": "Its contribution: L2D showed that subproblem selection in iterative VRP solving can be learned: a neural network identifies which spatially proximate subroutes are most promising to re-optimize, and delegates only that subproblem to a backbone solver, substan...	[ "oa_W2096305430", "arxiv_2212.08101", "arxiv_2401.06979", "arxiv_2105.02730", "arxiv_2205.00772", "arxiv_2012.13269", "arxiv_2102.11756", "arxiv_2310.04140", "arxiv_2310.15543", "arxiv_2501.03715", "arxiv_2311.13569" ]
q_2603.00873_core_query_0	[ { "id": "arxiv_2409.12959", "author_rationale": "MMSearch introduced a benchmark and search pipeline for evaluating large multimodal models as search engines, extending AI search beyond text-only queries and evidence to jointly support textual and visual information. It evaluated both individual search comp...	[ "arxiv_2410.03117", "arxiv_2408.13860", "arxiv_2405.03272", "arxiv_2408.12763", "arxiv_2502.17297", "arxiv_2104.06039", "arxiv_2411.16740", "arxiv_2309.08922", "arxiv_2402.19467", "arxiv_2405.18358" ]
q_2509.15221_core_query_0	[ { "id": "arxiv_2406.08451", "author_rationale": "GUIOdyssey introduced a comprehensive cross-app mobile navigation dataset containing 8,334 episodes across 212 applications and 1,357 app combinations, with each step annotated for screen comprehension, historical context, and decision rationale. Its central ...	[ "arxiv_2410.05243", "arxiv_2504.12679", "arxiv_2505.13909", "arxiv_2505.13227", "arxiv_2501.12326", "arxiv_2408.06327", "arxiv_2508.14040", "arxiv_2508.09123", "arxiv_2405.15341", "arxiv_2406.11317", "arxiv_2504.10127" ]
q_2511.05924_core_query_0	[ { "id": "arxiv_2410.16295", "author_rationale": "They do something very similar to my idea: permutation-equivariance of transformers used to go from particle-level information to the ensemble-level information (mean-field velocity in their case, density and score in my case)." }, { "id": "arxiv_2206...	[ "arxiv_2310.01762", "arxiv_2306.01993", "arxiv_2403.01189", "arxiv_2410.06986", "arxiv_2310.14458", "arxiv_2401.04856", "arxiv_2210.00726", "arxiv_2401.15604", "arxiv_2001.02728", "arxiv_2111.11010" ]
q_2603.05425_core_query_0	[ { "id": "arxiv_2511.16624", "author_rationale": "This paper proposes a feed-forward image-to-3D generation model that is able to complete occluded parts of objects. This open-sourced model serves as the backbone of our method, providing a pretrained latent space in which we are able to perform guidance fusi...	[ "arxiv_2411.15236", "arxiv_2410.11473", "arxiv_2307.13908", "arxiv_2312.06655", "arxiv_2404.11824", "arxiv_2312.17611", "arxiv_2309.03599", "arxiv_2412.12906", "arxiv_2308.06027", "arxiv_2410.21299" ]
q_2602.06291_core_query_0	[ { "id": "arxiv_2509.22099", "author_rationale": "This paper presents a systematic review of models that can judge problems they cannot solve. So it shows that judging capability > solving capability, and such a gap exists." }, { "id": "arxiv_2505.11855", "author_rationale": "One of my earlier pa...	[ "arxiv_2205.12910", "arxiv_2410.07523", "arxiv_2510.13888", "arxiv_2509.26076", "arxiv_2505.12575", "arxiv_2308.07921", "arxiv_2408.03492", "arxiv_2409.04168", "arxiv_2401.06961", "arxiv_2410.01920" ]
q_2511.19900_core_query_0	[ { "id": "arxiv_2504.11536", "author_rationale": "ReTool showed that reinforcement learning can teach models to strategically integrate external tools into their reasoning process, substantially improving both reasoning accuracy and training efficiency. Inspired by this, we extended learnable tool-integrated...	[ "arxiv_2503.06749", "arxiv_2411.17760", "arxiv_2509.00676", "arxiv_2509.02479", "arxiv_2404.06510", "arxiv_2511.15661", "arxiv_2510.24285", "arxiv_2412.17451", "arxiv_2404.04627", "arxiv_2501.01457", "arxiv_2511.16672" ]
q_2503.06623_core_query_0	[ { "id": "arxiv_2405.03376", "author_rationale": "This work uses a VAE architecture and entropy coding to substantially compress ERA5 reanalysis data. It inspired me to explore weather data compression from a representation perspective, aiming to reduce data size and consequently lower the costs of downloadi...	[ "arxiv_2111.03476", "arxiv_2304.08754", "arxiv_2406.19615", "arxiv_2410.09109", "arxiv_2012.09841", "arxiv_2210.12538", "arxiv_2407.11696", "arxiv_2205.04601", "arxiv_2301.10343", "arxiv_2406.10108", "arxiv_2312.00290" ]
q_2605.15622_core_query_0	[ { "id": "arxiv_2310.02025", "author_rationale": "DeepZero directly confronted ZO's dimensionality bottleneck in deep networks by exploiting sparsity and coordinate-wise structure, showing that carefully selecting perturbation dimensions can substantially improve ZO training and bring it closer to first-orde...	[ "arxiv_2511.07971", "arxiv_2502.01014", "arxiv_2509.15552", "arxiv_2506.09034", "arxiv_2406.18060", "arxiv_2402.01621", "arxiv_2410.07698", "arxiv_2405.16805", "arxiv_2506.14460" ]
q_2510.06824_core_query_0	[ { "id": "arxiv_2502.09741", "author_rationale": "This paper proposes converting numeric values into Fourier-like features and using these features as both inputs and prediction targets. Fourier features can represent a wide range of numbers continuously while maintaining a constant L2 norm for the encoding....	[ "arxiv_2401.03735", "arxiv_2410.11781", "arxiv_2411.16260", "arxiv_2310.06204", "arxiv_2205.06733", "arxiv_2411.03766", "s2_92a04a16a99eeec7d6bfc644e07c98589fe1cdf6", "arxiv_2308.01154", "arxiv_2102.13019", "arxiv_2411.02083", "arxiv_2006.01681" ]
q_2607.04293_core_query_0	[ { "id": "arxiv_2505.17968", "author_rationale": "This work framed scientific discovery as reverse-engineering a black-box system and directly compared learning from passive observations with learning through agent-chosen interventions. The idea we drew from it was that an AI scientist should be evaluated by...	[ "arxiv_2404.05545", "arxiv_2408.06849", "arxiv_2603.14575", "arxiv_2410.23884", "arxiv_2410.19923", "arxiv_2305.16183", "oa_W7163157186", "arxiv_2305.00050", "arxiv_2412.13667", "arxiv_2402.03941" ]
q_2512.20757_core_query_0	[ { "id": "arxiv_2310.08754", "author_rationale": "In this paper the authors trained their own tokenizers and models to study the effects of certain dimensions of the tokenizer design space. The main differences are as follows:\n1) They trained the tokenizers from scratch, while we study the applications that...	[ "arxiv_2402.18376", "arxiv_2404.13292", "arxiv_2305.17179", "arxiv_2109.02550", "arxiv_2502.12560", "arxiv_2306.09572", "arxiv_2304.10158", "arxiv_2204.04058", "arxiv_2403.06265", "arxiv_2404.18071", "arxiv_2204.08832" ]
q_2602.16682_core_query_0	[ { "id": "arxiv_2410.23266", "author_rationale": "TOMATO is a benchmark paper that aims to evaluate the state tracking capability in multimodal foundation models, which is the building block of SAW-Bench. We treat the state tracking capability as the major limitation in the current foundation models, and SAW...	[ "arxiv_2406.07544", "arxiv_2506.03135", "arxiv_2301.01949", "arxiv_2411.06048", "arxiv_2505.23764", "arxiv_2409.13929", "arxiv_2107.06011", "arxiv_2104.00387", "arxiv_2512.10863", "arxiv_2505.21500" ]
q_2605.17504_core_query_0	[ { "id": "arxiv_2510.14741", "author_rationale": "DEXTER introduced a data-free framework that uses diffusion models and textual reasoning to generate realistic images that strongly activate target visual concepts. Its results showed that a pretrained diffusion prior can preserve perceptual plausibility whil...	[ "arxiv_2004.13166", "arxiv_2307.05471", "arxiv_2401.06122", "arxiv_2209.14687", "arxiv_2407.01331", "arxiv_2306.06805", "arxiv_2309.17144", "arxiv_2106.12447", "arxiv_2404.14349", "arxiv_2306.04719", "arxiv_2409.01610" ]
q_2602.05999_core_query_0	[ { "id": "arxiv_2504.01871", "author_rationale": "This paper also trained an RNN as an RL policy. It showed that the RNN uses its iterations to predict planning related computations. It showed some evidence that the RL policy uses its iteration to form plans that causally affect how it acts. While this paper...	[ "arxiv_2004.12770", "arxiv_2005.07404", "arxiv_2301.13442", "arxiv_2102.07456", "arxiv_1901.03559", "arxiv_2201.12403", "arxiv_2005.02305", "arxiv_2210.06766", "arxiv_1602.02867", "arxiv_2406.08404", "arxiv_2311.06964" ]
q_2509.22352_core_query_0	[ { "id": "arxiv_2302.12749", "author_rationale": "SurvivalGAN demonstrated that synthetic data generation for survival analysis benefits from explicitly accounting for time-to-event outcomes and censoring rather than treating them as ordinary tabular variables. Its contribution was to introduce a survival-sp...	[ "arxiv_2402.12331", "arxiv_2410.16872", "arxiv_2202.03636", "arxiv_2410.16811", "arxiv_2104.10680", "arxiv_2106.05763", "arxiv_2102.09249", "arxiv_2302.14679", "arxiv_2504.16506", "arxiv_2410.01933" ]
q_2504.08837_core_query_0	[ { "id": "arxiv_2501.12948", "author_rationale": "Selected paper demonstrates how reinforcement learning with verifiable rewards incentivizes advanced reasoning ability in LLMs. On top of this, we fix the problem of vanishing advantages and handle the self-reflections in VLMs. I expected the idea to work bec...	[ "arxiv_2503.06749", "arxiv_1511.05952", "arxiv_2503.07365", "arxiv_2305.13267", "arxiv_2412.17451", "arxiv_2412.05479", "arxiv_2401.08025", "arxiv_2503.12937", "arxiv_2407.11422", "arxiv_2503.07536", "arxiv_2411.11930" ]
q_2605.14054_core_query_0	[ { "id": "arxiv_2505.15966", "author_rationale": "The first work for incentivizing tool-using behaviors in VLMs for visual reasoning. On top of this work, we realize that we might not need explicit tool-use to externalize visual perception behaviors in VLMs. We expect to use the model's internal visual perce...	[ "arxiv_2406.14544", "arxiv_2006.11524", "arxiv_2411.10440", "s2_4ef483f819e11873822416042a4b6dc4652e010c", "arxiv_2410.22995", "arxiv_2402.03300", "arxiv_2412.05479", "arxiv_2410.14138", "arxiv_2305.14985", "arxiv_2409.00106", "arxiv_2411.12591", "arxiv_2309.04461" ]
q_2605.24041_core_query_0	[ { "id": "arxiv_1909.01377", "author_rationale": "The deep equilibrium model explicitly casts network inference as the solution of a fixed-point system." }, { "id": "arxiv_1806.07366", "author_rationale": "The Neural ODE explicitly casts network inference as the solution of a continuous-time dyna...	[ "arxiv_2310.00120", "arxiv_2210.10890", "arxiv_2402.16845", "arxiv_2104.05512", "arxiv_2401.03492", "arxiv_1301.4498", "arxiv_2409.08477", "oa_W4408378430", "arxiv_2405.17211", "arxiv_2406.01857" ]
q_2602.01083_core_query_0	[ { "id": "arxiv_2301.12780", "author_rationale": "Navon et al. introduced Deep Weight-Space Networks (DWSNets), architectures designed to process neural-network parameters while respecting the neuron-permutation symmetries of deep weight spaces. Their work showed that explicitly incorporating these symmetrie...	[ "arxiv_2406.08966", "arxiv_2410.04213", "arxiv_2212.08648", "arxiv_2402.05232", "arxiv_2002.08599", "arxiv_1910.02421", "arxiv_2309.13736", "arxiv_2410.06665", "arxiv_2401.09235" ]
q_2606.00944_core_query_0	[ { "id": "arxiv_2410.20625", "author_rationale": "LoRA-RITE showed that equivalent LoRA factorizations should lead to the same effective update and introduced transformation invariance for LoRA optimization. This directly inspired us to ask whether the same invariance principle should also apply to the rando...	[ "arxiv_2405.06368", "arxiv_2006.15429", "arxiv_2409.17538", "arxiv_2207.00160", "arxiv_2302.01463", "arxiv_2310.15526", "arxiv_2202.08312", "arxiv_2106.09352", "arxiv_2207.02699", "arxiv_2412.17053" ]
q_2603.06957_core_query_0	[ { "id": "arxiv_2510.15020", "author_rationale": "This paper defines a notion of coverage for a base model which can guarantee the success of Best-of-N decoding. They then show that pre-training with empirical risk minimization or variants of stochastic gradient descent can efficiently achieve good coverage....	[ "arxiv_2510.11495", "arxiv_2507.14843", "arxiv_2506.14245", "arxiv_2410.03717", "arxiv_2106.04895", "arxiv_2402.10342", "arxiv_2106.15503", "arxiv_2406.01462", "arxiv_2510.14901" ]
q_2606.02133_core_query_0	[ { "id": "arxiv_2505.05755", "author_rationale": "This paper introduced a model that generates variable-length sequences by jointly selecting an insertion position and the token to insert. We took from it the view that left-to-right autoregression is only a special case of insertion-based generation, and ask...	[ "arxiv_1906.01604", "arxiv_2305.15562", "arxiv_2112.09097", "arxiv_2112.06295", "arxiv_2001.05540", "arxiv_2206.01370", "arxiv_1905.11006", "arxiv_2102.11008", "s2_753f2ba88fb4e50dd91c30db5264497f60c3f0e0", "arxiv_2106.03257" ]
q_2602.20433_core_query_0	[ { "id": "arxiv_2304.01373", "author_rationale": "Describe what this selected paper's own contribution was: Pythia introduced a suite of language models designed for controlled analysis of training dynamics and scaling, with models released at many intermediate checkpoints. This enabled researchers to study ...	[ "arxiv_2406.12159", "arxiv_2405.17767", "arxiv_2012.13255", "arxiv_2206.06072", "arxiv_2408.15417", "arxiv_2406.01468", "arxiv_2310.06977", "arxiv_2406.02449", "arxiv_2405.15471" ]
q_2510.00526_core_query_0	[ { "id": "arxiv_2508.05629", "author_rationale": "This paper analyzes standard SFT from a reinforcement-learning perspective and argues that its token-level gradients induce a problematic implicit reward structure. It proposes Dynamic Fine-Tuning, which rectifies this structure by rescaling each token’s obje...	[ "arxiv_2401.13586", "arxiv_2409.13641", "arxiv_2409.01369", "arxiv_2411.14797", "arxiv_2406.02756", "arxiv_2408.10642", "arxiv_2505.16984", "arxiv_2306.00398", "arxiv_2403.12017", "arxiv_2211.05826" ]
q_2601.22371_core_query_0	[ { "id": "oa_W2187471809", "author_rationale": "The main contribution of this paper is their theoretical foundation and approximation that enables accurate inference in heteroscedastic GPs, which inspires us to look into the heteroscedasticity in our data and how to learn better with heteroscedasticity. We t...	[ "arxiv_2206.04872", "arxiv_2402.18846", "arxiv_2006.15924", "arxiv_2305.04392", "arxiv_2407.15110", "arxiv_2002.02826", "arxiv_2207.01848", "arxiv_2502.17361", "arxiv_2310.03572" ]
q_2602.02886_core_query_0	[ { "id": "arxiv_2304.14068", "author_rationale": "Concept Embedding Models introduce concept embeddings to overcome the drop in accuracy which is usually experienced in concept bottleneck models. Those models preserve a certain level of interpretability and interventions. Nevertheless, the task predictor ope...	[ "arxiv_2503.04363", "arxiv_2202.01459", "arxiv_2406.17931", "arxiv_2401.14142", "arxiv_2209.09056", "arxiv_2504.03978", "arxiv_2401.08534", "arxiv_2410.06352", "arxiv_2407.03921", "arxiv_2406.19272" ]
q_2501.04126_core_query_0	[ { "id": "arxiv_1807.01622", "author_rationale": "Neural Processes showed that neural networks can learn distributions over functions directly from data and perform flexible regression conditioned on arbitrary observations. Despite these advantages, we observed important limitations in their probabilistic fo...	[ "arxiv_2404.02986", "arxiv_2303.00800", "arxiv_2206.03992", "arxiv_2302.00482", "arxiv_2212.00886", "arxiv_2011.01596", "arxiv_2106.05863", "arxiv_2404.13182", "arxiv_2202.01929", "arxiv_2302.07400", "arxiv_2002.06873" ]
q_2502.02513_core_query_0	[ { "id": "s2_6463cac46c43625ef78a9e0a3c2fc19affead187", "author_rationale": "This paper introduces the notion of generalized score matching, which is exactly the generalization we needed to define score matching on arbitrary representations of Lie algebras. Namely, such a generalization specifies which const...	[ "arxiv_2312.11707", "arxiv_2305.15586", "arxiv_2410.21853", "arxiv_2303.16169", "arxiv_2310.20030", "arxiv_2302.04411", "arxiv_2401.14131", "arxiv_2310.07216", "arxiv_2302.03660", "arxiv_2403.01430" ]
q_2504.10143_core_query_0	[ { "id": "s2_3750af62488e39c0595d097e399c970187be3428", "author_rationale": "CLAP was my earlier method-driven work showing that deliberately augmenting language prompts in contrastive learning can help separate semantic content from style, improving robustness and cross-domain generalization. The key idea I...	[ "arxiv_2405.18570", "arxiv_2411.02837", "arxiv_2412.07909", "arxiv_2304.03717", "arxiv_2305.05496", "arxiv_2310.04971", "arxiv_2303.05952", "arxiv_2404.01595", "arxiv_2302.06232" ]
q_2505.21955_core_query_0	[ { "id": "arxiv_2311.15596", "author_rationale": "This paper introduced a benchmark for evaluating whether LVLMs can effectively understand first-person views. Building on this idea, we extended the setting to investigate whether LVLMs can also integrate and understand additional third-person views when they...	[ "arxiv_2409.12969", "arxiv_2411.14869", "arxiv_2410.17385", "arxiv_2406.13642", "arxiv_2211.12436", "arxiv_2207.02202", "arxiv_2410.17856", "arxiv_2410.18962", "arxiv_2406.02537" ]
q_2504.05812_core_query_0	[ { "id": "arxiv_2411.16345", "author_rationale": "This paper uses majority voting as an effective learning objective for LLM reasoning. The new idea I brought into my own work is to implement this idea for a Base model without SFT." }, { "id": "arxiv_2302.09664", "author_rationale": "This paper u...	[ "arxiv_2407.13647", "arxiv_2407.08400", "arxiv_2412.17397", "arxiv_2501.12948", "arxiv_2411.16579", "arxiv_2412.02674", "arxiv_2404.07017", "arxiv_2410.23912", "arxiv_2504.16084", "arxiv_2411.04282", "arxiv_2504.13837" ]
q_2502.07861_core_query_0	[ { "id": "arxiv_2006.14009", "author_rationale": "This paper presents a very elegant algorithm with near-optimal guarantees for the vector balancing problem. This paper serves as a prior work to our result in the same way as for \"A Quasi-Monte Carlo Data Structure for Smooth Kernel Evaluations\". Its influe...	[ "arxiv_2412.08890", "arxiv_2405.14256", "arxiv_2403.04643", "arxiv_2403.05527", "arxiv_2309.16354", "arxiv_2402.02750", "arxiv_2410.15704", "arxiv_2311.14652", "arxiv_2412.16187", "arxiv_2406.03482" ]
q_2510.20733_core_query_0	[ { "id": "arxiv_2501.05707", "author_rationale": "The paper provides a framework for multi-agent fine-tuning, which inspired us to move from language to latents, based on the evidence that multi-agent collaboration works." }, { "id": "arxiv_2206.07751", "author_rationale": "The paper provides ide...	[ "arxiv_2402.18439", "arxiv_2310.10701", "arxiv_2407.12532", "arxiv_2212.01681", "arxiv_2211.04591", "arxiv_2203.03344", "arxiv_2407.11373", "arxiv_2307.01403", "arxiv_2107.10098", "arxiv_2110.15349", "arxiv_2002.09604" ]
q_2506.00635_core_query_0	[ { "id": "arxiv_2209.07522", "author_rationale": "Representative works on test-time training in recent years. It provides classic solutions for test-time training. The analysis of the theoretical effectiveness of test-time training inspired us to initiate this work." }, { "id": "arxiv_2202.11672", ...	[ "arxiv_2304.04795", "arxiv_2304.12566", "arxiv_2208.05117", "arxiv_2401.00989", "arxiv_2303.01630", "arxiv_2301.13018", "arxiv_2303.13899", "arxiv_2309.10109", "arxiv_2311.05858", "arxiv_2407.12492" ]
q_2508.17536_core_query_0	[ { "id": "arxiv_2305.14325", "author_rationale": "This work is one of the first works that proposed the Multi-Agent Debate protocol. The paper claimed that having multiple LLM agents engage in discussions improves task performance. It mostly compared with the single agent protocol and a single-agent iterativ...	[ "arxiv_2410.12853", "arxiv_2305.13160", "arxiv_2312.04854", "arxiv_2310.20151", "arxiv_2502.19130", "arxiv_2410.04663", "arxiv_2310.06272", "arxiv_2406.06561", "arxiv_2410.02506", "arxiv_2409.14051" ]
q_2511.03725_core_query_0	[ { "id": "arxiv_2304.06129", "author_rationale": "LF-CBM showed that human-understandable concepts can be obtained without manual concept annotations by leveraging LLMs and CLIP's dual-encoder representations. This inspired us to explore whether language-defined concepts could also make video models inherent...	[ "arxiv_2404.09490", "arxiv_2303.16268", "arxiv_2404.03645", "arxiv_2301.00436", "arxiv_2206.02846", "arxiv_2404.09067", "arxiv_2312.01431", "arxiv_2404.01591", "arxiv_2003.14285", "arxiv_2204.11929" ]
q_2602.03066_core_query_0	[ { "id": "arxiv_2008.00938", "author_rationale": "It defines the kernel target alignment of NTK. We use it to measure if shortcut features are top eigenfunctions of NTK." }, { "id": "arxiv_1912.01198", "author_rationale": "This shows that the network learns simple functions faster by NTK dynamics...	[ "arxiv_2105.05553", "arxiv_2310.18935", "arxiv_2206.08558", "arxiv_2410.10322", "arxiv_2312.07841", "arxiv_2206.01717", "arxiv_2010.02501", "arxiv_2205.06226", "arxiv_2405.18296" ]
q_2506.03642_core_query_0	[ { "id": "arxiv_2501.01428", "author_rationale": "GPT4Scene demonstrated that pre-trained vision-language models can be extended to understand 3D indoor scenes from videos without modifying their architectures. Its central contribution was to establish global–local correspondence by combining a reconstructed...	[ "arxiv_2403.13438", "arxiv_2408.06926", "arxiv_2410.16162", "arxiv_2501.00599", "arxiv_2405.03685", "arxiv_2412.00493", "arxiv_2109.01934", "arxiv_2411.16537" ]
q_2510.08132_core_query_0	[ { "id": "arxiv_2411.00409", "author_rationale": "This paper proposes a method to unlearn class-level information from pre-trained vision-language models without accessing the model's parameters. This paper made me think about the limitation of current machine unlearning on pre-trained vision-language models...	[ "arxiv_2410.03833", "arxiv_2303.05699", "arxiv_2407.11867", "arxiv_2404.12922", "arxiv_2405.15304", "arxiv_2407.07485", "arxiv_2312.00761", "arxiv_2104.02230", "arxiv_2406.06535", "arxiv_2410.23330", "arxiv_2510.11853" ]
q_2510.17515_core_query_0	[ { "id": "arxiv_2203.14328", "author_rationale": "This paper showed that randomly pruned neural networks can be analyzed with NTK theory in the infinite-width regime, where pruning preserves the dense NTK up to a sparsity-dependent scaling factor. Inspired by this, our work asks whether NTK analysis can be e...	[ "arxiv_2210.02412", "arxiv_2305.10550", "arxiv_2009.07439", "arxiv_2006.14548", "arxiv_2105.03703", "arxiv_2301.00327", "arxiv_2306.04495", "arxiv_2006.08228", "arxiv_2001.07301", "arxiv_2406.01820" ]
q_2508.02546_core_query_0	[ { "id": "arxiv_2402.09221", "author_rationale": "This paper was very relevant for the spectral analysis they used, and it was inspiring in general." }, { "id": "arxiv_2309.16588", "author_rationale": "The idea that attention sink could have a function, that in this case the author claims could b...	[ "arxiv_2211.11052", "arxiv_2106.01506", "arxiv_2010.09697", "arxiv_2111.05498", "arxiv_2412.02682", "arxiv_2306.13596", "arxiv_2410.19931" ]
q_2510.16582_core_query_0	[ { "id": "arxiv_2111.09266", "author_rationale": "It proposes the GFlownet framework, which is very inspiring to my work. The new idea I brought is to extend it to GraphRAG and explore how to implement it efficiently for graph retrieval. I think this will work because learning a policy network for GraphRAG h...	[ "arxiv_2405.13021", "arxiv_2502.14361", "arxiv_2412.15118", "arxiv_2412.01572", "arxiv_2108.12472", "arxiv_2502.20317", "arxiv_2405.13640", "arxiv_2402.11163", "arxiv_2503.00845", "arxiv_2202.13296" ]
q_2506.01320_core_query_0	[ { "id": "arxiv_2402.15194", "author_rationale": "They formulated sampling from the desired target distribution under a stochastic optimal control (SOC) framework and derived the optimal initial distribution required to achieve this goal. In particular, they identified this distribution as a reward-tilted Ga...	[ "arxiv_2302.06504", "arxiv_2409.04832", "arxiv_2408.08252", "arxiv_2405.13726", "arxiv_2502.00968", "arxiv_2309.06373", "arxiv_2501.00467", "s2_3ce93c1a772061161948a6728716f6f4d752809b", "arxiv_2310.01762", "arxiv_2209.14593" ]
q_2505.18600_core_query_0	[ { "id": "arxiv_2312.02149", "author_rationale": "The \"Generative Powers of Ten\" paper introduces a multi-scale diffusion sampling method that uses a pretrained text-to-image model to generate consistent content across extreme zoom levels. The proposed method is a novel technique for creating seamless zoom...	[ "arxiv_2412.09013", "arxiv_2402.03300", "arxiv_2306.00714", "arxiv_2411.12072", "arxiv_2411.18824", "arxiv_2104.14951", "arxiv_2405.10014", "arxiv_2411.16969", "arxiv_2311.15658", "arxiv_2403.01124", "arxiv_2310.01110", "arxiv_2410.14279", "arxiv_2302.07864" ]
q_2505.18943_core_query_0	[ { "id": "arxiv_2303.17760", "author_rationale": "Introduced inception prompting for communicative multi-agent cooperation, proving that assigning distinct, specialized roles to LLMs enables autonomous multi-agent dialogue and prevents goal drift. Deconstructed monolithic social processing by assigning disti...	[ "arxiv_2301.06926", "arxiv_2402.09836", "arxiv_2405.11841", "arxiv_2306.00924", "arxiv_2407.06813", "arxiv_2404.15515", "arxiv_2310.03051", "arxiv_2107.05697", "arxiv_2306.14325" ]
q_2504.11453_core_query_0	[ { "id": "arxiv_2111.08819", "author_rationale": "CleanRL is a repository of online RL algorithm implementations in single files, making them easy for researchers to vendor and modify. An early idea behind Unifloral was to create a similar repository for offline RL algorithms. This was guaranteed to work as ...	[ "arxiv_2210.07105", "arxiv_2406.09329", "arxiv_2110.14000", "arxiv_2110.02304", "arxiv_2110.04135", "arxiv_2212.08131", "arxiv_2306.00972", "arxiv_2106.06860", "s2_0ab3f7ecbdc5a33565a234215604a6ca9d155a33", "arxiv_2201.05433", "arxiv_2109.10813" ]
q_2509.22281_core_query_0	[ { "id": "arxiv_2406.03866", "author_rationale": "LLplace demonstrated that 3D indoor layout generation can be formulated as structured language modeling by fine-tuning an open-source LLM on dialogue-formatted 3D-FRONT data to generate room layouts and support object addition and removal, without relying on ...	[ "arxiv_2001.08481", "arxiv_2505.02836", "arxiv_2407.05425", "arxiv_2211.04604", "arxiv_2502.12894", "arxiv_2011.01126", "arxiv_2312.04533", "arxiv_2404.09465", "arxiv_2109.10460", "arxiv_2304.14391" ]
q_2510.21834_core_query_0	[ { "id": "arxiv_2310.15213", "author_rationale": "The task function vectors contain task-relevant semantics and can guide the model to generate task-relevant tokens during generation, and hence, improving performance. The lost components during pruning may also be represented as similar function vectors for ...	[ "arxiv_2410.14720", "arxiv_2408.14398", "arxiv_2405.17506", "arxiv_2407.13331", "arxiv_2312.15230", "arxiv_2204.09656", "arxiv_2310.13191", "arxiv_2407.16286", "arxiv_2404.03147", "arxiv_2406.15524" ]
q_2509.18376_core_query_0	[ { "id": "arxiv_2303.17651", "author_rationale": "This paper presented an approach for improving initial LLM outputs through iterative feedback and refinement prompts, operating at test time without requiring modifications to the model's weights. We adapted this framework to address two key technical challen...	[ "arxiv_2210.17159", "arxiv_2407.01979", "arxiv_2406.04548", "arxiv_2405.12654", "s2_09ebe119cc955c804e2a5f742a8483e725213eaf", "arxiv_2209.07924", "arxiv_2410.19978", "arxiv_2210.11695", "arxiv_2208.10609", "oa_W2086618114", "arxiv_2311.05764" ]
q_2510.27172_core_query_0	[ { "id": "arxiv_2410.10014", "author_rationale": "Safety-Aware Fine-Tuning (SAFT) introduced a data-centric defense that automatically identifies and removes potentially harmful examples before fine-tuning. Specifically, it constructs a harmfulness-related subspace from the model's internal representations, ...	[ "arxiv_2410.02220", "arxiv_2408.09600", "arxiv_2505.16559", "arxiv_2402.14968", "arxiv_2510.10085", "arxiv_2308.10544", "arxiv_2501.18100", "arxiv_2506.03850", "arxiv_2307.02222", "arxiv_2311.02105", "arxiv_2410.09760", "arxiv_2402.16382" ]
q_2512.02932_core_query_0	[ { "id": "arxiv_2403.17888", "author_rationale": "2DGS is an important work showing that replacing 3D Gaussians with 2D surfels can preserve real-time rendering while significantly improving geometric accuracy. By introducing 2D surfels together with a ray-splat intersection rasterizer, it achieves consisten...	[ "arxiv_2412.12734", "arxiv_2412.04955", "arxiv_2411.08508", "arxiv_2411.18966", "arxiv_2001.06782", "arxiv_2410.09467", "arxiv_2407.11840", "arxiv_2412.01823", "arxiv_2406.11672", "arxiv_2411.18625", "arxiv_2407.09733", "arxiv_2408.16982" ]
q_2501.06425_core_query_0	[ { "id": "arxiv_2305.13245", "author_rationale": "GQA showed that the KV-cache cost of multi-head attention can be substantially reduced by using fewer key-value heads than query heads, interpolating between MHA and MQA. Importantly, its experiments demonstrated that this structured reduction could retain qu...	[ "arxiv_2401.03462", "arxiv_2406.09297", "arxiv_2407.12866", "arxiv_2310.08152", "arxiv_2410.18517", "arxiv_2412.08890", "arxiv_2409.10593", "arxiv_2409.15012", "arxiv_2402.07616", "arxiv_2401.18079" ]
q_2505.17677_core_query_0	[ { "id": "arxiv_2211.16835", "author_rationale": "This work showed that hand motion naturally provides multiple views of a held object, enabling detailed 3D reconstruction without a learned object-shape prior. We took from it the idea that hand–object motion can provide useful geometric constraints, rather t...	[ "s2_48db41826f18061406c585eb0cd4318069a3cbc4", "arxiv_2106.08059", "arxiv_2012.09856", "arxiv_2409.02598", "arxiv_2108.07044", "arxiv_2309.10748", "arxiv_2308.11617", "arxiv_2307.10387", "arxiv_2106.11725" ]
q_2505.13878_core_query_0	[ { "id": "arxiv_2305.18290", "author_rationale": "They propose a preference alignment (PA) stage for LLMs and a direct preference alignment method, i.e., DPO. The new idea is PA for model fusion. It works because DPO works for one reference model, and it should possibly work for multiple reference models." ...	[ "arxiv_2502.14272", "arxiv_2404.12150", "arxiv_2406.18853", "arxiv_2302.08215", "arxiv_2505.13893", "arxiv_2410.13828", "arxiv_2408.06266", "arxiv_2411.17792", "arxiv_2407.04181", "arxiv_2402.00742" ]
q_2502.17159_core_query_0	[ { "id": "arxiv_2307.13269", "author_rationale": "Important baseline for parameter-efficient model merging." }, { "id": "arxiv_2212.04089", "author_rationale": "Fundamental baseline for model merging." }, { "id": "arxiv_2306.01708", "author_rationale": "Important baseline for full fin...	[ "arxiv_2412.12153", "arxiv_2410.17961", "arxiv_2408.13656", "arxiv_2410.17146", "arxiv_2405.17461", "arxiv_2411.14064", "arxiv_2310.02575", "arxiv_2412.17023", "arxiv_2402.16842", "arxiv_2410.13025", "arxiv_2312.04339" ]
q_2505.14460_core_query_0	[ { "id": "arxiv_2503.22679", "author_rationale": "RL for IQA." }, { "id": "arxiv_2501.12948", "author_rationale": "GRPO method." } ]	[ "arxiv_1707.08347", "arxiv_2406.09397", "arxiv_2307.09668", "arxiv_2412.05479", "arxiv_2402.03300", "arxiv_2402.03681", "arxiv_2412.06760", "arxiv_2404.01911", "arxiv_2401.06591", "arxiv_2402.02651", "arxiv_2411.12791" ]
q_2502.02770_core_query_0	[ { "id": "arxiv_2410.16179", "author_rationale": "This prior work explores the shortage of top-k attention, introducing sampling-based attention to solve the problem." }, { "id": "arxiv_2501.01005", "author_rationale": "This work built a high performance attention operator library FlashInfer. Twi...	[ "arxiv_2410.08584", "arxiv_2209.13802", "arxiv_2404.02690", "arxiv_2410.12876", "arxiv_1904.09751", "arxiv_2110.11299", "arxiv_2409.14846", "arxiv_2412.04652", "arxiv_2406.15486", "arxiv_2404.09336" ]
q_2510.20607_core_query_0	[ { "id": "arxiv_2206.15448", "author_rationale": "This paper introduces the framework of energy minimization to solve reasoning problems. This leads to an approach that expresses reasoning as an optimization problem, allowing problems to be solved through Langevin dynamics at inference time, instead of the s...	[ "arxiv_1702.08367", "arxiv_2301.03094", "arxiv_2006.07054", "arxiv_2206.00787", "arxiv_2205.10157", "arxiv_1703.00443", "arxiv_2203.06578", "arxiv_2007.12101", "arxiv_2305.07617", "arxiv_2303.00466" ]
q_2505.23758_core_query_0	[ { "id": "arxiv_2412.09611", "author_rationale": "This paper introduces the concept that semantic changes can be performed with attention-level blending. The paper utilizes this for global and local visual edits, such as attribute addition and style changes induced by text. Taking this observation as a start...	[ "arxiv_2212.04488", "arxiv_2404.03913", "arxiv_2404.05268", "arxiv_2403.19776", "arxiv_2410.04891", "arxiv_2406.13175", "arxiv_2412.09622", "arxiv_2305.01644", "arxiv_2408.03632" ]
q_2510.24461_core_query_0	[ { "id": "arxiv_2404.14964", "author_rationale": "This article motivates the analysis of the effect of surrogate gradient settings in deeper networks. They describe few neuron networks, and show that there is potential for gradient sign reversal, which essentially breaks training. The new idea for my researc...	[ "arxiv_2305.17650", "arxiv_2109.01905", "arxiv_2106.01862", "arxiv_2111.01750", "arxiv_2005.05941", "arxiv_2106.07854", "arxiv_2102.06408", "arxiv_2201.09754", "arxiv_2106.02681", "arxiv_2412.12858" ]
q_2505.19210_core_query_0	[ { "id": "s2_2377d7cf88966949ffdac9677930fca52108d3bf", "author_rationale": "Contribution: This paper shows that the eigendirections associated with differences between the covariance structures of two datasets can reveal features that are distinctive to one dataset relative to the other. These directions, r...	[ "arxiv_2407.02687", "arxiv_2411.16738", "arxiv_2410.13746", "arxiv_2207.12598", "arxiv_2402.03201", "arxiv_2208.11970", "arxiv_2312.07586", "arxiv_2406.02507", "arxiv_2406.08070", "arxiv_2404.07724" ]
q_2605.11387_core_query_0	[ { "id": "arxiv_1802.06070", "author_rationale": "DIAYN showed that mutual information can be used as an objective to learn diverse, distinguishable skills through reinforcement learning, even without an external reward function. In particular, it demonstrated that different latent skill variables could be a...	[ "arxiv_2205.03484", "arxiv_2207.05631", "arxiv_2403.09930", "arxiv_2007.13134", "arxiv_2406.00681", "arxiv_2205.13521", "arxiv_2204.08573", "arxiv_2412.01245", "arxiv_2305.18738" ]
q_2510.01265_core_query_0	[ { "id": "arxiv_2506.08007", "author_rationale": "This paper introduced a reasoning-based pretraining paradigm that reformulates next token prediction as a standard RLVR task with binary rewards. Using insights from this work, we designed a different reasoning-based algorithm which extends the supervision to...	[ "arxiv_2305.17077", "arxiv_2205.12496", "arxiv_2409.11402", "arxiv_2110.02370", "arxiv_2503.24290", "arxiv_2411.04282", "arxiv_2402.05808", "arxiv_2406.13763", "arxiv_2409.08642", "arxiv_2304.03984" ]
q_2505.14615_core_query_0	[ { "id": "arxiv_2410.07432", "author_rationale": "This is the paper that directly evaluates whether LLMs can solve SAT problems. This motivates us to believe that SAT is an interesting domain, and that benchmarking LLM capabilities can be interesting to show the community." }, { "id": "arxiv_2502.011...	[ "arxiv_2310.15164", "arxiv_2402.11291", "arxiv_2309.05452", "arxiv_2310.05993", "arxiv_2406.17169", "arxiv_2401.00757", "arxiv_2308.07336", "arxiv_2402.08064", "arxiv_2406.11035", "arxiv_2402.11903" ]
q_2510.01146_core_query_0	[ { "id": "arxiv_2505.13388", "author_rationale": "Rubrics but the English version, from here we know rubrics can be used for both LLM-as-a-judge and RL training, and SFT seems to be enough rather than redoing RL on reward models." }, { "id": "arxiv_2310.08491", "author_rationale": "They used a Li...	[ "arxiv_2406.03368", "arxiv_2401.03401", "arxiv_2402.16694", "arxiv_2403.13787", "arxiv_2309.07462", "arxiv_2003.11080", "arxiv_2506.20544", "arxiv_2401.07817", "arxiv_2308.16797" ]
q_2606.00995_core_query_0	[ { "id": "arxiv_2509.23886", "author_rationale": "We can localize subliminal learning to a set of sparse tokens in the training data, where an unbiased teacher would have selected a divergent response. Subliminal learning is fragile and early layers are important, which suggests we can find a real mechanism ...	[ "arxiv_2205.16004", "arxiv_2112.02505", "arxiv_2402.03119", "arxiv_2305.10010", "arxiv_2604.25783", "arxiv_2308.10248", "arxiv_2606.00831", "arxiv_2211.01071", "arxiv_2012.03236", "arxiv_2603.09517", "arxiv_2403.09053" ]
q_2510.00219_core_query_0	[ { "id": "arxiv_2405.16039", "author_rationale": "Ways of stabilizing and allocating additional computation. We leveraged the empirical and measurement approaches for this paper for our structural baseline. This was just read for evals, but pinned down how we were going to do our science." }, { "id":...	[ "arxiv_2107.00910", "arxiv_2006.09286", "arxiv_2410.06024", "arxiv_2111.12701", "arxiv_2203.12861", "arxiv_2305.12689", "arxiv_2002.07106", "arxiv_2402.13934", "arxiv_2402.11917", "arxiv_2203.13240" ]
q_2505.17353_core_query_0	[ { "id": "arxiv_1605.01710", "author_rationale": "The main contribution of this paper was their theoretical proof of the fixed point convergence of ADMM with any off-the-shelf pretrained denoising model. It is highly relevant to our paper since we also wanted to show a rigorous theoretical proof of our devel...	[ "arxiv_2304.11751", "arxiv_2302.05290", "arxiv_2405.10748", "arxiv_2407.20784", "arxiv_2206.00941", "arxiv_2410.05646", "arxiv_2407.16125", "arxiv_2412.16748", "arxiv_2402.02149", "arxiv_2405.17673", "arxiv_2411.15551" ]
q_2607.01208_core_query_0	[ { "id": "arxiv_2507.14805", "author_rationale": "Cloud et al. showed that behavioral traits can transfer between language models through semantically unrelated data, even after explicit references to the trait are filtered out. This suggested to us that the effect of a hidden conditioning prompt can remain ...	[ "arxiv_2406.12670", "arxiv_2305.13302", "arxiv_2207.10245", "arxiv_2410.22517", "arxiv_2109.04095", "arxiv_2406.03631", "arxiv_2207.04546", "arxiv_2006.07710", "arxiv_2403.00180", "arxiv_2305.07378", "arxiv_2103.00453" ]
q_2603.22435_core_query_0	[ { "id": "arxiv_2305.16291", "author_rationale": "This paper developed the concept of a \"skill library\" and how \"skill library is an evolving concept\". We applied a skill library when developing cap-agent0. It seems that high level skills are just compositions of low level skills. Thus this should work."...	[ "arxiv_2306.08647", "arxiv_2402.12275", "arxiv_2402.19299", "arxiv_2308.06810", "arxiv_2403.13801", "arxiv_2405.16450", "arxiv_2311.11183", "arxiv_2405.01534", "arxiv_2310.11604", "arxiv_2406.03757" ]
q_2606.19496_core_query_0	[ { "id": "arxiv_2510.10020", "author_rationale": "Calibrating Generative Models (CGM) was our group's prior work on a finetuning method for finetuning a generative model so that the means of features under the finetuned model match user-specified target means. Our idea was to replace the mean matching object...	[ "arxiv_2212.04473", "arxiv_2511.22640", "arxiv_2409.08861", "arxiv_2309.14054", "arxiv_2302.10688", "arxiv_2202.08937", "arxiv_2107.02212", "arxiv_2410.11646", "arxiv_2311.11973", "arxiv_2303.05699" ]
q_2509.25424_core_query_0	[ { "id": "arxiv_2505.22617", "author_rationale": "The paper contributes theoretical analysis of how entropy changes under standard RL gradient updates. This analysis (and the proof technique) was used in both deriving and analysing our method." }, { "id": "arxiv_2410.01679", "author_rationale": "...	[ "arxiv_2310.05324", "arxiv_2010.14484", "arxiv_2006.01419", "arxiv_2410.17126", "arxiv_2207.05631", "arxiv_2106.00707", "arxiv_2002.00632", "arxiv_2310.14509", "arxiv_2410.10431", "arxiv_2205.13521" ]
q_2510.20328_core_query_0	[ { "id": "arxiv_2502.13923", "author_rationale": "This technical report explains the capabilities of the Qwen2.5-VL family of the open-source vision language models (VLMs). One of the most important capabilities of this model for this project is its ability to understand long-context videos. The training dat...	[ "arxiv_2008.07451", "arxiv_2209.04899", "arxiv_2010.09170", "arxiv_2309.15278", "arxiv_2310.08581", "arxiv_2110.12810", "arxiv_2012.08977", "arxiv_2409.13682", "arxiv_2210.15629", "arxiv_2105.14039" ]
q_2505.09561_core_query_0	[ { "id": "arxiv_2010.14876", "author_rationale": "They tackle the copycat problem by looking into the correlations of the experts. This helped us think about action correlations and we started looking into generated action correlations and we identified that they were less correlated than the expert, making ...	[ "arxiv_2012.02788", "arxiv_2306.15156", "arxiv_2302.12422", "arxiv_2105.05484", "arxiv_2412.15427", "arxiv_2411.07954", "arxiv_2310.06171", "arxiv_2408.02024", "arxiv_2207.09705", "arxiv_2202.01312" ]
q_2505.08078_core_query_0	[ { "id": "arxiv_2401.12963", "author_rationale": "AutoRT is an orchestration pipeline that uses a VLM for scene understanding/grounding and an LLM to propose diverse, novel natural-language instructions for a fleet of real robots, so that data collection can be scaled up \"in the wild\" with minimal human su...	[ "arxiv_2312.11374", "arxiv_2109.10813", "arxiv_2208.00478", "arxiv_2304.10573", "arxiv_2407.20635", "arxiv_2310.15145", "arxiv_2303.01488", "arxiv_2302.02948", "arxiv_2402.04539", "arxiv_2108.03298", "arxiv_2205.03353" ]

End of preview. Expand in Data Studio

ScholarCatalyst: A Benchmark for Retrieving Papers that Inspire New Research

Project Page • Paper • Code

Benchmark Contributors

Adir Dayan, Ajay Sridhar, Alberta Longhini, Alberto Rota, Alexandru Oarga, Ali Hatamizadeh, Alireza Mousavi-Hosseini, Angana Borah, Anjiang Wei, Atharva Kulkarni, Bhuvan Sachdeva, Bo Liu, Bryan Sangwoo Kim, Burouj Armgaan, Camila Blank, Caroline Choi, Changho Shin, Changyeon Kim, Chaofan Lin, Chengyang He, Chuhan Li, David Anugraha, Dayoon Ko, Dilara Soylu, Dongjin Kang, Dongyang Fan, Ekaterina Kochetkova, Eunkyu Park, Fan Nie, Fanhu Zeng, Fatemeh Pesaran Zadeh, Francesco De Santis, Gaotang Li, Giung Nam, Guancheng Zhou, Guangnian Wan, Guijin Son, Gul Sena Altıntas, Gyuhyeon Seo, Han Luo, Hanane Moussa, Hangoo Kang, Haochen Tian, Haoyu Zhang, Haozhe Chen, Haozhe Wang, Himanshu Gupta, Ho Sy Tuyen, Hoang Pham, Houjun Liu, Hwan Chang, Hyeong Kyu Choi, Hyunji Lee, Hyunjun Lee, Ifdita Hasan Orney, Ilia Sucholutsky, Insu Lee, Itay Itzhak, Jaewoo Ahn, Jarek Liesen, Jason S Lucas, Jen-Yen Chang, Jiaju Ma, Jiaqi Liu, Jiayin Zhu, Jingzhe Shi, Jinheon Baek, Jinkun Hao, Jinwoo Lim, Jiwoo Chung, Johannes Kirmayr, Johnny Tian-Zheng Wei, Jongseo Lee, Joonghyuk Shin, Jubayer Ibn Hamid, Junchi Yu, Jungwoo Kim, Junyoung Lim, Kallol Saha, Kevin Qinghong Lin, Kodai Kawamura, Korneel Van den Berghe, Kunwoong Kim, Letian Fu, Linus Kreitner, Lucy Xiaoyang Shi, Lynnette Hui Xian Ng, Madhav Kanda, Mahyar Ghazanfari, Marcel Torne, Marco Bertolini, Marco Fumero, Maria Parelli, Marie Brockschmidt, Markus Gross, Mert Cemri, Michael Rizvi-Martel, Ming Hu, Mingtong Zhang, Mingyu Kim, Minki Kang, Minkyu Kim, Minseo Kim, Na Min An, Nathaniel L. Diamant, Niels Mündler, Nilesh Jain, Nilesh Prasad Pandey, Omin Kwon, Peng Xia, Pengcheng Wang, Pratik Sachdeva, Qingyang Zhang, Qizheng Zhang, Rahul Sharma, Raj Ghugare, Richard Cornelius Suwandi, Rohit Jena, Rosen Ting-Ying Yu, Roshen Sanjay Nair, Roussel Desmond Nzoyem, Rulin Shao, Runhan Huang, Sanghyun Jo, Seohong Park, Seokwon Song, Seonghyeon Ye, Seungone Kim, Shayan Talaei, Shihao Wang, Shiyin Jiang, Sijia Liu, Sijie Zhao, Siqi Zhu, Taehoon Yoon, Tanvir Ahmed Sijan, Tianhe Wu, Tongtian Zhu, Truong Buu Phan, Valentin Hofmann, Valeria Ruscio, Vasily Ilin, Wei Chen, Weiheng Liu, Wenbin Ouyang, Wenjie Zhu, Woomin Song, Woongyeong Yeo, Xiang Li, Xiaotian Liu, Xuan Luo, Xuanming Zhang, Xuying Ning, Yancheng Zhang, Yang Li, Yangcen Liu, Yanggan Gu, Yangtian Zhang, Yanjiang Guo, Yaozhong Shi, Yasaman Haghighi, Yichao Cai, Yifan Zhang, Yoonho Lee, Yoonkyo Jung, Yoonsang Lee, Yuante Li, Yuejiang Liu, Yujia Zheng, Yumin Choi, Yusuf Dalva, Zeyuan Chen, Zhaoyang Liu, Zheng Huang, Zhenhao Chen, Zhongxing Xu, Zijian Feng, Zixuan Hu

ScholarCatalyst is a literature inspiration retrieval benchmark grounded in researchers' firsthand knowledge of their own projects. 184 researchers who led 207 recent computer science projects verified 894 research questions as they stood before each project's key findings, then labeled which prior papers did or could have advanced their work and explained why, including papers they had not encountered at the time.

Dataset Structure

corpus (191k documents)

id: the document id
title: the paper title
text: the paper title followed by its abstract
primary_category: the paper's primary arXiv category, when the source paper is from arXiv
categories: every arXiv category the paper is listed under, when the source paper is from arXiv
published: the paper's publication date

queries

id: the query id
paper_id: the id of the source paper in corpus
question: the query text
type: core_query or subfield_query
paper_published: the source paper's publication date
paper_domain: the source paper's primary arXiv category

core_query and subfield_query, the relevance judgments, one query per source paper for core_query and one query per idea thread inside a source paper for subfield_query

query_id: the id of the query in queries
positive_docs: for core_query, a list of {id, author_rationale}, the papers the author identified as the key inspirations for the whole paper, each with the author's own rationale. For subfield_query, a list of document ids the author identified as inspirations for that idea thread
hard_negatives: a list of document ids that are topically similar to the query but are not an inspiration
author_rationale: present only in subfield_query, the author's explanation for why the positive documents were selected

Data Loading

from datasets import load_dataset

corpus = load_dataset("ScholarCatalyst/ScholarCatalyst", "corpus")["test"]
queries = load_dataset("ScholarCatalyst/ScholarCatalyst", "queries")["test"]
core_query = load_dataset("ScholarCatalyst/ScholarCatalyst", "core_query")["test"]
subfield_query = load_dataset("ScholarCatalyst/ScholarCatalyst", "subfield_query")["test"]

Dataset Statistics

Citation

@article{kim2026scholarcatalyst,
    title   = {ScholarCatalyst: A Benchmark for Retrieving Papers that Inspire New Research},
    author  = {Kim, Sohyeon and Lee, Yoonho and Liu, Bo and Ko, Dayoon and Shao, Rulin and Kim, Seungone and Neubig, Graham and Koh, Pang Wei and Chowdhery, Aakanksha and Asai, Akari and Khattab, Omar and Choi, Yejin and Kim, Gunhee and Finn, Chelsea},
    journal = {arXiv preprint arXiv:2610.02202},
    year    = {2026}
}

Downloads last month: 45

Paper for ScholarCatalyst/ScholarCatalyst

ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research

Paper • 2610.02202 • Published 1 day ago • 1