year int64 2.03k 2.03k | id stringlengths 10 10 | rating listlengths 0 9 | decision stringclasses 1
value | reviewer_comments listlengths 0 9 | _raw_metadata dict |
|---|---|---|---|---|---|
2,026 | 0b6a2SE23v | [
4,
8,
6,
6
] | [
{
"content": "This paper adapts pre-trained self-supervised representations into a single continuous latent token space and trains a lightweight latent generator with a generative decoder. This idea is potentially impactful for compute-efficient generation. However, at the reported compute and model sizes, the ... | {
"cdate": 1758285556767,
"content": {
"TLDR": {
"value": "We show that fine-tuned self-supervised tokens can serve as compact latents, enabling faithful single-token reconstruction and efficient generation."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025adapting,\ntitle={Adapting ... | |
2,026 | 0bGqD9hEcB | [
2,
4,
6,
6
] | [
{
"content": "This paper proposes TimeLT to tackle long-tail time series classification by considering multi-scal temporal encoding, data augmentation, and variant representation learning strategies. TimeLT demonstrates effectiveness across 4 selected datasets with ablation study and parameter analysis.",
"... | {
"cdate": 1758122923169,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025toward,\ntitle={Toward Robust Feature Space in Long-Tailed Time Series Classification: A Multi-Scale Perspective},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Confer... | |
2,026 | 0bPrfRIwPI | [
4,
2,
2,
8
] | [
{
"content": "This paper proposes the Fine-Grained Safety Neurons (FGSN) with Training-Free Continual Projection to mitigate safety risks arising from the fine-tuning of large language models (LLMs). Existing post-fine-tuning defense methods often rely on coarse-grained safety layer mapping, failing to comprehe... | {
"cdate": 1758271637393,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025finegrained,\ntitle={Fine-Grained Safety Neurons with Training-Free Continual Projection to Reduce {LLM} Fine Tuning Risks},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth Internatio... | |
2,026 | 0bgCa3XAnT | [
4,
4,
6,
4
] | [
{
"content": "The authors propose a heuristic method to approximately solve a wide class of combinatorial optimization problems, which seek to minimize a real-valued function $F(x)$ over a binary vector $x\\in\\\\{0,1\\\\}^n$. If such a function is replaced with its multilinear extension $f(x)$, the relaxation ... | {
"cdate": 1757856973932,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025smoothing,\ntitle={Smoothing Binary Optimization: A Primal-Dual Perspective},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\ny... | |
2,026 | 0blfYtdJES | [
6,
4,
6,
8
] | [
{
"content": "The paper proposes Sparkle, a unified intermediate representation for point cloud–based human motion capture that combines skeletal joints and surface anchors. Built into the SparkleMotion framework, it achieves state-of-the-art accuracy and robustness across 11 diverse datasets, with real-time pe... | {
"cdate": 1757825281523,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025sparkle,\ntitle={Sparkle: A Robust and Versatile Representation for Point Cloud-based Human Motion Capture},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference o... | |
2,026 | 0bvYEPH1O5 | [
2,
4,
4,
4
] | [
{
"content": "The paper investigates Low-Rank Adaptation (LoRA) and its variants. It introduces new, computationally more efficient extensions, such as Cheap LoRA (cLA) and its variants, along with a chained circulant variant LA. The study conducts theoretical analyses, including novel information-theoretic gen... | {
"cdate": 1756934655253,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\ncadenhead2025lora,\ntitle={Lo{RA}: The Past, Present, and Future},\nauthor={Elijah Cadenhead and Cristian McGee and Xin Li and El houcine Bergou and Aritra Dutta},\nyear={2025},\nurl={https://openreview.net/forum?id=0b... | |
2,026 | 0bwjkwSTuk | [
2,
4,
4
] | [
{
"content": "This paper addresses the limitation of text-to-image generation models in handling simple or underspecified prompts, which often result in suboptimal image-text alignment and visual quality. The authors propose a prompt rewriting framework that leverages large language models (LLMs) to refine user... | {
"cdate": 1758312016184,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nchen2025improving,\ntitle={Improving Text-to-Image Generation with Input-Side Inference-Time Scaling},\nauthor={Ruibo Chen and Jiacheng Pan and Heng Huang and Zhenheng Yang},\nyear={2025},\nurl={https://openreview.net/... | |
2,026 | 0bxjCu1trd | [
2,
2,
4,
8
] | [
{
"content": "Reward Model trained with human preferences are important for aligning LLMs but can break with distribution shift. The paper introduces a preference distribution agnostic approach to use the reward model to guide control decoding toward mis-specified responses. Using this approach, they train on ... | {
"cdate": 1758331719482,
"content": {
"TLDR": {
"value": "\\textsc{REFORM} lets a reward model find and fix its own blind spots, improving robustness on HH and Beavertails without hurting in distribution or downstream quality."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025teach,\... | |
2,026 | 0c7nAZjyr5 | [
6,
4,
8,
4
] | [
{
"content": "The paper proposes Seeing-to-Experiencing (S2E), a training recipe for goal-conditioned navigation that keeps visual priors from offline videos and adds interaction skills with reinforcement learning in simulation. It introduces a new action representation, using an anchor-conditioned Gaussian mix... | {
"cdate": 1758130752440,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025from,\ntitle={From Seeing to Experiencing: Scaling Navigation Foundation Models with Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference ... | |
2,026 | 0c9AtxLj3T | [
2,
4,
4,
4
] | [
{
"content": "The paper introduces an algorithm named PEARL-Prox which is the proximal counterpart of the PEARL-SGD algorithm used in the multiplayer federated learning problem. The authors define the conception of player drift and analyze the proposed algorithm in the exact / inexact setting under assumptions ... | {
"cdate": 1758205535827,
"content": {
"TLDR": {
"value": "Proposes Per-Player Local Proximal Algorithm (PEARL-Prox) resolving player drift in Multiplayer Federated Learning (MpFL)."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025pearlprox,\ntitle={{PEARL}-Prox: Proximal Algorithm f... | |
2,026 | 0cBlhTTHfx | [
2,
4,
2,
4
] | [
{
"content": "The paper proposes DynamicEval, a benchmark consisting of systematically curated prompts emphasizing dynamic camera motion.",
"id": "B5YRpEFwb6",
"rating": 2
},
{
"content": "This paper introduces **DynamicEval**, a new benchmark and metric suite for evaluating *text-to-video (T2... | {
"cdate": 1758346265734,
"content": {
"TLDR": {
"value": "DynamicEval is a benchmark for camera-motion video generation, introducing interpretable background and foreground consistency metrics that align better with human preferences than existing deep-feature methods."
},
"_bibtex": {
"val... | |
2,026 | 0cKUfYFeaf | [
6,
2,
4,
2
] | [
{
"content": "This manuscript investigates how the difficulty composition of training data affects the performance and cost efficiency of neural PDE solvers. The experiments and analysis take the example of 2D incompressible N.-S. equation along two axes: geometry (number and complexity of obstacles) and physic... | {
"cdate": 1757870923288,
"content": {
"TLDR": {
"value": "Pre-generating lower difficulty data improves the few-shot performance of neural PDE solvers."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025pregenerating,\ntitle={Pre-Generating Multi-Difficulty {PDE} Data For Few-Shot Neu... | |
2,026 | 0cbUKCyBsH | [
4,
2,
4,
4
] | [
{
"content": "Paper attributes the performance bottleneck of TSF to the “self-excitation” assumption and introduces the IATSF paradigm along with the lightweight model FIATS. By incorporating external influences such as text as conditional inputs, it seeks to lower the theoretical lower bound of prediction erro... | {
"cdate": 1758192238422,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025influenceaware,\ntitle={Influence-Aware Forecasting: Breaking the Self-Stimulation Barrier in Time Series},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on... | |
2,026 | 0cdXElXkk6 | [
2,
4,
4,
4
] | [
{
"content": "This paper analyzes the role of Classifier-Free Guidance (CFG) in conditional diffusion models, specifically examining its differential impact on low- and high-frequency components during image generation. The authors posit that an excessively high CFG scale applied to the low-frequency domain res... | {
"cdate": 1758211843751,
"content": {
"TLDR": {
"value": "We show that applying classifier-free guidance in the frequency domain substantially improves the quality at low guidance scales, while inherently avoiding the shortcomings associated with high guidance values."
},
"_bibtex": {
"valu... | |
2,026 | 0cmuYj3WeG | [
6,
6,
6,
4
] | [
{
"content": "This paper presents PALC (Preference Alignment via Logit Calibration), a novel and lightweight method for preference alignment during LLM inference time. By employing a bottleneck architecture between the final hidden states and output logits, PALC learns to generate position-dependent calibration... | {
"cdate": 1758351174908,
"content": {
"TLDR": {
"value": "PALC: preference alignment via logit calibration. Learns compact calibrations for frozen LLMs, achieving strong alignment without external rewards or fine-tuning. Outperforms most test-time methods with minimal latency."
},
"_bibtex": {
... | |
2,026 | 0crU7lZV8n | [
4,
4,
2,
2
] | [
{
"content": "This paper proposes RAPCal (Rank-Preserving Adaptive Pseudo-Calibration), a two-stage framework for calibrating LLMs under both domain and model shifts, even when target-domain labels are unavailable and the source model is black-box.\nRAPCal first performs source pre-calibration using an EM-based... | {
"cdate": 1757771714834,
"content": {
"TLDR": {
"value": "A calibration method for LLMs under both model and domain shift."
},
"_bibtex": {
"value": "@misc{\nhu2025rankpreserving,\ntitle={Rank-Preserving Calibration of {LLM}s Under Model and Distribution Shifts},\nauthor={Jian Hu and Qunli ... | |
2,026 | 0czAcXMBNO | [
2,
4,
2,
4
] | [
{
"content": "The paper studies length bias in Group Relative Policy Optimization (GRPO) style RL with verifiable rewards for LLM reasoning tasks. The authors propose λ-GRPO, a unified view of GRPO-style objectives for RLVR, and introduce a learnable sample-level weighting over responses, adjusting the length p... | {
"cdate": 1756736295666,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025lambdagrpo,\ntitle={\\${\\textbackslash}lambda\\$-{GRPO}: Unifying the {GRPO} Frameworks with Learnable Token Preferences},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth Internation... | |
2,026 | 0dHrYUd17W | [
6,
2,
4,
4
] | [
{
"content": "This paper proposes MambaSIC, a novel stereo image compression framework that leverages the Mamba-based Stereo Visual State Space Block (Stereo VSSB) to efficiently capture long-range inter-view correlations with linear complexity. It further introduces a bi-directional multi-reference entropy mod... | {
"cdate": 1757234058929,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nqin2025mambasic,\ntitle={Mamba{SIC}: Mamba-based Stereo Image Compression with Bi-directional Multi-reference Entropy Model},\nauthor={Shiyu Qin and Xinjie Zhang and Zhening Liu and Jinpeng Wang and Bin Chen and Jiawei... | |
2,026 | 0dOrtl8hwj | [
2,
2,
2,
2
] | [
{
"content": "The authors propose an extension of MoEs by which they modulate features to route data more effectively through the network. They validate this approach through experiments on MT10 and MT50 from Meta-World, alongside experiments in vision and language modelling.",
"id": "4wb2DqFAgX",
"rati... | {
"cdate": 1757315998586,
"content": {
"TLDR": {
"value": "We propose global neural inhibition network and empirically shows improvements across mixture-of-experts architectures, particularly in multi-task reinforcement learning."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025globa... | |
2,026 | 0dnSbhNoTB | [
4,
8,
2,
4
] | [
{
"content": "This paper proposes selecting a subset of motion-influential data for training video generative models. The goal is to choose data that strongly affects motion and to guide data curation in a way that improves temporal consistency and physical plausibility. A computational method is used to compar... | {
"cdate": 1757812169610,
"content": {
"TLDR": {
"value": "Our method, MOTIVE, is a scalable, motion-centric data attribution framework for video generative models."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025where,\ntitle={Where is Motion From? Scalable Motion Attribution for V... | |
2,026 | 0eEtTsnmyo | [
4,
2,
6
] | [
{
"content": "The paper connects Deep SVDD and normalizing flows for anomaly detection by introducing a subclass called Uniformly Scaling Flows (USFs), where the Jacobian determinant is constant across inputs. Under this setup, the maximum-likelihood objective of a flow becomes equivalent to the Deep SVDD loss,... | {
"cdate": 1758209496628,
"content": {
"TLDR": {
"value": "We uncover a theoretical bridge between deep one-class classification (e.g., Deep SVDD) and a class of normalizing flows called Uniformly Scaling Flows (USFs)."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025on,\ntitle={On U... | |
2,026 | 0eM74HjPQA | [
6,
4,
6,
6,
4
] | [
{
"content": "This paper presents new convergence guarantees for stepsized Newton methods under Holder continuity assumptions on the Hessian or third derivatives. The authors reinterpret the classical Newton method as a third-order tensor method and propose a family of stepsize schedules (RN), as well as linese... | {
"cdate": 1756982748131,
"content": {
"TLDR": {
"value": "We analyze the stepsized Newton method under various Holder continuity assumptions, including Holder continuity of third derivatives. We present the first stepsize schedule with $\\mathcal O(k^{-3})$ global convergence rate."
},
"_bibtex":... | |
2,026 | 0eenZ5FvH1 | [
2,
2,
2,
4
] | [
{
"content": "The paper proposes SampleCLR, a contrastive learning framework for learning sample representations. The learned sample representations are benchmarked on the COVID-19 dataset in unsupervised and supervised settings.",
"id": "9WpDIz3V9l",
"rating": 2
},
{
"content": "This manuscript... | {
"cdate": 1758330163563,
"content": {
"TLDR": {
"value": "We create a contrastive learning framework for efficient self-supervised learning of sample representations from single-cell data"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025great,\ntitle={Great patients embed alike: con... | |
2,026 | 0elvad3gEu | [
6,
6,
6,
4
] | [
{
"content": "The paper shows that the test error of infinite and finite-width networks is upper bound by a value related to the smallest eigenvalue of the (empirical) NTK. Based on the theory, the authors provide a training-free method to determine the cardinal width, which is the critical width where the test... | {
"cdate": 1757984683396,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025trainingfree,\ntitle={Training-Free Determination of Network Width via Neural Tangent Kernel},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Rep... | |
2,026 | 0etzWKrS4F | [
4,
4,
4,
6
] | [
{
"content": "The paper investigates why VLMs fail on abstract visual reasoning tasks such as the Bongard problems. The paper proposes the Linear Separability Ceiling (LSC), a diagnostic measure of how well a simple linear classifier performs on a VLM’s visual embeddings. It serves as a baseline to test whether... | {
"cdate": 1757007063068,
"content": {
"TLDR": {
"value": "next token + contrastive objective in vlms helps resolve alignment gap between perception and reasoning"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025beyond,\ntitle={Beyond the Linear Separability Ceiling: Aligning Represe... | |
2,026 | 0fJKg3fX41 | [
2,
4,
2,
2
] | [
{
"content": "The paper proposes DLM-3D, a framework that brings diffusion language models (discrete diffusion in token space) to 3D point-cloud generation. The method (i) tokenizes a point cloud into discrete “semantic” tokens with a permutation-invariant tokenizer based on FPS anchors + PointNet patch encoder... | {
"cdate": 1758187156320,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025dlmd,\ntitle={{DLM}-3D: Diffusion Language Models for 3D Point Clouds Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations... | |
2,026 | 0fNQCOWKc1 | [
2,
6,
6,
4
] | [
{
"content": "The paper introduces a system called UniVA to help with complex video tasks. UniVA is designed to combine many different video tasks, such as understanding, cutting, editing, and creating videos, into one workflow. UniVA uses two types of AI agents to work. A planner agent receives a high-level re... | {
"cdate": 1757594291015,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025univa,\ntitle={Uni{VA}: Universal Video Agents towards Next-Generation Video Intelligence},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repres... | |
2,026 | 0fcVDzkGK2 | [
4,
4,
0
] | [
{
"content": "The paper proposes a game-theoretic approach to combine multiple different text-to-image diffusion models. A crucial constraint is that the models must have the same latent dimensions. Ensuring fairness when dividing the noise maps among the models prevents the collapse into a single concept and o... | {
"cdate": 1758183743243,
"content": {
"TLDR": {
"value": "a game-theoretic approach to compositional sampling from multiple pre-trained diffusion models"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025divideanddenoise,\ntitle={{DIVIDE}-{AND}-{DENOISE}: A {GAME} {THEORETIC} {METHOD}... | |
2,026 | 0fg3OTEUFF | [
4,
2,
4,
2
] | [
{
"content": "The paper introduces a framework for LLM self-correction, where the main idea is to generate an intermediate task abstraction before refining the answer. This is opposed to the vanilla setup that critiques the output directly, the propose approach distills the problem and output into a structured ... | {
"cdate": 1758307901968,
"content": {
"TLDR": {
"value": "We show the advantages of task abstraction to self-correct the initial responses on large and small LLMs"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025selfcorrection,\ntitle={Self-Correction via Task Distillation},\nauthor... | |
2,026 | 0fgsHvmBBI | [
8,
2,
6,
2
] | [
{
"content": "The compilation of a PyTorch 2.0 model goes through multiple stages. Dynamo, the stage responsible for running the model for the first time and recording the computation graph, can be split into two passes: Torch-IR (a higher-level graph that still resembles model layers) and Aten-IR (a lower-leve... | {
"cdate": 1758325264827,
"content": {
"TLDR": {
"value": "An automated approach for lifting Sequence Parallelism, and other targeted memory optimisations for long-context training, into the compiler."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025autosp,\ntitle={Auto{SP}: Unlockin... | |
2,026 | 0fib2BYc0L | [
6,
6,
4,
4
] | [
{
"content": "This paper presents a vision-only framework for indoor scene spatial understanding, GPT4Scene. The paper introduces two main innovations: (1) feeding BEV images into the VLM to provide global scene perception, and (2) assigning consistent object-level markers across the BEV view and multiple frame... | {
"cdate": 1757665042923,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025gptscene,\ntitle={{GPT}4Scene: Understand 3D Scenes from Videos with Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repr... | |
2,026 | 0fk3GVbJPm | [
2,
6,
6,
4
] | [
{
"content": "This paper presents SylCipher, a syllable-based unsupervised ASR (UASR) system that jointly predicts syllable boundaries and embedding tokens from raw speech using a unified self-supervised objective. The authors conduct experiments across domains (LibriSpeech, SpokenCOCO) and languages (English, ... | {
"cdate": 1757424169294,
"content": {
"TLDR": {
"value": "A syllable-level automatic speech recognizer trained with unpaired speech and text for langauge-universal speech technology"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025towards,\ntitle={Towards Unsupervised Speech Recogni... | |
2,026 | 0fuYOuJyzl | [
4,
6,
6,
6
] | [
{
"content": "This work identifies the vulnerabilities of existing alignment schemes and the limitations of deep alignment solutions. Based on observations of intrinsic safety signals, it proposes an inference-time alignment scheme. By injecting a designed safety checkpoint during token generation, it recalls t... | {
"cdate": 1758336477980,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025anydepth,\ntitle={Any-Depth Alignment: Unlocking Innate Safety Alignment of {LLM}s to Any-Depth},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning ... | |
2,026 | 0fvVI2rORC | [
2,
2,
2,
2
] | [
{
"content": "This paper investigates whether large language models can automatically model executable programs into formal specifications suitable for model checking.\nThe authors propose MODEL-BENCH, a benchmark and pipeline that converts Python programs into TLA+ specifications to test LLMs’ ability to produ... | {
"cdate": 1758360179100,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025can,\ntitle={Can Large Language Models Model Programs Formally?},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\... | |
2,026 | 0g5Dk4Qfh0 | [
10,
4,
6,
8
] | [
{
"content": "The paper tackles personalized federated learning (PFL) under realistic heterogeneity: data heterogeneity where each client has distinct multi-modal tasks with temporal shifts, and model heterogeneity where clients use different model families and sizes. It introduces FedMosaic, which combines rel... | {
"cdate": 1758249305511,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025not,\ntitle={Not All Clients Are Equal: Collaborative Model Personalization on Heterogeneous Multi-Modal Clients},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Confer... | |
2,026 | 0gDKeKwFj6 | [
2,
6,
2
] | [
{
"content": "This paper aims to address the episodic knowledge binding problem in LLMs; that is, enhancing their ability to interlink and recall multiple related events, each defined by a tuple of time, space, entity, and content. The authors generate a series of synthetic episodic events and evaluate model pe... | {
"cdate": 1758210559017,
"content": {
"TLDR": {
"value": "We characterize a new challenge in LLM continual learning where models trained on separate episodic events, fail to semantically bind them. We provide a benchmark and human-inspired baseline which we termed generative cued recall."
},
"_bi... | |
2,026 | 0gKCVb8jpl | [
6,
6,
2,
4
] | [
{
"content": "This paper challenges the “frozen world” assumption underlying most reasoning-model evaluations—the idea that model reasoning unfolds in static, unchanging contexts. It systematically investigates how Large Reasoning Models (LRMs) behave when their reasoning process is interrupted mid-inference or... | {
"cdate": 1757908634870,
"content": {
"TLDR": {
"value": "This paper demonstrates that large reasoning models, can lose up to 60% accuracy when subjected to interruptions and in-flight context updates, due to reasoning leakage, self-doubt, and panic."
},
"_bibtex": {
"value": "@inproceeding... | |
2,026 | 0gh7haE5tc | [
4,
4,
6
] | [
{
"content": "This paper introduces a framework for evaluating the influence of training data in the context of MAML, which is formulated as a bilevel optimization problem. The authors propose task influence functions (task-IF) and instance influence functions (instance-IF). These methods are designed to accura... | {
"cdate": 1757693819953,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025evaluating,\ntitle={Evaluating Data Influence in Meta Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\n... | |
2,026 | 0h5ohpUGY4 | [
4,
6,
4,
4
] | [
{
"content": "This paper proposes a novel framework, UniDD, that unifies existing Dataset Distillation (DD) methods from the perspective of spectral filtering. The authors interpret different distillation objective functions as filtering functions applied to the feature-feature correlation matrix (FFC) and the ... | {
"cdate": 1758203375115,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025understanding,\ntitle={Understanding Dataset Distillation via Spectral Filtering},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations... | |
2,026 | 0hLuQAT3fV | [
8,
4,
4,
4
] | [
{
"content": "The paper introduces a universal adversarial perturbation approach for protecting images from diffusion-based editing. Instead of optimizing perturbations per image, it learns a single perturbation that can be applied universally. The method combines a semantic injection loss that aligns perturbed... | {
"cdate": 1757677827083,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025universal,\ntitle={Universal Image Immunization against Diffusion-based Image Editing via Semantic Injection},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference... | |
2,026 | 0hck6JFV1n | [
2,
2,
2,
4
] | [
{
"content": "The authors propose a special network architecture to reconstruct Cryo-ET images and correct for the missing wedge in the measurement setup. They test this network on a dataset and report better results.",
"id": "wYHEZ8EXN7",
"rating": 2
},
{
"content": "This paper proposes the Lat... | {
"cdate": 1758346711248,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025latent,\ntitle={Latent Back-projection Network: a missing-wedge generative model for cryo-electron tomography},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conferenc... | |
2,026 | 0hdSiZ1br0 | [
2,
4,
4
] | [
{
"content": "This paper proposes a watermarking method for discrete diffusion language models, where the decoding order is controlled based on binary hash rules to embed watermark signals. This method is compatible with common decoding strategies and can be further enhanced by beam search. Experimental results... | {
"cdate": 1758149435093,
"content": {
"TLDR": {
"value": "dMARK embeds robust watermarks in discrete diffusion LLMs by guiding decoding order with a parity key, achieving strong detectability without degrading text quality."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025dmark,\nti... | |
2,026 | 0hmBDnWeEK | [
6,
4,
2,
2
] | [
{
"content": "This paper proposes AttPortrait, a dual-branch diffusion framework for identity-attribute conditional face generation. In addition to the standard denoising branch, a disentanglement branch with explicit attribute supervision is introduced to decouple identity and attribute representations.\nThe m... | {
"cdate": 1758251712537,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025a,\ntitle={A Dual-Branch Disentanglement Diffusion for {ID}-Attribute Conditional Face Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learnin... | |
2,026 | 0hnnPy6vt2 | [
2,
2,
2
] | [
{
"content": "In this paper, the authors proposed p-Mark, a distortion-free watermarking method for LLM. p-Mark use beta distribution to adjust the distribution of the original LM. It selects the top-p tokens and add bias to them. Experiments on multiple LLMs show that p-Mark improves perplexity compared to bas... | {
"cdate": 1756897285084,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025distortionfree,\ntitle={Distortion-free Watermarking for Large Language Models via Adaptive Top-\\$p\\$ Sampling},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Confer... | |
2,026 | 0hy9kJ1ULB | [
6,
8,
6,
8
] | [
{
"content": "This paper introduces Mixture-of-Groups Attention (MoGA), a sparse attention mechanism designed to alleviate the quadratic complexity of Diffusion Transformers in long video generation. MoGA employs a lightweight, learnable token router to assign tokens into semantically coherent groups and perfor... | {
"cdate": 1756740189766,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025moga,\ntitle={Mo{GA}: Mixture-of-Groups Attention for End-to-End Long Video Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Represent... | |
2,026 | 0iLBGsGTS9 | [
2,
4,
4,
2
] | [
{
"content": "This paper studies knowledge distillation for hypergraph neural networks. It proposes HSelKD, a selective knowledge distillation framework that transfers task-relevant knowledge from an HGNN teacher to an MLP student. HSelKD leverages inverse optimal transport to distill the most informative parts... | {
"cdate": 1758205980546,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025hselkd,\ntitle={{HS}el{KD}: Selective Knowledge Distillation for Hypergraphs using Optimal Transport},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Lear... | |
2,026 | 0iN4UKZwgn | [
6,
6,
4,
6
] | [
{
"content": "This paper proposes AnomReason, a large-scale benchmark annotated with structured quadruples to capture commonsense, physical, and relational inconsistencies in generated images. The dataset is built using AnomAgent, a modular multi-agent framework that decomposes anomaly detection into entity par... | {
"cdate": 1757938821566,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025semantic,\ntitle={Semantic Visual Anomaly Detection and Reasoning in {AI}-Generated Images},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repre... | |
2,026 | 0ifNNBqBlQ | [
2,
4,
4
] | [
{
"content": "The paper proposes GiPFE (Gini-guided Progressive Frequency Extraction), a model-agnostic add-on for time-series forecasting focusing on frequency modeling. It claims that spectral energy imbalance causes models to overfit high-energy frequency components. Experiments on five benchmarks demonstrat... | {
"cdate": 1758116202853,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025addressing,\ntitle={Addressing Spectral Energy Imbalance in Time-Series Forecasting with Gini-Guided Progressive Frequency Extraction},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth... | |
2,026 | 0ik0xKYQd1 | [
0,
4,
2,
2,
4
] | [
{
"content": "The paper suggests a method for network training based on noisy labels. The technique gradually 'corrects' the labels by shifting them from the original label to the predicted class.",
"id": "Hxplj7jER4",
"rating": 0
},
{
"content": "This paper introduced a noise correction techniq... | {
"cdate": 1757908438590,
"content": {
"TLDR": {
"value": "assign each sample a trainable trust parameter that shifts from label to prediction to correct noise"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025linc,\ntitle={Li{NC}: Lightweight Noise Correction via Adaptive Label Refin... | |
2,026 | 0imrI7UXdu | [
4,
6,
6,
4
] | [
{
"content": "This work introduces a specialized Gaussian Splatting framework for ultrasound imaging, designed to reconstruct 3D volumes from traditional 2D scans. Instead of using conventional camera-style projection, it models the acoustic image-formation process via probe-plane intersection rendering. This a... | {
"cdate": 1758121461701,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025ultragauss,\ntitle={UltraGauss: Ultrafast Gaussian Reconstruction of 3D Ultrasound Volumes},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repre... | |
2,026 | 0izu8IWJ5z | [
4,
4,
4,
4
] | [
{
"content": "This paper proposes Neural Gaussian Radio Fields (nGRF), an explicit neural field model for channel estimation that represents wireless environments with 3D Gaussian primitives. Each Gaussian acts as a localized “radio modulator,” and the channel is rendered via complex-valued aggregation that mod... | {
"cdate": 1758354742697,
"content": {
"TLDR": {
"value": "This work introduces nGRF, a framework using explicit 3D Gaussian primitives for MIMO channel estimation by directly modeling the physical superposition of electromagnetic waves, achieving state-of-the-art accuracy with a 220x speedup."
},
... | |
2,026 | 0j0MmK7EMA | [
4,
4,
4
] | [
{
"content": "This paper introduces SimpleFold, a novel protein folding model based on flow-matching and a standard Transformer architecture. The central thesis challenges the necessity of complex, domain-specific architectural components (like MSAs, explicit pair representations, triangle updates, equivariant ... | {
"cdate": 1758213653984,
"content": {
"TLDR": {
"value": "We tackle protein folding as it was a text-to-3D generative model"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025simplefold,\ntitle={SimpleFold: Folding Proteins is Simpler than You Think},\nauthor={Anonymous},\nbooktitle={... | |
2,026 | 0jHyEKHDyx | [
8,
4,
6,
8
] | [
{
"content": "This paper investigates a long-standing training instability in transformer models when using flash attention with BF16 precision, which manifests as catastrophic loss explosions during training. It provides a mechanistic explanation for this failure as (1) the emergence of structurally similar lo... | {
"cdate": 1758264196512,
"content": {
"TLDR": {
"value": "For the first time, we mechanistically explain why low-precision training with flash attention fails, identifying a vicious cycle of rounding errors and proposing a simple, effective fix."
},
"_bibtex": {
"value": "@inproceedings{\na... | |
2,026 | 0k0RfnsEh3 | [
4,
4,
4,
6
] | [
{
"content": "The paper presents PointLAM for point-based 3D object detection. First, the paper proposes Dynamic Point Sampler (DPS) that curates an information-rich and structurally representative subset of raw points. Second, synergizes Bi-Directional Mamba (BDM) layers for global context modeling, and Loc... | {
"cdate": 1758203603121,
"content": {
"TLDR": {
"value": "An effective and efficient 3D backbone for point-based 3D object detection."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025pointlam,\ntitle={Point{LAM}: Local Attentive Mamba for Efficient Point-based 3D Object Detection},\... | |
2,026 | 0k5w8O0SNg | [
6,
6,
4,
6
] | [
{
"content": "This paper introduces Cartridges, a memory-efficient, high-throughput, and general-purpose solution for processing long contexts in LLMs, promising to replicate the functionality of in-context learning (ICL). The authors argue that existing LLMs rely on costly KV caches for long-context processing... | {
"cdate": 1758259072436,
"content": {
"TLDR": {
"value": "We show how to use offline synthetic data generation and training to reduce long context memory consumption."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025cartridges,\ntitle={Cartridges: Lightweight and general-purpose lon... | |
2,026 | 0kHbD6ad07 | [
4,
4,
6,
6
] | [
{
"content": "This is a primarily theory paper that discusses decoder only transformers being injective. Specifically that each unique prompt maps to a unique last token embedding (from final transformer layer). Paper provides theoretical setup and justification for this argument under acceptable assumptions at... | {
"cdate": 1758266731499,
"content": {
"TLDR": {
"value": "We prove that transformers are (a.s.) injective and propose an algorithm that provably inverts their hidden representations back to the original input prompt."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025language,\ntitle=... | |
2,026 | 0kZKtWleMS | [
4,
4,
6,
8
] | [
{
"content": "This work presents MoTIF, an interpretable framework for video classification that adapts the concept bottleneck model (CBM) idea to temporal data. The method introduces a transformer-like architecture with per-concept temporal attention, enabling explanations at global, local, and temporal levels... | {
"cdate": 1757484524248,
"content": {
"TLDR": {
"value": "MoTIF extends concept bottlenecks to video with per-channel temporal attention, enabling interpretable classification for arbitrary-length clips."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025concepts,\ntitle={Concepts in ... | |
2,026 | 0kiFgLo5al | [
2,
2,
2,
6,
4
] | [
{
"content": "This work trains LLMs with GRPO/DPO on synthetic graph connectivity and shortest-path problems, and finds that the reasoning ability can generalize to real-world multi-hop QA and structured planning. The further analysis shows the misalignment between single-hop and multi-hop results.",
"id": ... | {
"cdate": 1758229443110,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nzhang2025generalizable,\ntitle={Generalizable {LLM} Learning of Graph Synthetic Data with Post-training Alignment},\nauthor={Yizhuo Zhang and Heng Wang and Shangbin Feng and Zhaoxuan Tan and Xinyun Liu and Yulia Tsvetk... | |
2,026 | 0kjCyq9Hbo | [
2,
4,
8,
6,
2
] | [
{
"content": "In this paper, the authors study the capability of sequential enumeration of five different LLMs using two tasks: numerosity naming and numerosity production. In addition, the paper also studies whether an “internal counter mechanism” emerges in the model by looking into the latent space of them. ... | {
"cdate": 1758277710560,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025sequential,\ntitle={Sequential Enumeration in Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={202... | |
2,026 | 0klioDjSVM | [
4,
6,
4,
6
] | [
{
"content": "This paper introduces Delta-Triplane Transformers (DTT), a new 4D occupancy world model (OWM) for autonomous driving. Unlike previous works (e.g., DOME, OccWorld), DTT does not predict the full occupancy state but instead models changes (deltas) in a compact triplane representation (xy/xz/yz). By ... | {
"cdate": 1758208778705,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025deltatriplane,\ntitle={Delta-Triplane Transformers as Occupancy World Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\n... | |
2,026 | 0knDtnbYGd | [
4,
2,
4
] | [
{
"content": "This paper proposes RDAR , an RL-based framework for estimating which surrounding agents truly influence the ego vehicle’s behavior in autonomous driving. Instead of processing all scene entities with quadratic attention cost, RDAR formulates relevance estimation as an MDP, where actions are binar... | {
"cdate": 1758353059480,
"content": {
"TLDR": {
"value": "We propose a reiforcement learning framework to train a scoring model that ranks agents in a driving scene based on their importance for the decision-making process."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025rdar,\ntit... | |
2,026 | 0lGVMSAazo | [
4,
4,
4,
2
] | [
{
"content": "This paper introduces TFCA-Attention, a training-free dynamic sparse attention mechanism designed to solve the $O(L^2)$ computational bottleneck for long-context LLMs. The authors identify that existing methods are \"siloed,\" accelerating either prefilling or decoding, but not both. TFCA-Attentio... | {
"cdate": 1757490065027,
"content": {
"TLDR": {
"value": "A dynamic sparse attention method for unified acceleration of both prefilling, decoding and KV cache reduction in LLM inference."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025one,\ntitle={One Stone Three Birds: Training-fr... | |
2,026 | 0lW2UBiEWN | [
4,
6,
4,
4
] | [
{
"content": "The paper introduces MESA&MASK, a deception‐oriented benchmark that contrasts neutral system prompts (MESA) with pressure-inducing system prompts (MASK) across 2,100 scenarios spanning 6 deception types × 6 domains, and evaluates 22 LLMs with a rubric-guided LLM-as-judge (GPT-4.1) on both final an... | {
"cdate": 1758123535691,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025mesa,\ntitle={Mesa and Mask: A Benchmark for Detecting and Classifying Deceptive Behaviors in {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Lear... | |
2,026 | 0lct7PrPgS | [
6,
6,
6
] | [
{
"content": "This paper addresses the problem of learning linear state-space models (LSSMs) with sparse system matrices, a setting relevant to many real-world dynamical systems where interactions are limited. The authors propose a maximum a posteriori (MAP)–EM algorithm that introduces Student-t sparsity-promo... | {
"cdate": 1756860486655,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025learning,\ntitle={Learning linear state-space models with sparse system matrices},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations... | |
2,026 | 0lsidbAjNW | [
6,
4,
6,
2
] | [
{
"content": "This paper tackles the challenge of accurately predicting the multi-scale scientific impact (specifically future citation counts at yearly and monthly horizons) of research papers, a task complicated by heterogeneous factors and the limitations of existing methods in leveraging domain knowledge. C... | {
"cdate": 1758281469468,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025modeling,\ntitle={Modeling Multi-Scale Scientific Impact via Heterogeneous Networks and {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning R... | |
2,026 | 0mBCl2goJr | [
6,
4,
6,
6
] | [
{
"content": "This paper introduces a new task called Text-to-ImageSet (T2IS) generation, which aims to create coherent sets of images that share visual consistency while following diverse text instructions. To support research on this task, the authors propose a benchmark dataset named T2IS-Bench, along with a... | {
"cdate": 1758261801244,
"content": {
"TLDR": {
"value": "We propose Text-to-ImageSet (T2IS) task and develop unified framework for both T2IS evaluation and generation."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025why,\ntitle={Why Settle for One? Text-to-ImageSet Generation and... | |
2,026 | 0mCMKqgXZI | [
4,
2,
4,
4
] | [
{
"content": "This paper introduces Resolving Interference (RI), a lightweight framework to reduce cross-task interference in model merging, defined as merged models’ representation drift from constituent task-specialized models. RI disentangles expert models into functionally orthogonal subspaces using only un... | {
"cdate": 1757927180011,
"content": {
"TLDR": {
"value": "Lightweight adaptation stratergy to reduce cross-task interference to improve the performance of existing merging methods."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025resolving,\ntitle={Resolving Interference ({RI}): Dis... | |
2,026 | 0mNnINd2z5 | [
6,
2,
4
] | [
{
"content": "This paper addresses the inefficiency of uniform test-time compute allocation for LLMs and proposes a bandit learning framework for strategic compute distribution. It formulates test-time compute allocation as a bandit problem, treating each query as an action and allocating compute sequentially t... | {
"cdate": 1758330070137,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025strategic,\ntitle={Strategic Scaling of Test-Time Compute: A Bandit Learning Approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representa... | |
2,026 | 0mUiXz1TNq | [
2,
6,
4,
6
] | [
{
"content": "This paper proposes a new dataset, VUDG, to test domain generalisation in video understanding models. It includes 11 domains across three kinds of shifts: **semantic** (e.g., cartoon), **viewpoint** (e.g., egocentric), and **environmental conditions** (e.g., foggy). VUDG has the following qualitie... | {
"cdate": 1758252698046,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025vudg,\ntitle={{VUDG}: A Dataset for Video Understanding Domain Generalization},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\... | |
2,026 | 0mYcWbQyo7 | [
2,
4,
6,
2
] | [
{
"content": "This study propose a method which is effective under the continual unlearning situation where the unlearning requests accumulate over the course of time.\n\nThe authors create an unlearning sentence embedder with a synthetically generated dataset designed to enable the formation of sharp decision ... | {
"cdate": 1758175275969,
"content": {
"TLDR": {
"value": "We present a novel method for achieving effective continual unlearning in large language models in real-time."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025care,\ntitle={Ca{RE}: Continual Real-time Unlearning with Ensured ... | |
2,026 | 0miO9v1jeC | [
4,
4,
2,
2
] | [
{
"content": "Inspired by the structure of the human brain, this paper proposes TAR, a framework for token-level semantic correction. Based on the authors’ experiments, the proposed framework appears to yield certain performance improvements.\n\nHowever, overall, I find the claimed connection between the method... | {
"cdate": 1758356571640,
"content": {
"TLDR": {
"value": "We propose a brain-inspired Token Adaptive Routing framework that enables LLMs to self-correct token-level semantic errors, improving reasoning accuracy while reducing inference tokens."
},
"_bibtex": {
"value": "@inproceedings{\nano... | |
2,026 | 0miqobfGyv | [
4,
4,
4
] | [
{
"content": "This paper introduces AnomalyLMM, a training-free framework that leverages the advanced reasoning capabilities of existing LMMs for text-based person anomaly search. The authors propose a novel cloze-based re-ranking method, which encourages LMMs to fill in generated cloze-style text queries and r... | {
"cdate": 1757000559575,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025anomalylmm,\ntitle={Anomaly{LMM}: Bridging Generative Knowledge and Discriminative Retrieval for Text-based Person Anomaly Search},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth Int... | |
2,026 | 0mqsIlMtfm | [
4,
2,
2,
4
] | [
{
"content": "This paper proposes PTQTP, a simple data-free quantization scheme. The core idea is to decompose weight vectors into linear combinations of two trit-plane vectors. The optimization is performed using an alternating greedy algorithm: given the trit-plane vectors, the linear combination coefficients... | {
"cdate": 1758185401637,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nxiao2025ptqtp,\ntitle={{PTQTP}: Post-Training Quantization to Trit-Planes for Large Language Models},\nauthor={He Xiao and RUNMING YANG and Qingyao Yang and Wendong XU and Zhen Li and Yupeng Su and Zhengwu Liu and Hong... | |
2,026 | 0n1YcK0yuQ | [
6,
2,
2,
4
] | [
{
"content": "The authors study how moral reasoning in LLMs changes depending on the language used to prompt the model. To do this, they translate the MoralExceptQA and ETHICS datasets into 5 languages and compare the behaviors of 7 models on these datasets. They demonstrate that LLMs reason differently and com... | {
"cdate": 1758296106996,
"content": {
"TLDR": {
"value": "LLMs show inconsistent moral judgments across languages due to English-centric training, revealing cultural misalignments and the need for more culturally aware AI."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025one,\ntitle... | |
2,026 | 0n7dDguNeJ | [
2,
2,
4,
6
] | [
{
"content": "This work proposes a novel monolithic large vision-language model based on embedding the visual modality inside an LLM through lower-rank adaptation layers. Coupled with this architecture, the work also proposes a new alignment training stage for aligning the earlier LLM blocks' representations wi... | {
"cdate": 1758188499574,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025vision,\ntitle={Vision as Lo{RA}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/fo... | |
2,026 | 0ngAxtR0IH | [
4,
6,
6,
2
] | [
{
"content": "This paper proposes a novel multimodal framework for generating accurate CAD models from minimal inputs—a brief text prompt and a single sketch or image. It enhances user prompts into detailed CAD descriptions using a fine-tuned LLM. Also it proposes a hierarchical multi-modality fusion module to ... | {
"cdate": 1757299165477,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025hmfusion,\ntitle={{HMF}usion: Hierarchical Multi-Modality Fusion for {CAD} Representation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repres... | |
2,026 | 0nmGpN7Qqa | [
2,
4,
6,
4
] | [
{
"content": "This paper introduces PiCSAR, a training-free best-of-n reranking method that scores each candidate reasoning chain by the joint log-likelihood of the chain and its final answer, combining “reasoning confidence” log p(r | x) with “answer confidence” log p(y | r,x). PiCSAR operationalizes this wit... | {
"cdate": 1758294544463,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025picsar,\ntitle={Pi{CSAR}: Probabilistic Confidence Selection And Ranking for Reasoning Chains},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Re... | |
2,026 | 0nvQ5kHXf4 | [
6,
6,
6
] | [
{
"content": "This paper proposes Weight–Activation Subspace Iteration (WASI), a method that performs model training entirely within a low-rank subspace of both weights and activation. Experimental results demonstrate that WASI significantly reduces memory usage and roughly halves the computational cost (FLOPs)... | {
"cdate": 1758206892672,
"content": {
"TLDR": {
"value": "We propose a novel method that enables training vision transformer models within a low-rank subspace to optimize computational resources, making on-device learning practically feasible."
},
"_bibtex": {
"value": "@inproceedings{\nano... | |
2,026 | 0oHaazjMUX | [
6,
4,
6,
6
] | [
{
"content": "The paper introduces LeSTD (Learning-based Sparse Tensor Decomposition), a data-free, post-training compression framework for large language models. It addresses the dense core bottleneck in tensor decomposition methods by learning a shared basis across attention heads and then applying a theoreti... | {
"cdate": 1758132569495,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025lestd,\ntitle={Le{STD}: {LLM} Compression via Learning-based Sparse Tensor Decomposition},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Represe... | |
2,026 | 0oXyMbPMtP | [
2,
8,
4,
6
] | [
{
"content": "This paper presents MambaVoiceCloning (MVC), a text-to-speech framework that removes all attention and recurrent components from the encoder and conditioning path at inference, relying solely on SSMs. The proposed system includes three Mamba structured modules. \nThe key claimed is that MVC is the... | {
"cdate": 1758168690784,
"content": {
"TLDR": {
"value": "SSM-only TTS conditioning at inference (no attention/RNN); a gated Bi-Mamba improves long-form stability/streaming and gives ~1.6× encoder speed with modest, statistically significant quality gains."
},
"_bibtex": {
"value": "@inproc... | |
2,026 | 0obDxMfeYu | [
2,
2,
4,
2,
6
] | [
{
"content": "This paper introduces Medix, a method for Out-of-Distribution (OOD) detection that utilizes unlabeled \"in-the-wild\" data. Its core contribution is a two-stage process: (1) filtering potential OOD samples from the unlabeled data by identifying points whose gradient's element-wise median deviates ... | {
"cdate": 1758218712188,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025a,\ntitle={A Median Perspective on Unlabeled Data for Out-of-Distribution Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representati... | |
2,026 | 0oqEBQA0UD | [
2,
4,
4,
4
] | [
{
"content": "This work proposes a new method for de novo sequencing, called PeakNovo. The method involves two novel components: a masked self-distillation approach that is designed to help the model cope with missing peaks, and an \"MS fusion\" approach that makes use of spectra derived from an external datab... | {
"cdate": 1757965034364,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025peaknovo,\ntitle={PeakNovo: Towards the Robust De Novo Peptide Sequencing for Missing Spectral Peaks},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Lear... | |
2,026 | 0oxkxG9cCo | [
2,
2,
2
] | [
{
"content": "This work presents a new 3D scene dataset (IL3D) for LLM-driven 3D scene generation. This dataset is constructed by integrating, cleaning, and supplementing existing popular datasets (i.e., 3D-FRONT and HSSD) and then adding their own synthetic data to enhance scene diversity and cover underrepres... | {
"cdate": 1757067142609,
"content": {
"TLDR": {
"value": "IL3D, a large-scale dataset with 27,816 indoor layouts and 29,215 3D assets, supports LLM-driven 3D scene generation. It includes natural language annotations and rigorous benchmarks, enhancing the 3D scene generation task."
},
"_bibtex": ... | |
2,026 | 0pCAQoNE5E | [
4,
2,
6,
4
] | [
{
"content": "The paper proposes an interaction-aware 4D Gaussian splatting method for dynamic hand–object interaction reconstruction from RGB video without object shape priors. It decomposes the scene into hand, object, and background fields; conditions the object field on the hand; augments Gaussians with lea... | {
"cdate": 1758024048345,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\ntian2025interactionaware,\ntitle={Interaction-Aware 4D Gaussian Splatting for Dynamic Hand-Object Interaction Reconstruction},\nauthor={Hao Tian and Chenyangguang Zhang and Rui Liu and Wen Shen and Xiaolin Qin},\nyear=... | |
2,026 | 0pFcKF2li1 | [
4,
2,
2
] | [
{
"content": "This paper proposes Sandbox-RL, a reinforcement learning framework for optimizing multiple large language models through structured sandbox environments organized as workflow graphs. The method enables heterogeneous models to co-train under controlled cooperation and competition, supported by a sc... | {
"cdate": 1757328486987,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nliu2025sandboxrl,\ntitle={Sandbox-{RL}: Scalable Multi-{LLM}s Optimization through Sandbox-Based Reinforcement Learning},\nauthor={Dong Liu and Yanxuan Yu and Ying Nian Wu and Xuhong Wang},\nyear={2025},\nurl={https://... | |
2,026 | 0pGKVsri2B | [
4,
4,
4
] | [
{
"content": "While existing graph generation methods can achieve realistic results, the generated graphs lack domain-specific patterns. This paper proposes a method that uses first-order logic to enhance VGAE, making the domain-specific patterns in VGAE-generated graphs more closely match those in the original... | {
"cdate": 1758252093936,
"content": {
"TLDR": {
"value": "Shows how deep graph generation can be enhanced with domain knowledge represented by first-order logic rules with a novel semantic loss function"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025enhancing,\ntitle={Enhancing Gr... | |
2,026 | 0pRvnSBNGx | [
6,
4,
4,
4
] | [
{
"content": "Tackling hallucinations of code generation models with automatically generated unit tests using dynamic code analysis tools. They propose SCG to abstain from uncertain generations. Defines a probabilistic notion of correctness called α-code entailment, leveraging dynamic analysis tools to approxim... | {
"cdate": 1758244406762,
"content": {
"TLDR": {
"value": "We propose a learning algorithm for selective code generation that controls the rate of hallucination by exploiting a code analysis method, called fuzzing, to generate unit tests for a code correctness measure."
},
"_bibtex": {
"valu... | |
2,026 | 0pVKknV9nM | [
4,
4,
6,
2
] | [
{
"content": "Large language models (LLMs) exhibit advanced text generation abilities, underscoring the need for reliable detection to mitigate misuse. However, existing zero-shot detectors struggle with two key issues: style imitation (SIC), where LLMs successfully mimic human writing styles, and content inter... | {
"cdate": 1758122596332,
"content": {
"TLDR": {
"value": "We propose a novel framework termed SaFT to improve LLM-generated text detection accuracy by spotting style imitation and filtering content interference."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025saft,\ntitle={Sa{FT}: ... | |
2,026 | 0psy3EVkT7 | [
4,
2,
6
] | [
{
"content": "The paper proposes a self-supervised way of learning discrete action codes through predicting how point clouds evolve during manipulation rather than reconstructing visual observations. The paper argues that the spatial geometry changing through time is important and useful to spatial understandin... | {
"cdate": 1757034779635,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025geomola,\ntitle={GeoMoLa: Geometry-Aware Motion Latents for Learning Robust Manipulation Policies},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learnin... | |
2,026 | 0pw5Qmfynp | [
2,
2,
4,
2
] | [
{
"content": "In this paper, the authors examine \"regret\" behavior in large language models from a mechanistic interpretability perspective. The work centers on defining and applying several metrics to identify neurons associated with regret expression. The authors propose a Supervised Compression-Decoupling ... | {
"cdate": 1757594199403,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\ncui2025compositional,\ntitle={Compositional Architecture of Regret in Large Language Models},\nauthor={Xiangxiang Cui and Shu Yang and Tianjin Huang and Wanyu Lin and Lijie Hu and Di Wang},\nyear={2025},\nurl={https://... | |
2,026 | 0qFN7Ugt4k | [
4,
8,
4,
2
] | [
{
"content": "This paper investigates an interesting and important problem, namely multi-source-free domain adaptation for object detection. The authors propose a new framework that effectively aggregates knowledge from multiple sources and mitigates distribution discrepancies across domains. Experimental resul... | {
"cdate": 1758332408059,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025multisource,\ntitle={Multi-Source Knowledge-Fusion for Source-Free Domain Adaptation in Object Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on ... | |
2,026 | 0qVrS2XdOW | [] | [] | {
"cdate": 1757754362701,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nzhang2025brainintheloop,\ntitle={Brain-in-the-Loop Generation: Test-Time Scaling of {EEG} Signals to Steer Large Language Models},\nauthor={Junzi Zhang and Yue Yao and Jianing Shen and Yi Zhang and Hailin Zhang and Tom... | |
2,026 | 0qVu2WsDle | [
4,
6,
4,
4
] | [
{
"content": "This paper looks at \"dataset condensation,\" which is the task of shrinking a giant dataset (like ImageNet) into a tiny set of synthetic images that can be used to train a model just as well. The authors' key insight is that everyone has been focused on shrinking the number of images (e.g., 10 im... | {
"cdate": 1758210412232,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025toward,\ntitle={Toward Bit-Efficient Dataset Condensation: A General Framework},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},... | |
2,026 | 0qgcZvtQx0 | [
6,
0,
2,
0
] | [
{
"content": "The paper describes an approach to certifying the faithfulness of a given Chain of Thought used to perform a numerical calculation, by translating a given CoT into a computational flowgraph of type operators in a functional language. Metrics are defined that are used to determine if the analyzed C... | {
"cdate": 1758257116294,
"content": {
"TLDR": {
"value": "We treat Chain-of-Thought as a formal proof using the Curry-Howard correspondence to mechanistically verify its faithfulness"
},
"_bibtex": {
"value": "@misc{\nperrier2025typed,\ntitle={Typed Chain-of-Thought: A Curry-Howard Framewor... | |
2,026 | 0qrPON6rIN | [
0,
2,
4,
2
] | [
{
"content": "This paper introduces Identity-Projection, a geometric hypothesis that semantic concepts in transformers persist as invariant prototype directions, and evince themselves to different degrees in model representations across contexts depending on the semantic relevance. The paper develops two method... | {
"cdate": 1758354711278,
"content": {
"TLDR": {
"value": "The paper introduces identity-projection and Head2Feat, unsupervised methods for analyzing and steering transformer-based language models by identifying and aligning influential attention heads with semantic features."
},
"_bibtex": {
... | |
2,026 | 0r1KU3dlps | [
8,
8,
4,
6
] | [
{
"content": "The paper proposes two algorithms for solving the “egalitarian”/“group DRO objective” of multi-group regression: one algorithm for the “robust” version of it—i.e., the maximum over each group’s squared prediction error—and one which interpolates from the “robust” version to the fully “nonrobust” v... | {
"cdate": 1758213659886,
"content": {
"TLDR": {
"value": "We give algorithms for optimizing a distributionally robust/multidistributional loss for least squares linear regression."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025distributionally,\ntitle={Distributionally Robust Line... | |
2,026 | 0rHEudxV8K | [
4,
2,
6,
6
] | [
{
"content": "The paper addresses text-based image editing with pre-trained rectified-flow models and argues that the usual corruption-then-restoration pipeline constructs a target-agnostic intermediate state, which limits editability and/or source consistency when the target edit departs from the source image.... | {
"cdate": 1757592045883,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nwang2025flowcycle,\ntitle={FlowCycle: Pursuing Cycle-Consistent Flows for Text-based Editing},\nauthor={Yanghao Wang and Zhen Wang and Long Chen},\nyear={2025},\nurl={https://openreview.net/forum?id=0rHEudxV8K}\n}"
... | |
2,026 | 0rJUulYnow | [
6,
6,
4,
2
] | [
{
"content": "The paper presents **EvoMAS**, an evolutionary framework that formulates multi-agent workflow automation as a constrained single-objective optimization problem. It models the “variation–selection–reflection” cycle as a non-homogeneous Markov process and employs a meta-controller, **Cyber Creator**... | {
"cdate": 1758284365231,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025evomas,\ntitle={Evo{MAS} : Heuristics in the Loop{\\textemdash}Evolving Smarter Agentic Workflows},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learnin... | |
2,026 | 0rlc34xAhz | [
2,
2,
4,
6
] | [
{
"content": "In many classification tasks, the target label is not just a single value from the output space but rather a collection of values. In such “multi-label” classification problems, there is no way yet to directly leverage ideas from conformal prediction to enable distribution-free uncertainty estimat... | {
"cdate": 1758280429231,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025efficient,\ntitle={Efficient Multilabel Uncertainty Quantification with Conformal Ensembles},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Repr... | |
2,026 | 0sCyk9Tr5J | [
4,
8,
6,
8
] | [
{
"content": "This paper introduces semantic calibration for LLM, where a model’s probability over semantic answer classes matches its empirical accuracy. It proposes a theoretical analysis to link calibration and local loss optimality. Experiments show that instruction tuning and Chain-of-thought (CoT) hurt LL... | {
"cdate": 1758069984795,
"content": {
"TLDR": {
"value": "We show that LLMs can be semantically calibrated, and we develop theory for when and why."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025trained,\ntitle={Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic ... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.