pair_id int64 0 99 | insight stringlengths 650 2.74k | paper1_scores float64 -5,484.72 -835.95 | paper2_scores float64 -6,548.31 -950.63 | joint_scores float64 -5,539.25 -981.31 | no_context_scores float64 -6,717.06 -578.01 | contrastive_loss float64 1.38k 14k | paper1_scores_avg float64 -15.41 -7.71 | paper2_scores_avg float64 -15.35 -7.96 | joint_scores_avg float64 -15.56 -7.5 | no_context_scores_avg float64 -15.55 -5.78 | abstract sequencelengths 2 2 | valid_abstracts bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | <insight>
Advanced AI systems often need to operate within complex behavioral boundaries, such as adhering to safety protocols or aligning with nuanced human preferences. One approach involves learning these boundaries implicitly as latent structures from available data, allowing policies to optimize objectives while r... | -1,362.5 | -2,230.5625 | -1,856.304688 | -2,224.21875 | 3,960.976563 | -7.921512 | -12.968387 | -10.792469 | -12.931504 | [
"LATENT SAFETY-CONSTRAINED POLICY APPROACH\nFOR SAFE OFFLINE REINFORCEMENT LEARNING\n\nPrajwal Koirala, Zhanhong Jiang, Soumik Sarkar & Cody Fleming\nIowa State University\nAmes, Iowa, USA\n{prajwal,zhjiang,soumiks,flemingc}@iastate.edu\n\nABSTRACT\n\nIn safe offline reinforcement learning (RL), the objective is to... | true |
0 | <insight>
Effectively guiding intelligent systems requires learning complex objectives that often involve balancing desired outcomes with critical constraints, such as adhering to safety protocols or aligning with nuanced human values. Two distinct strategies are emerging to tackle this challenge. One approach involves... | -1,440.859375 | -2,338.046875 | -1,824.90625 | -2,236.835938 | 4,190.835938 | -8.426078 | -13.672789 | -10.671967 | -13.080912 | [
"LATENT SAFETY-CONSTRAINED POLICY APPROACH\nFOR SAFE OFFLINE REINFORCEMENT LEARNING\n\nPrajwal Koirala, Zhanhong Jiang, Soumik Sarkar & Cody Fleming\nIowa State University\nAmes, Iowa, USA\n{prajwal,zhjiang,soumiks,flemingc}@iastate.edu\n\nABSTRACT\n\nIn safe offline reinforcement learning (RL), the objective is to... | true |
0 | <insight>Guiding complex agent behavior in reinforcement learning often requires more than just maximizing a predefined reward. Two distinct but complementary strategies emerge for learning these crucial guiding signals, especially when direct supervision or comprehensive data is lacking. One strategy focuses on inferr... | -1,409.46875 | -1,568.875 | -1,117.25 | -1,579.492188 | 3,440.585938 | -9.787977 | -10.894965 | -7.75868 | -10.968696 | [
"LATENT SAFETY-CONSTRAINED POLICY APPROACH\nFOR SAFE OFFLINE REINFORCEMENT LEARNING\n\nPrajwal Koirala, Zhanhong Jiang, Soumik Sarkar & Cody Fleming\nIowa State University\nAmes, Iowa, USA\n{prajwal,zhjiang,soumiks,flemingc}@iastate.edu\n\nABSTRACT\n\nIn safe offline reinforcement learning (RL), the objective is to... | true |
1 | <insight>
Achieving effective simultaneous alignment of representations from multiple distinct sources (like vision, audio, and text) within a shared high-dimensional embedding space requires addressing challenges beyond just the alignment mechanism itself. While geometric approaches, such as minimizing the volume span... | -1,300.796875 | -1,941.8125 | -1,987.75 | -1,831.402344 | 3,086.261719 | -8.446733 | -12.609172 | -12.907468 | -11.892223 | [
"GRAMIAN MULTIMODAL REPRESENTATION\nLEARNING AND ALIGNMENT\n\nGiordano Cicchetti∗, Eleonora Grassucci∗, Luigi Sigillo, Danilo Comminiello\nDept. of Information Engineering, Electronics, and Telecomm., Sapienza University of Rome, Italy\n{name.surname}@uniroma1.it\n\nABSTRACT\n\nHuman perception integrates multiple ... | true |
1 | <insight>
Developing advanced machine learning systems that operate effectively in high-dimensional latent or embedding spaces necessitates addressing two distinct but complementary challenges. Firstly, when integrating information from multiple sources (like different senses or data types), ensuring that their respect... | -1,865.929688 | -2,542.195313 | -2,586.5625 | -2,524.289063 | 4,345.851563 | -10.140923 | -13.816278 | -14.057405 | -13.718963 | [
"GRAMIAN MULTIMODAL REPRESENTATION\nLEARNING AND ALIGNMENT\n\nGiordano Cicchetti∗, Eleonora Grassucci∗, Luigi Sigillo, Danilo Comminiello\nDept. of Information Engineering, Electronics, and Telecomm., Sapienza University of Rome, Italy\n{name.surname}@uniroma1.it\n\nABSTRACT\n\nHuman perception integrates multiple ... | true |
1 | <insight>
Advancements are simultaneously being made in how multiple distinct information streams (like vision, audio, and text) are integrated and how generative models operate within abstract representational spaces. One line of work focuses on ensuring that the vector representations derived from different sources a... | -2,318.25 | -2,876.84375 | -2,973.78125 | -2,852.25 | 5,073.5625 | -11.199275 | -13.897796 | -14.366093 | -13.778986 | [
"GRAMIAN MULTIMODAL REPRESENTATION\nLEARNING AND ALIGNMENT\n\nGiordano Cicchetti∗, Eleonora Grassucci∗, Luigi Sigillo, Danilo Comminiello\nDept. of Information Engineering, Electronics, and Telecomm., Sapienza University of Rome, Italy\n{name.surname}@uniroma1.it\n\nABSTRACT\n\nHuman perception integrates multiple ... | true |
4 | <insight>While foundational theoretical work establishes guarantees for how learning algorithms can converge to optimal policies, even in complex, long-term scenarios, the practical assessment of sophisticated AI systems presents its own distinct set of challenges. Specifically, verifying the true capabilities of syste... | -1,819.6875 | -1,999.492188 | -2,005.96875 | -1,771.75 | 3,584.960938 | -12.378826 | -13.601988 | -13.646046 | -12.052721 | [
"GLOBAL CONVERGENCE OF POLICY GRADIENT IN AV-\nERAGE REWARD MDPS\n\nNavdeep Kumar∗\nElectrical and Computer Engineering\nTechnion - Israel Institute of Technology\nnavdeepkumar@campus.technion.ac.il\n\nYashaswini Murthy∗\nECE & CSL\nUniversity of Illinois Urbana-Champaign\nymurthy2@illinois.edu\n\nItai Shufaro\nEle... | true |
4 | <insight>
Significant effort is dedicated to rigorously proving the theoretical convergence and performance guarantees of complex learning algorithms, ensuring they can, in principle, reach optimal solutions even in challenging settings like infinite-horizon average-reward problems. However, the practical evaluation of... | -1,742.203125 | -1,939.3125 | -1,947.25 | -1,528.787109 | 3,263.052734 | -12.015194 | -13.374569 | -13.429311 | -10.54336 | [
"GLOBAL CONVERGENCE OF POLICY GRADIENT IN AV-\nERAGE REWARD MDPS\n\nNavdeep Kumar∗\nElectrical and Computer Engineering\nTechnion - Israel Institute of Technology\nnavdeepkumar@campus.technion.ac.il\n\nYashaswini Murthy∗\nECE & CSL\nUniversity of Illinois Urbana-Champaign\nymurthy2@illinois.edu\n\nItai Shufaro\nEle... | true |
4 | <insight>
Advancing artificial intelligence involves tackling complexity on two complementary fronts. On one hand, establishing foundational understanding requires rigorous theoretical analysis of learning algorithms, particularly under challenging, long-term scenarios (like optimizing for average rewards over infinite... | -2,230.78125 | -2,445.34375 | -2,432.21875 | -2,158.222656 | 4,402.128906 | -13.199889 | -14.46949 | -14.391827 | -12.770548 | [
"GLOBAL CONVERGENCE OF POLICY GRADIENT IN AV-\nERAGE REWARD MDPS\n\nNavdeep Kumar∗\nElectrical and Computer Engineering\nTechnion - Israel Institute of Technology\nnavdeepkumar@campus.technion.ac.il\n\nYashaswini Murthy∗\nECE & CSL\nUniversity of Illinois Urbana-Champaign\nymurthy2@illinois.edu\n\nItai Shufaro\nEle... | true |
5 | <insight>
Optimizing neural networks for specific goals, such as maximizing efficiency through compression or enhancing capability via large-scale pretraining for downstream tasks, reveals a common challenge: naive application of seemingly beneficial techniques can unexpectedly degrade the ultimate desired performance.... | -2,102.28125 | -2,250.40625 | -1,994.140625 | -2,299.359375 | 4,657.90625 | -12.013036 | -12.859465 | -11.395089 | -13.139196 | [
"QP-SNNS: QUANTIZED AND PRUNED SPIKING\nNEURAL NETWORKS\n\nWenjie Wei1, Malu Zhang1∗, Zijian Zhou1, Ammar Belatreche2,\nYimeng Shan3, Yu Liang1, Honglin Cao1, Jieyuan Zhang1, Yang Yang1\n\n1University of Electronic Science and Technology of China\n2Northumbria University, 3Liaoning Technical University\n\nABSTRACT\... | true |
5 | <insight>
Achieving optimal outcomes when refining complex computational models involves navigating subtle trade-offs beyond simple scaling or reduction. While techniques like quantization and pruning aim to create more efficient, lightweight models by reducing complexity, they risk performance degradation if not caref... | -1,423.164063 | -1,514.804688 | -1,342.328125 | -1,586.558594 | 3,182.199219 | -10.165458 | -10.820033 | -9.588058 | -11.332561 | [
"QP-SNNS: QUANTIZED AND PRUNED SPIKING\nNEURAL NETWORKS\n\nWenjie Wei1, Malu Zhang1∗, Zijian Zhou1, Ammar Belatreche2,\nYimeng Shan3, Yu Liang1, Honglin Cao1, Jieyuan Zhang1, Yang Yang1\n\n1University of Electronic Science and Technology of China\n2Northumbria University, 3Liaoning Technical University\n\nABSTRACT\... | true |
5 | <insight>
Achieving both high performance and computational efficiency in neural networks requires navigating complexities beyond optimizing either aspect in isolation. Techniques designed to make models smaller and faster, such as reducing parameter precision or removing redundant components, can have their effectiven... | -1,406.414063 | -1,535.28125 | -1,420.25 | -1,464.265625 | 2,985.710938 | -10.045815 | -10.966294 | -10.144643 | -10.459041 | [
"QP-SNNS: QUANTIZED AND PRUNED SPIKING\nNEURAL NETWORKS\n\nWenjie Wei1, Malu Zhang1∗, Zijian Zhou1, Ammar Belatreche2,\nYimeng Shan3, Yu Liang1, Honglin Cao1, Jieyuan Zhang1, Yang Yang1\n\n1University of Electronic Science and Technology of China\n2Northumbria University, 3Liaoning Technical University\n\nABSTRACT\... | true |
6 | <insight>
Hierarchical approaches to complex sequential decision-making, which involve breaking down problems into layers (e.g., strategic resource allocation followed by tactical item selection), may implicitly interact with the underlying geometric structure of the problem's representation space, particularly when le... | -2,208.296875 | -1,622.65625 | -1,689.3125 | -2,165.492188 | 4,307.132813 | -12.618839 | -9.272322 | -9.653214 | -12.374241 | [
"SEQUENTIAL STOCHASTIC COMBINATORIAL OPTI-\nMIZATION USING HIERARCHICAL REINFORCEMENT\nLEARNING\n\nXinsong Feng1, Zihan Yu2, Yanhai Xiong3, Haipeng Chen3\n1UCLA, 2The University of Hong Kong, 3William & Mary\nxsfeng@ucla.edu, u3634664@connect.hku.hk, {yxiong05, hchen23}@wm.edu\n\nABSTRACT\n\nReinforcement learning ... | true |
6 | <insight>
Hierarchical approaches designed to tackle complex, sequential decision-making problems may gain significant advantages in stability and effectiveness by aligning their structure with the intrinsic geometry of the underlying state space. The observation that states necessitating similar optimal actions or yie... | -2,947.65625 | -2,272.28125 | -2,314.109375 | -2,977.59375 | 5,883.421875 | -14.309011 | -11.030492 | -11.233541 | -14.454339 | [
"SEQUENTIAL STOCHASTIC COMBINATORIAL OPTI-\nMIZATION USING HIERARCHICAL REINFORCEMENT\nLEARNING\n\nXinsong Feng1, Zihan Yu2, Yanhai Xiong3, Haipeng Chen3\n1UCLA, 2The University of Hong Kong, 3William & Mary\nxsfeng@ucla.edu, u3634664@connect.hku.hk, {yxiong05, hchen23}@wm.edu\n\nABSTRACT\n\nReinforcement learning ... | true |
6 | <insight>
Complex sequential decision-making problems, often tackled with hierarchical learning structures that separate high-level strategy (like resource allocation over time) from low-level execution (like specific actions at each step), may benefit significantly from inherent geometric properties within the problem... | -2,238.070313 | -2,423.820313 | -2,583.34375 | -2,592.84375 | 4,671.390625 | -12.097677 | -13.101731 | -13.964021 | -14.015371 | [
"SEQUENTIAL STOCHASTIC COMBINATORIAL OPTI-\nMIZATION USING HIERARCHICAL REINFORCEMENT\nLEARNING\n\nXinsong Feng1, Zihan Yu2, Yanhai Xiong3, Haipeng Chen3\n1UCLA, 2The University of Hong Kong, 3William & Mary\nxsfeng@ucla.edu, u3634664@connect.hku.hk, {yxiong05, hchen23}@wm.edu\n\nABSTRACT\n\nReinforcement learning ... | true |
7 | <insight>
Advanced intelligent systems face distinct but complementary challenges related to temporal dynamics. One challenge involves learning robustly from external feedback sources whose quality or characteristics evolve over time, such as adapting to a human supervisor who becomes more proficient. This necessitates... | -1,261.994141 | -1,375.148438 | -1,535.21875 | -1,447.386719 | 2,549.310547 | -9.417867 | -10.262301 | -11.456857 | -10.801394 | [
"LEARNING FROM IMPERFECT HUMAN FEEDBACK: A\nTALE FROM CORRUPTION-ROBUST DUELING\n\nYuwei Cheng\nDepartment of Statistics\nUniversity of Chicago\nChicago, IL 60637, USA\nyuweicheng@uchicago.edu\n\nXuefeng Liu∗\nDepartment of Computer Science\nUniversity of Chicago\nChicago, IL 60637, USA\nxuefeng@uchicago.edu\n\nFan... | true |
7 | <insight>
Designing machine learning systems often involves tackling distinct challenges: one is modeling complex, structured data, such as sequences exhibiting rich temporal dependencies, requiring powerful architectures. Another is learning effectively from imperfect or evolving training signals, like human feedback ... | -1,402.859375 | -1,512.921875 | -1,611.960938 | -1,649.603516 | 2,953.423828 | -10.315143 | -11.124426 | -11.852654 | -12.129437 | [
"LEARNING FROM IMPERFECT HUMAN FEEDBACK: A\nTALE FROM CORRUPTION-ROBUST DUELING\n\nYuwei Cheng\nDepartment of Statistics\nUniversity of Chicago\nChicago, IL 60637, USA\nyuweicheng@uchicago.edu\n\nXuefeng Liu∗\nDepartment of Computer Science\nUniversity of Chicago\nChicago, IL 60637, USA\nxuefeng@uchicago.edu\n\nFan... | true |
7 | <insight>
Developing advanced intelligent systems requires solving two fundamentally different types of problems regarding information fidelity. First, systems must learn effectively even when guidance signals, such as human preferences, are inherently flawed, inconsistent, or change dynamically over time (potentially ... | -1,831.515625 | -1,976.296875 | -2,004.96875 | -2,118.816406 | 3,921.660156 | -11.305652 | -12.199364 | -12.37635 | -13.079114 | [
"LEARNING FROM IMPERFECT HUMAN FEEDBACK: A\nTALE FROM CORRUPTION-ROBUST DUELING\n\nYuwei Cheng\nDepartment of Statistics\nUniversity of Chicago\nChicago, IL 60637, USA\nyuweicheng@uchicago.edu\n\nXuefeng Liu∗\nDepartment of Computer Science\nUniversity of Chicago\nChicago, IL 60637, USA\nxuefeng@uchicago.edu\n\nFan... | true |
8 | <insight>
The internal activation states within large language models represent a key leverage point for achieving distinct goals without costly retraining. It's possible to manipulate these activations during inference in multiple ways simultaneously. For instance, one can introduce sparsity by zeroing out less signif... | -1,320.015625 | -2,039.234375 | -1,498.109375 | -1,859.4375 | 3,720.578125 | -8.919024 | -13.77861 | -10.12236 | -12.563766 | [
"TRAINING-FREE ACTIVATION SPARSITY IN\nLARGE LANGUAGE MODELS\n\nJames Liu1,2∗ Pragaash Ponnusamy2 Tianle Cai3 Han Guo1 Yoon Kim1 Ben Athiwaratkun2\n\n1 Massachusetts Institute of Technology 2 Together AI\n\n3 Princeton University\n\n(cid:135) https://github.com/FasterDecoding/TEAL\n\nABSTRACT\n\nActivation sparsity... | true |
8 | <insight>
The internal activation states within large language models possess a structure that enables different kinds of beneficial, training-free manipulation during inference. Widespread computational efficiency can be achieved by zeroing out a significant fraction of low-magnitude activation values globally across ... | -1,440.5625 | -2,095.375 | -1,681.90625 | -1,938.617188 | 3,792.648438 | -9.234375 | -13.431891 | -10.78145 | -12.427033 | [
"TRAINING-FREE ACTIVATION SPARSITY IN\nLARGE LANGUAGE MODELS\n\nJames Liu1,2∗ Pragaash Ponnusamy2 Tianle Cai3 Han Guo1 Yoon Kim1 Ben Athiwaratkun2\n\n1 Massachusetts Institute of Technology 2 Together AI\n\n3 Princeton University\n\n(cid:135) https://github.com/FasterDecoding/TEAL\n\nABSTRACT\n\nActivation sparsity... | true |
8 | <insight>
The internal activation states within large language models possess a remarkable degree of malleability that can be exploited *during inference* without any need for retraining or fine-tuning. This flexibility allows for fundamentally different goals to be achieved through direct, targeted manipulation of the... | -1,436.1875 | -2,213.5 | -1,711.953125 | -2,075.859375 | 4,013.59375 | -8.757241 | -13.496951 | -10.438739 | -12.65768 | [
"TRAINING-FREE ACTIVATION SPARSITY IN\nLARGE LANGUAGE MODELS\n\nJames Liu1,2∗ Pragaash Ponnusamy2 Tianle Cai3 Han Guo1 Yoon Kim1 Ben Athiwaratkun2\n\n1 Massachusetts Institute of Technology 2 Together AI\n\n3 Princeton University\n\n(cid:135) https://github.com/FasterDecoding/TEAL\n\nABSTRACT\n\nActivation sparsity... | true |
10 | <insight>
Advancements in complex machine learning domains, whether dealing with structured data like directed graphs or large opaque models like language models accessed via APIs, increasingly rely on developing highly specialized techniques to inject critical contextual information. For graphs, particularly directed ... | -1,457.984375 | -2,062.140625 | -1,572.4375 | -1,880.039063 | 3,827.726563 | -8.890148 | -12.574028 | -9.588034 | -11.463653 | [
"WHAT ARE GOOD POSITIONAL ENCODINGS FOR\nDIRECTED GRAPHS?\n\nYinan Huang\nGeorgia Institute of Technology\nyhuang903@gatech.edu\n\nPan Li\nGeorgia Institute of Technology\npanli@gatech.edu\n\nHaoyu Wang\nGeorgia Institute of Technology\nhaoyu.wang@gatech.edu\n\nABSTRACT\n\nPositional encodings (PEs) are essential f... | true |
10 | <insight>
Effectively interfacing with complex computational systems, whether they are graph neural networks processing intricate directed structures or large language models accessed as black boxes, increasingly relies on the sophisticated design and optimization of specialized input representations. Techniques are be... | -1,818.234375 | -2,650.5625 | -2,168.453125 | -2,591.4375 | 4,891.78125 | -9.420903 | -13.733484 | -11.235508 | -13.427137 | [
"WHAT ARE GOOD POSITIONAL ENCODINGS FOR\nDIRECTED GRAPHS?\n\nYinan Huang\nGeorgia Institute of Technology\nyhuang903@gatech.edu\n\nPan Li\nGeorgia Institute of Technology\npanli@gatech.edu\n\nHaoyu Wang\nGeorgia Institute of Technology\nhaoyu.wang@gatech.edu\n\nABSTRACT\n\nPositional encodings (PEs) are essential f... | true |
10 | <insight>
Effectively leveraging complex computational models requires tailored approaches depending on whether the primary challenge lies in representing intricate inherent structure or in optimizing interaction under constraints. For systems with rich internal relationships, such as directed graphs, advancements focu... | -1,659.804688 | -2,881.375 | -2,164.417969 | -2,846.441406 | 5,223.203125 | -8.340727 | -14.479271 | -10.876472 | -14.303725 | [
"WHAT ARE GOOD POSITIONAL ENCODINGS FOR\nDIRECTED GRAPHS?\n\nYinan Huang\nGeorgia Institute of Technology\nyhuang903@gatech.edu\n\nPan Li\nGeorgia Institute of Technology\npanli@gatech.edu\n\nHaoyu Wang\nGeorgia Institute of Technology\nhaoyu.wang@gatech.edu\n\nABSTRACT\n\nPositional encodings (PEs) are essential f... | true |
11 | <insight>
Evaluating the sophisticated reasoning capabilities of advanced AI systems requires moving beyond traditional static benchmarks that primarily test final answer accuracy. A more comprehensive assessment emerges from combining two complementary approaches. First, rigorously testing the system's understanding *... | -2,658.890625 | -2,241.9375 | -2,756.828125 | -2,656.363281 | 4,800.363281 | -13.776635 | -11.616257 | -14.284083 | -13.76354 | [
"IS YOUR MODEL REALLY A GOOD MATH REASONER?\nEVALUATING MATHEMATICAL REASONING WITH\nCHECKLIST\n\nZihao Zhou12∗ Shudong Liu3∗ Maizhen Ning126 Wei Liu4\nDerek F. Wong3 Xiaowei Huang2 Qiufeng Wang1† Kaizhu Huang6\n1Xi’an Jiaotong-liverpool University\n4HKUST\nhttps://mathcheck.github.io/\n\n5William & Mary 6Duke Kuns... | true |
11 | <insight>Evaluating the reasoning abilities of advanced AI systems requires moving beyond traditional static benchmarks that primarily measure problem-solving accuracy. A more comprehensive assessment strategy is emerging, characterized by two complementary thrusts. One focuses on deepening the evaluation within specif... | -3,147.25 | -2,583.945313 | -3,208.1875 | -3,190.59375 | 5,713.601563 | -14.503456 | -11.907582 | -14.784274 | -14.703197 | [
"IS YOUR MODEL REALLY A GOOD MATH REASONER?\nEVALUATING MATHEMATICAL REASONING WITH\nCHECKLIST\n\nZihao Zhou12∗ Shudong Liu3∗ Maizhen Ning126 Wei Liu4\nDerek F. Wong3 Xiaowei Huang2 Qiufeng Wang1† Kaizhu Huang6\n1Xi’an Jiaotong-liverpool University\n4HKUST\nhttps://mathcheck.github.io/\n\n5William & Mary 6Duke Kuns... | true |
11 | <insight>
Evaluating the reasoning abilities of advanced AI models requires moving beyond simply checking if they get the right answer on static problems. Two complementary strategies are emerging to achieve a deeper assessment. One involves systematically testing models across a variety of related tasks and conditions... | -3,011.75 | -2,393.3125 | -3,068.359375 | -3,023.742188 | 5,360.445313 | -14.549517 | -11.561896 | -14.822992 | -14.60745 | [
"IS YOUR MODEL REALLY A GOOD MATH REASONER?\nEVALUATING MATHEMATICAL REASONING WITH\nCHECKLIST\n\nZihao Zhou12∗ Shudong Liu3∗ Maizhen Ning126 Wei Liu4\nDerek F. Wong3 Xiaowei Huang2 Qiufeng Wang1† Kaizhu Huang6\n1Xi’an Jiaotong-liverpool University\n4HKUST\nhttps://mathcheck.github.io/\n\n5William & Mary 6Duke Kuns... | true |
14 | <insight>
The way language models apply safety restrictions can be brittle, as demonstrated by their differing responses to harmful requests depending on whether the request is phrased in the present or past tense. This suggests that the models' internal grasp of semantic meaning may not consistently generalize across ... | -1,277.828125 | -1,318.84375 | -1,336.189453 | -1,242.361328 | 2,502.84375 | -8.935861 | -9.222684 | -9.343982 | -8.687841 | [
"DOES REFUSAL TRAINING IN LLMS GENERALIZE TO\nTHE PAST TENSE?\n\nMaksym Andriushchenko\nEPFL\n\nNicolas Flammarion\nEPFL\n\nABSTRACT\n\nRefusal training is widely used to prevent LLMs from generating harmful, unde-\nsirable, or illegal outputs. We reveal a curious generalization gap in the current\nrefusal training... | true |
14 | <insight>
Large language models can exhibit surprising brittleness in their safety training, where simple linguistic modifications, such as changing a request's verb tense from present to past (e.g., "How to do X?" vs. "How did people do X?"), can circumvent safety guardrails designed to prevent harmful outputs. This s... | -3,453.03125 | -3,412.75 | -3,528.8125 | -3,520.65625 | 6,857.625 | -14.151768 | -13.98668 | -14.462346 | -14.428919 | [
"DOES REFUSAL TRAINING IN LLMS GENERALIZE TO\nTHE PAST TENSE?\n\nMaksym Andriushchenko\nEPFL\n\nNicolas Flammarion\nEPFL\n\nABSTRACT\n\nRefusal training is widely used to prevent LLMs from generating harmful, unde-\nsirable, or illegal outputs. We reveal a curious generalization gap in the current\nrefusal training... | true |
14 | <insight>
The observation that large language models' safety refusals can often be circumvented by merely changing the tense of a request (e.g., from present to past) reveals a subtle brittleness in current alignment techniques. This vulnerability becomes particularly salient when considered alongside efforts to develo... | -1,437.804688 | -1,564.546875 | -1,550.09375 | -1,481.179688 | 2,933.4375 | -9.585364 | -10.430312 | -10.333959 | -9.874531 | [
"DOES REFUSAL TRAINING IN LLMS GENERALIZE TO\nTHE PAST TENSE?\n\nMaksym Andriushchenko\nEPFL\n\nNicolas Flammarion\nEPFL\n\nABSTRACT\n\nRefusal training is widely used to prevent LLMs from generating harmful, unde-\nsirable, or illegal outputs. We reveal a curious generalization gap in the current\nrefusal training... | true |
15 | <insight>
Neural architectures demonstrate a strong capacity to approximate complex, continuous mathematical operators, such as those mapping inputs to the solutions of differential equations, especially when the target operator is well-defined and potentially mirrors classical iterative solution structures. However, w... | -1,614.53125 | -1,532.1875 | -1,597.34375 | -1,801.007813 | 3,350.382813 | -9.610305 | -9.120164 | -9.507998 | -10.720284 | [
"JAILBREAKING AS A REWARD MISSPECIFICATION\nPROBLEM\n\nJiahui Gao1† Lei Li1 Zhenguo Li2 Qi Liu1 Lingpeng Kong1†\n\nZhihui Xie1\n1The University of Hong Kong\n{zhxieml,ggaojiahui,nlp.lilei}@gmail.com\n{li.zhenguo}@huawei.com\nWARNING: This paper contains examples of harmful language.\n\n{liuqi,lpk}@cs.hku.hk\n\n2Hua... | true |
15 | <insight>
Neural systems are trained via optimization to approximate complex transformations, mapping inputs to desired outputs. This optimization process can be remarkably effective, sometimes resulting in models whose internal structure implicitly learns or mirrors classical iterative algorithms used for solving rela... | -1,887.71875 | -1,426.117188 | -1,726.53125 | -2,067.539063 | 3,654.84375 | -10.545915 | -7.967135 | -9.645426 | -11.550498 | [
"JAILBREAKING AS A REWARD MISSPECIFICATION\nPROBLEM\n\nJiahui Gao1† Lei Li1 Zhenguo Li2 Qi Liu1 Lingpeng Kong1†\n\nZhihui Xie1\n1The University of Hong Kong\n{zhxieml,ggaojiahui,nlp.lilei}@gmail.com\n{li.zhenguo}@huawei.com\nWARNING: This paper contains examples of harmful language.\n\n{liuqi,lpk}@cs.hku.hk\n\n2Hua... | true |
15 | <insight>
Neural networks exhibit a remarkable ability to learn approximations of complex functions and operators, whether these represent solutions to well-defined mathematical problems like differential equations or more nuanced behavioral objectives like safety alignment. However, the success and reliability of thes... | -2,550.8125 | -2,559.09375 | -2,364.570313 | -3,410.71875 | 6,156.054688 | -11.286781 | -11.323423 | -10.462701 | -15.091676 | [
"JAILBREAKING AS A REWARD MISSPECIFICATION\nPROBLEM\n\nJiahui Gao1† Lei Li1 Zhenguo Li2 Qi Liu1 Lingpeng Kong1†\n\nZhihui Xie1\n1The University of Hong Kong\n{zhxieml,ggaojiahui,nlp.lilei}@gmail.com\n{li.zhenguo}@huawei.com\nWARNING: This paper contains examples of harmful language.\n\n{liuqi,lpk}@cs.hku.hk\n\n2Hua... | true |
17 | <insight>
The development of advanced AI systems reveals a notable contrast: while one strand of research meticulously analyzes the theoretical properties and fundamental limitations of powerful model classes, uncovering potential mathematical discontinuities or convergence issues when examined under specific theoretic... | -1,549.9375 | -1,902.0625 | -1,942.539063 | -1,568.4375 | 3,077.898438 | -10.13031 | -12.431781 | -12.696334 | -10.251225 | [
"HIGHER-ORDER GRAPHON NEURAL NETWORKS:\nAPPROXIMATION AND CUT DISTANCE\n\nDaniel Herbst∗ and Stefanie Jegelka∗,†,‡\n∗TUM, School of CIT †TUM, MCML and MDSI\n{daniel.herbst, stefanie.jegelka}@tum.de\n\n‡MIT, Department of EECS and CSAIL\n\nABSTRACT\n\nGraph limit models, like graphons for limits of dense graphs, hav... | true |
17 | <insight>
Advancing artificial intelligence capabilities involves tackling distinct types of challenges. On one hand, increasing the theoretical expressive power of models, particularly those designed for complex structured data, can lead to nuanced mathematical issues, such as discontinuities with respect to certain c... | -1,976.4375 | -2,574.875 | -2,617.53125 | -2,212.871094 | 4,146.652344 | -11.041551 | -14.384776 | -14.623079 | -12.362409 | [
"HIGHER-ORDER GRAPHON NEURAL NETWORKS:\nAPPROXIMATION AND CUT DISTANCE\n\nDaniel Herbst∗ and Stefanie Jegelka∗,†,‡\n∗TUM, School of CIT †TUM, MCML and MDSI\n{daniel.herbst, stefanie.jegelka}@tum.de\n\n‡MIT, Department of EECS and CSAIL\n\nABSTRACT\n\nGraph limit models, like graphons for limits of dense graphs, hav... | true |
17 | <insight>
Achieving cutting-edge performance in complex AI tasks, whether analyzing intricate relational structures or enabling seamless multimodal interaction, often involves a trade-off between theoretical expressivity and practical system stability and efficiency. While research continually pushes the boundaries of ... | -1,594.6875 | -2,143.671875 | -2,177.96875 | -1,803.835938 | 3,364.226563 | -9.904891 | -13.314733 | -13.527756 | -11.20395 | [
"HIGHER-ORDER GRAPHON NEURAL NETWORKS:\nAPPROXIMATION AND CUT DISTANCE\n\nDaniel Herbst∗ and Stefanie Jegelka∗,†,‡\n∗TUM, School of CIT †TUM, MCML and MDSI\n{daniel.herbst, stefanie.jegelka}@tum.de\n\n‡MIT, Department of EECS and CSAIL\n\nABSTRACT\n\nGraph limit models, like graphons for limits of dense graphs, hav... | true |
18 | <insight>
Significant advancements in AI performance, spanning both complex reasoning and robust real-world identification tasks (including handling previously unseen examples), can be unlocked by integrating fundamentally distinct types of information. Rather than relying solely on text or isolated data streams, power... | -1,984.640625 | -1,365.546875 | -2,031.703125 | -1,529.765625 | 2,848.25 | -13.319736 | -9.164744 | -13.635592 | -10.266884 | [
"THINK-ON-GRAPH 2.0: DEEP AND FAITHFUL LARGE\nLANGUAGE MODEL REASONING WITH KNOWLEDGE-\nGUIDED RETRIEVAL AUGMENTED GENERATION\n\nShengjie Ma1,2∗, Chengjin Xu1∗†, Xuhui Jiang1∗, Muzhi Li3, Huaren Qu4,\nCehao Yang1, Jiaxin Mao2†, Jian Guo1†\n1 IDEA Research, International Digital Economy Academy, Shenzhen, Guangdong,... | true |
18 | <insight>Advanced AI systems demonstrate significant performance gains by integrating diverse information sources tailored to the task at hand. For complex reasoning challenges, combining structured knowledge graphs (representing entities and relationships) with unstructured text documents through iterative, tightly-co... | -2,493.625 | -1,499 | -2,647.125 | -2,131.367188 | 3,476.867188 | -14.49782 | -8.715117 | -15.390262 | -12.391669 | [
"THINK-ON-GRAPH 2.0: DEEP AND FAITHFUL LARGE\nLANGUAGE MODEL REASONING WITH KNOWLEDGE-\nGUIDED RETRIEVAL AUGMENTED GENERATION\n\nShengjie Ma1,2∗, Chengjin Xu1∗†, Xuhui Jiang1∗, Muzhi Li3, Huaren Qu4,\nCehao Yang1, Jiaxin Mao2†, Jian Guo1†\n1 IDEA Research, International Digital Economy Academy, Shenzhen, Guangdong,... | true |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 7