paper_id
stringlengths
10
10
review
stringlengths
0
16.9k
scores
listlengths
2
6
avg_score
float64
1
10
04RGjODVj3
Now I have all the information I need. Let me write the final consolidated review. ## Summary This paper proposes HyperEEGNet, which combines a hypernetwork with EEGNet to generate adaptive weights for motor imagery (MI) classification. The hypernetwork takes connectivity features (COH/PLV) extracted from pre-stimulu...
[ 1, 5, 3, 3 ]
3
04qx93Viwj
Now I have all the information I need. Let me construct the final consolidated review. --- ## Summary This paper provides an empirical measurement of the environmental impact (carbon emissions, water consumption, and embodied manufacturing costs) of developing and training a series of LLMs from 20M to 7B parameters....
[ 8, 8, 6 ]
7.33
07yvxWDSla
Now I have all the information I need. Let me compose the final consolidated review. ## Summary This paper proposes **synthetic continued pretraining**: using a small domain-specific corpus (~1.3M tokens) to synthesize a much larger corpus via LLM prompting, then performing continued pretraining on the synthetic data...
[ 8, 8, 8, 8 ]
8
0CvJYiOo2b
Here is my final consolidated review. --- ## Summary This paper proposes using PCA as a preprocessing step to reduce the temporal dimension of time series data before feeding it into deep-learning models for classification, forecasting, and regression. The core claim is that PCA-based temporal compression maintains ...
[ 5, 5, 3, 5 ]
4.5
0F1rIKppTf
Now I have all the evidence I need. Let me write the consolidated review. ## Summary This paper studies the "mirror Schrödinger bridge" — the Schrödinger bridge problem from a probability distribution to itself — and proposes an alternating minimization procedure (AMP) to solve it. The key algorithmic insight is that...
[ 6, 5, 6, 6 ]
5.75
0JwxMqKGxa
Now let me write the consolidated final review. ## Summary This paper proposes using reinforcement learning on procedurally generated synthetic environments (NavIndoor) to train navigation features for sensory substitution devices for visually impaired individuals. The authors train a compact CNN via Q-learning on se...
[ 6, 5, 1, 3, 3, 1 ]
3.17
0Q1mBvUgmt
I have thoroughly verified the reviewer's claims against the paper. Let me now produce the final consolidated review. --- ## Summary VIPER proposes a time series forecasting framework that decomposes the input into trend and seasonal components, applies multi-period 2D transformations (FFT-based period extraction fo...
[ 3, 3, 3 ]
3
0T8vCKa7yu
Now I have all the information I need. Let me construct the final consolidated review. ## Summary This paper proposes CVXQ, a framework for post-training weight quantization of LLMs that formulates the mixed-precision bit allocation problem as a constrained least-squares problem and solves it via dual ascent, using r...
[ 3, 3, 3, 3 ]
3
0aTIvSJ83I
## Summary This paper proposes Agnostic-SAM, a variant of Sharpness-Aware Minimization (SAM) that adds a validation-gradient correction to the perturbation step. The training perturbation is modified by subtracting (a momentum-smoothed version of) the validation gradient direction, which the Taylor expansion reveals e...
[ 3, 3, 3, 3 ]
3
0bmGL4q7vJ
This is the final consolidated review. --- ## Summary This paper proposes a pipeline for synthesizing multi-modal tool-usage training data (MM-Traj, 20K tasks with trajectories across 10 tools) and uses it to fine-tune VLMs (MiniCPM-V-8.5B, Qwen2-VL-7B) into a tool-using agent (T3-Agent). The pipeline inverts the ty...
[ 6, 8, 8, 8 ]
7.5
0ctvBgKFgc
Here is the final consolidated review. ## Summary This paper presents ProtComposer, a method for conditioning protein backbone generation on 3D ellipsoid layouts annotated with secondary structure type and residue counts. It fine-tunes the Multiflow flow-matching model with an SE(3)-equivariant Invariant Cross Attent...
[ 8, 8, 8, 8 ]
8
0fpLLsAynh
Now I have all the information needed. Let me compose the final consolidated review. ## Summary This paper proposes DSpodFL (Decentralized Sporadic Federated Learning), which introduces probabilistic skipping of both SGD computations and model exchanges in decentralized FL. Devices independently decide whether to com...
[ 5, 3, 3 ]
3.67
0iAZYF9hrl
Now let me write the final consolidated review. ## Summary This paper proposes a Disentangled Representation Learning (DRL) methodology for microscopy image classification. The method uses Ada-GVAE trained on a synthetic source dataset (TexturedSprite) with DINO ViT features as input, then transfers the disentangled ...
[ 3, 3, 1, 3 ]
2.5
0iscEAo2xB
Now I have all the information needed. Let me construct the final consolidated review. ## Summary This paper empirically compares risk-based targeting (treating individuals with highest predicted baseline risk) against CATE-based targeting (treating those with highest estimated treatment effects) across 5 diverse RCT...
[ 6, 5, 6, 10 ]
6.75
0mdUV1pLGP
## Summary This paper proposes a neural Hawkes process variant (ENHP) that replaces the traditional parametric exponential-decay impact function with a neural-network-based impact kernel operating in a learned embedding space. The embedding formulation reduces the kernel output dimension from O(M²) to O(D²) (D << M), ...
[ 3, 3, 5, 3 ]
3.5
0pLCDJVVRD
Now let me prepare the final consolidated review. ## Summary This paper proposes a phenomenological definition of emergence in neural networks (requiring discontinuous improvement, simultaneity across tasks, and correlation with learning of data regularities), constructs a controlled formal-language experimental syst...
[ 6, 6, 8, 8 ]
7
0rS9o1uKqu
Now I have all the information I need. Let me write the final consolidated review. ## Summary This paper proposes TLDR (Training-Like Data Reconstruction), a method that uses a conditioned generator to invert CNN classifiers and reconstruct images that semantically resemble the training data. The approach combines ne...
[ 3, 3, 3, 1 ]
2.5
0zGvf2yRMQ
Now I have all the information needed. Let me produce the final consolidated review. --- ## Summary MeshGen proposes a pipeline for generating PBR-textured 3D meshes from a single image. It introduces three main technical components: (1) a render-enhanced point-to-shape auto-encoder that uses perceptual render loss ...
[ 3, 5, 6, 5, 6 ]
5
0ziGSo4uWp
Now I have all the information I need. Let me compile the final review and anchor list.
[ 3, 5, 3 ]
3.67
10JOlFIPjt
Now let me write the consolidated final review, carefully verifying each claim against the paper. ## Summary This paper introduces NEMO (Neuronal Embeddings via MultimOdal Contrastive Learning), a CLIP-style contrastive learning framework that jointly embeds extracellular waveforms and activity autocorrelograms (ACGs...
[ 8, 8, 8, 6 ]
7.5
11xgiMEI5o
Now I have all the information I need to write the final consolidated review. Let me construct it carefully. ## Summary The paper proposes OmniRe, a 3D Gaussian Splatting framework with a dynamic neural scene graph that models static backgrounds, rigid vehicles, and—critically—non-rigid actors (pedestrians, cyclists)...
[ 6, 8, 8 ]
7.33
12iSWNLDzj
Now let me write the final consolidated review. ## Summary This paper proposes DAFR (Diffusion Attack against Facial Recognition), a method that uses diffusion-based adversarial guidance (from AdvDiff) combined with 3D rendering to generate stealthy adversarial face masks with text-promptable style control (e.g., "pu...
[ 3, 3, 3, 3 ]
3
14rn7HpKVk
Now I have all the information needed. Here is the final consolidated review: --- ## Summary SALMONN is an end-to-end multimodal LLM that integrates a Vicuna-13B backbone with dual auditory encoders (Whisper for speech, BEATs for audio events/music) via a window-level Q-Former and LoRA adaptors. It identifies a "tas...
[ 6, 8, 6 ]
6.67
15lk4nBXYb
I have all the information I need. Let me produce the final consolidated review. ## Summary This paper introduces CCM-DiT, a method for adding camera-pose controllability to DiT-based video generators. It proposes a Sparse Motion Encoding Module that converts dense pixel-wise Plücker coordinate fields into a regular-...
[ 3, 3, 3, 3 ]
3
1DEEVAl5QX
Here is my final consolidated review. --- ## Summary This paper introduces the first mini-batch algorithm for maximizing non-negative monotone decomposable submodular functions. It shows that resampling a small sparsifier at each greedy iteration (mini-batch) improves over the prior sparsifier approach (sampling onc...
[ 3, 6, 5 ]
4.67
1FY1apsMxc
Now I have a thorough understanding of the paper and have verified all the reviewer's claims against the actual paper content. Let me produce the final consolidated review. --- ## Summary PromptGFM proposes a graph foundation model with two modules: (1) a **Graph Understanding Module** that uses GPT-3.5 prompts to r...
[ 5, 3, 3, 3 ]
3.5
1MHgMGoqsH
Now I have all the information I need. Let me write the authoritative final review. ## Summary This paper introduces a Model Predictive Control (MPC) framework for training deep neural networks, unifying Back-Propagation (horizon *h=T*) and Forward-Forward (horizon *h=1*) as two extremes of a spectrum of look-forward ...
[ 3, 3, 3 ]
3
1Nwsqw0sTm
Now I have all the information I need. Let me write the final consolidated review. ## Summary This paper proposes VOCAL, an open-vocabulary object detection (OVDet) approach that constructs classifiers from text descriptions (generated by GPT-3), visual exemplars (with a prototype-discovery mechanism), and a multi-mo...
[ 5, 6, 5 ]
5.33
1Ogw1SHY3p
Now I have a thorough understanding of the paper. Let me produce the final consolidated review. --- ## Summary This paper introduces Monet, a Mixture-of-Experts architecture that scales the number of experts to 262K per layer via a novel product-key composition method (horizontal and vertical decomposition), achievi...
[ 8, 8, 6, 6 ]
7
1P6AqR6xkF
Now I have all the information needed. Let me construct the final consolidated review. ## Summary This paper introduces ACID, a large-scale dataset for AI-created image detection comprising ~13M images (4.1M real-world, 8.9M AI-generated) sourced from over 50 generative models. It also proposes ACIDNet, a two-branch ...
[ 6, 3, 5, 3 ]
4.25
1PPjf4wife
Now let me write the final consolidated review. ## Summary This paper introduces FAMA (Functionally-Aligned Multi-Agents), a framework that uses a single LLM as a shared backbone for multiple agents in cooperative MARL, combined with a centralized MAPPO critic and discrete natural-language inter-agent communication. ...
[ 5, 3, 6, 3 ]
4.25
1PXEY7ofFX
Here is the final consolidated review. --- ## Summary This paper introduces "Bespoke solvers," a framework for learning custom ODE solvers tailored to individual pre-trained generative flow models. The core idea is to parameterize a family of scale-time transformations over sampling paths (using a tiny number of lea...
[ 6, 8, 8, 8, 6 ]
7.2
1TXDtnDIsV
Now I have all the verification I need. Let me finalize the review. --- ## Summary This paper investigates whether Mamba (an attention-free selective state space model) can serve as an effective continual learner under the meta-continual learning (MCL) framework, where learning is cast as sequence prediction. The au...
[ 3, 6, 5 ]
4.67
1V1QQYARmd
Now I have all the evidence needed. Let me write the consolidated review. --- ## Summary This paper proposes an unsupervised OOD detection method that uses the k-NN radius on a neural network's intermediate embeddings, normalized by the expected in-distribution radius estimated via cross-validation. The key insight ...
[ 8, 5, 6, 3 ]
5.5
1X1R7P6yzt
Good. I have all the information needed. Here is the final consolidated review. --- ## Summary DGPPO proposes a framework that simultaneously learns a discrete graph CBF (DGCBF) via constraint-value functions and a safe distributed policy for multi-agent systems under unknown discrete-time dynamics, partial observab...
[ 8, 6, 6 ]
6.67
1hLFLNu4uy
Now I have a thorough understanding of the paper and all the reviewer's claims. Let me compose the final consolidated review. ## Summary PORTIA is a prompt-rewriting system that mitigates position bias in LLM-based pairwise evaluators by splitting candidate answers into aligned segments, then merging corresponding se...
[ 6, 6, 3, 5 ]
5
1iuaxjssVp
Now I have all the information needed. Let me produce the final consolidated review. ## Summary This paper introduces InvMSAFold, an inverse folding method that uses a neural network to generate the parameters of a low-rank pairwise (Potts) model conditioned on a protein structure, rather than predicting sequences au...
[ 8, 8, 5, 8 ]
7.25
1k4yZbbDqX
Now I have all the information needed. Let me produce the final consolidated review. ## Summary This paper introduces InstaFlow, a one-step text-to-image generator derived from Stable Diffusion (SD) via a text-conditioned Rectified Flow pipeline. The core idea is that direct distillation of SD fails because its proba...
[ 6, 6, 8, 8 ]
7
1nHQRsb3Ze
Now I have all the information needed. Let me produce the final consolidated review. ## Summary This paper analyzes the stability of intermediate neural network layers in continual learning and proposes using auxiliary classifiers (ACs) — additional classifiers attached to intermediate layers — to leverage this stabi...
[ 5, 5, 5 ]
5
1qzUPE5QDZ
Now I have all the information needed. Let me write the consolidated review. ## Summary This paper identifies "group irregularities" as a pathology in existing shift explanation methods (K-cluster transport, optimal transport, DiCE) where dataset-level explanations can break subpopulation structure by mapping source ...
[ 6, 5, 5, 5 ]
5.25
1vDArHJ68h
Now let me write the final consolidated review. ## Summary This paper introduces Recall to Imagine (R2I), which replaces the GRU in DreamerV3's RSSM world model with a Structured State-Space Model (S3M, based on S4/S5 with parallel scan). The goal is to improve the agent's ability to handle long-term dependencies in ...
[ 1, 8, 6 ]
5
28TLorTMnP
Now I have all the verification I need. Let me write the final consolidated review. --- ## Summary This paper proposes Soft Preference Optimization (SPO) and SPO-abs, two algorithms for aligning LLMs using explicit scalar-reward datasets (lists of responses with scalar scores) as well as pairwise preference data. SP...
[ 3, 3, 3, 1 ]
2.5
293V3bJbmE
Here is the final consolidated review. --- ## Summary This paper presents HELMET, a benchmark for evaluating long-context language models (LCLMs) across seven application-centric categories (RAG, generation with citations, passage re-ranking, long-document QA, summarization, many-shot ICL, and synthetic recall) at c...
[ 6, 6, 6, 6 ]
6
2DD4AXOAZ8
Now I have all the information I need. Let me write the final consolidated review. ## Summary This paper empirically evaluates MixAttention, an architectural modification combining sliding window attention with KV cache sharing across layers (closely related to a Character.AI blog post). Using 24-layer MoE models (1....
[ 1, 3, 3, 1 ]
2
2GcR9bO620
Now I have all the information I need. Let me produce the final consolidated review. ## Summary This paper introduces DeepFakeVox-HQ, a large-scale deepfake audio dataset (1.3M samples, 40 vocoders), and F-SAT (Frequency-Selective Adversarial Training), which applies adversarial training only in the high-frequency ran...
[ 8, 8, 6, 6 ]
7
2LhCPowI6i
Now I have all the information needed. Let me produce the final consolidated review. --- ## Summary This paper proposes filtering generated replay samples (pseudodata) by classifier confidence to mitigate error propagation in the generator during generative replay for class-incremental learning. The method is simple...
[ 3, 3, 1 ]
2.33
2aebB2mf0q
Now let me write the final consolidated review. --- ## Summary This paper proposes SemiAugIR, a semi-supervised framework for single-frame infrared small target (SIRST) detection. The method has three main components: (1) non-uniform chromaticity augmentation (NUC), which generates smooth random perturbation maps in...
[ 3, 5, 1 ]
3
2dnO3LLiJ1
Now let me produce the final consolidated review. ## Summary This paper identifies and characterizes a real phenomenon in modern Vision Transformers (ViTs): high-norm outlier tokens that appear in feature maps during inference, predominantly in low-informative background patches. The authors show that these tokens co...
[ 8, 8, 8, 8 ]
8
2fSyBPBfBs
Now I'll write the final consolidated review. --- ## Summary This paper studies bilevel optimization (BLO) when the lower-level function lacks strong convexity (LLSC), a regime that is poorly understood theoretically despite being common in practice. The paper provides: (i) hardness results showing BLO without LLSC ...
[ 3, 5, 3, 3, 6, 5 ]
4.17
2fojNANZSv
Now let me write the final consolidated review. ## Summary This paper proposes MixturePFN, a framework to scale Prior-Fitted Networks (specifically TabPFN) to larger tabular datasets (beyond the ~3,000-sample limit). The method has two components: (1) MICP (Mixture of In-Context Prompters), which routes test samples ...
[ 6, 8, 6 ]
6.67
2mqb8bPHeb
Now I have thoroughly verified every claim. Let me construct the final consolidated review. ## Summary This paper proposes T-Stitch, a training-free technique for accelerating diffusion model sampling by using a smaller model for early denoising steps and switching to a larger model for later steps. The key insight —...
[ 8, 8, 6, 6 ]
7
2oWRumm67L
Now I have all the information I need. Let me produce the final consolidated review. --- ## Summary This paper proposes Light-MILPopt, a four-stage ML-guided framework for solving large-scale Mixed Integer Linear Programs (MILPs). The pipeline: (1) FENNEL graph partition to divide the MILP bipartite graph into small...
[ 6, 6, 5, 3 ]
5
2uPZ4aX1VV
Now I have all the information I need. Let me compose the final consolidated review. ## Summary This paper introduces NCII (Null Counterfactual Interaction Inference) and HInt (Hindsight Relabeling using Interactions) for goal-conditioned RL. NCII defines interactions between state factors via a null-counterfactual t...
[ 6, 5, 8, 8 ]
6.75
2yqAzFPT4F
Now I have all the verification I need. Let me produce the final consolidated review. --- ## Summary This paper proposes Zer0-Jack, a method for jailbreaking black-box Multi-modal Large Language Models (MLLMs) using zeroth-order optimization (SPSA) with patch coordinate descent (SPSA-P) to generate adversarial image...
[ 5, 3, 5, 8 ]
5.25
381rZinzJE
Now I have all the information needed. Let me produce the final consolidated review. ## Summary This paper proposes a Physics-Informed Autoencoder (PIAE) that integrates established biophysical NEE models (Lloyd-Taylor respiration, Michaelis-Menten GPP light-response) into a stochastic differential equation (SDE) fra...
[ 6, 5, 5, 3 ]
4.75
3EeyQNgKTP
## Summary This paper proposes TCTO, a framework for automated feature transformation that maintains an evolving transformation "roadmap" (a graph tracking the history of feature transformations) and uses three cascading RL agents (head-cluster, operation, operand-cluster) guided by an RGCN-based state representation ...
[ 5, 6, 6 ]
5.67
3GTtZFiajM
Now I have everything I need. Let me write the consolidated review. --- ## Summary This paper introduces **Calm**, a framework for automatically quantifying 12 types of bias in LLM-as-a-Judge systems. The core idea is an "attack-and-detect" approach: a perturbation generator (GPT-4o) modifies response content or ins...
[ 6, 5, 8, 8 ]
6.75
3IFRygQKGL
Now I have all the information needed. Let me write the final consolidated review. ## Summary This paper proposes OptionZero, an extension of MuZero that integrates autonomous option discovery into the learned-dynamics planning framework. The key ideas are: (1) an option network that predicts a "dominant option" (the...
[ 8, 8, 6, 8 ]
7.5
3Vw7DQqq7U
Now I have all the information needed. Let me produce the final consolidated review. ## Summary LEMON proposes a lossless model expansion method for Transformers that handles indivisible width/depth increments. The core technical ideas are (i) breaking symmetry of replicated neurons by using unequal fan-out weights (...
[ 6, 6, 8, 8 ]
7
3ZdGSTxKuy
Now let me write the final consolidated review. ## Summary This paper introduces a dataset of 5,486 "atypical" videos (sci-fi, animation, unintentional actions, and abnormal/ surveillance events) and studies whether using this data as outlier exposure (OE) during fine-tuning improves out-of-distribution (OOD) detecti...
[ 3, 3, 1, 1 ]
2
3f8556SIEn
Now I have all the information needed. Let me write the consolidated final review. --- ## Summary This paper introduces MEDIC, a training-free framework for zero-shot music editing that combines two innovations: (1) Harmonized Attention Control (HAC), which adds a "Harmonic Branch" between the standard source and ta...
[ 5, 5, 3, 3 ]
4
3fl1SENSYO
Now I have all the information I need. Let me compose the final consolidated review. ## Summary 2-3 sentence summary of the paper's contribution. This paper proposes DiffPutter, an iterative missing-data imputation method that couples the Expectation-Maximization (EM) algorithm with diffusion models. In each iteratio...
[ 6, 8, 8, 8 ]
7.5
3qo1pJHabg
Now I have all the evidence I need. Let me write the final consolidated review. ## Summary This paper proposes LRR (Language-Driven Resamplable Continuous Representation), a preprocessing-based defense against adversarial attacks on visual object trackers. It consists of two modules: (1) a Spatial-Temporal Implicit R...
[ 6, 6, 5 ]
5.67
473sH8qki8
Now let me compose the final consolidated review. ## Summary This paper proposes a reward-based policy that conditions solely on the history of rewards and actions (disregarding observations), enabling zero-shot transfer across environments with the same dynamics but arbitrarily different observation spaces. The auth...
[ 1, 1, 3, 3 ]
2
49qqV4NTdy
Now I have a thorough understanding of the paper. Let me construct the final consolidated review. ## Summary This paper presents a systematic, controlled study of preference alignment for Multimodal LLMs (MLLMs), disentangling confounded variables (base model, dataset construction, alignment algorithm, dataset size) ...
[ 8, 6, 6 ]
6.67
49ti6LOUw5
Now let me write the consolidated review. --- ## Summary This paper proposes UnoLoRA, a method that uses a single shared LoRA adapter (rather than per-task adapters) for multi-task fine-tuning of LLMs, and UnoLoRA*, an enhanced variant with a shared hypernetwork that generates task-specific scaling vectors. The core...
[ 3, 3, 3, 3 ]
3
4AlNpszv66
Now I have all the information needed. Let me compile the final consolidated review. ## Summary This paper introduces Feedback Controllability Components Analysis (FCCA), a dimensionality reduction method that identifies subspaces of neural population dynamics that maximize a measure of feedback controllability (the ...
[ 8, 5, 3, 3 ]
4.75
4F1a8nNFGK
Now I have all the information. Let me write the final consolidated review. ## Summary This paper introduces "Context is Key" (CiK), a time-series forecasting benchmark designed so that the accompanying natural-language context is *essential* — not merely helpful — for accurate forecasts. It contributes a taxonomy of ...
[ 3, 6, 6, 5 ]
5
4Hf5pbk74h
Now I have all the information I need to produce the consolidated review. ## Summary This paper proposes LaSeNN (LAtent-SElf-kNN), a post-hoc inference-time method that averages a classifier's prediction with the predictions of k-nearest neighbors found in latent space (layer activations). The method requires no retr...
[ 1, 3, 3 ]
2.33
4KqkizXgXU
Now I have everything I need. Here is the consolidated final review. --- ## Summary The paper proposes Curiosity-driven Red Teaming (CRT), which augments the standard RL-based red teaming objective for LLMs with an entropy bonus and two novelty rewards (based on SelfBLEU and cosine similarity) to improve the explora...
[ 8, 8, 8, 8 ]
8
4MWUdp6deL
## Summary This paper introduces CodeFavor, a framework for training pairwise code preference models from synthetic evolution data derived from two sources: code commits (CI) converting pre/post-commit diffs into preference pairs, and code critiques (CE) pairing draft-model outputs with critic-model revisions. To eval...
[ 5, 6, 5, 6 ]
5.5
4UiLqimGm5
Now I have all the information needed. Let me compose the final consolidated review. ## Summary This paper proposes Coordinate-Aware Modulation (CAM), an architectural module that injects grid-derived scale/shift modulation parameters into intermediate MLP features for neural fields. Unlike the dominant sequential pa...
[ 6, 6, 8, 8 ]
7
4UxXe3JZta
Now I have verified all claims against the actual paper. Here is the consolidated final review. --- ## Summary HRVMamba combines Mamba-based state space model blocks with HRNet's multi-resolution parallel architecture for dense prediction tasks. The core technical contribution is the Dynamic Visual State Space (DVSS...
[ 5, 5, 5, 3 ]
4.5
4YESQqIys7
Now I'll produce the final review. ## Summary This paper proposes NfgTransformer, a permutation-equivariant transformer architecture for representing normal-form games (NFGs) as action embeddings. The architecture uses a sequence of self- and cross-attention operations to iteratively refine action embeddings from zer...
[ 3, 8, 5, 8 ]
6
4ZX2a3OKEV
Now I have all the information needed. Let me construct the final consolidated review. ## Summary This paper extends surrogate-loss techniques — previously limited to scalar minimization — to variational inequality (VI) problems with hidden monotone structure, which includes min-max optimization and projected Bellman...
[ 8, 8, 6, 6 ]
7
4aywmeb97I
Now let me write the consolidated review. ## Summary The paper studies how asynchronous delay and data heterogeneity jointly degrade convergence in asynchronous federated learning. It proposes CA²FL (Cache-Aided Asynchronous Federated Learning), where the server caches each client's latest update and reuses it to cal...
[ 8, 8, 6 ]
7.33
4bOCP1GtX4
Now I have all the information needed. Let me write the consolidated review. ## Summary This paper proposes WenXinGPT, a decoder-only language model (7B parameters) designed for orthopedic expert consultations in Chinese medical contexts. It is trained on data from Peking Union Medical College Hospital spanning 16 ca...
[ 3, 3, 3 ]
3
4dtwyV7XyW
## Summary This paper proposes Knowledge Tracing Set Transformers (KTSTs), a transformer-based architecture for knowledge tracing that centers on (i) a learnable variant of ALiBi (attention with linear biases) for positional information, (ii) permutation-invariant set representations of student interactions that addre...
[ 5, 1, 3, 3 ]
3
4h1apFjO99
Now I have everything I need. Let me write the final consolidated review. ## Summary This paper proposes Diffusion-TS, a diffusion-based generative model for multivariate time series that incorporates seasonal-trend decomposition into the denoising network architecture. The model predicts the clean signal directly (r...
[ 8, 5, 6 ]
6.33
4iFSBgxvIO
Now let me produce the final consolidated review. ## Summary This paper studies multi-LoRA composition in text-to-image diffusion models. It first provides a Fourier-domain analysis showing that different LoRA categories (Character, Style, Background, etc.) contribute differently to high-frequency vs. low-frequency c...
[ 6, 5, 5 ]
5.33
4l3AH8Bhmt
Now I have all the information needed to verify each claim. Let me compose the final review. ## Summary This paper identifies and addresses "Specificity Failure" in knowledge-edited LLMs—where the presence of an edited subject in context corrupts the model's predictions for unrelated entities. The authors diagnose th...
[ 8, 6, 8, 6 ]
7
4mqt6QxSUO
Now let me write the consolidated review. ## Summary This paper claims to present a unified Riemannian-geometric framework for SARS-CoV-2 detection from CT scans, integrating submodular optimization, Fisher Information Metric-based feature integration, proper scoring rules for decision-making, and adversarial domain ...
[ 6, 1, 3, 3 ]
3.25
50RNY6uM2Q
Now I have all the information needed. Let me construct the final consolidated review. ## Summary MG-LLaVA extends the LLaVA architecture with a multi-granularity vision flow that processes low-resolution (ViT), high-resolution (ConvNeXt), and object-level (RoI from external detector) visual features jointly. A Conv-...
[ 5, 3, 3, 5, 5 ]
4.2
51cjeYcXjs
Now let me write the final consolidated review. ## Summary This paper proposes DDG Fingerprinting, a method for representing binary executable programs as sets of hashed data dependency graphs (DDGs) to enable search and retrieval by behavioral similarity. The pipeline segments programs into basic blocks, extracts DD...
[ 3, 1, 3, 3 ]
2.5
52x04chyQs
I have verified all claims against the paper. Here is the final consolidated review. --- ## Summary This paper provides a theoretical characterization of the expressiveness of invariant geometric deep learning models for point clouds. It introduces an $\mathcal{A}$-symmetry framework to rigorously bound DisGNN's uni...
[ 6, 6, 6, 6 ]
6
55oi1LCdDL
Now I have all the information I need. Let me write the consolidated review. ## Summary This paper proposes Duct (DUal ConsolidaTion), a method for Domain-Incremental Learning (DIL) with pre-trained Vision Transformers that addresses forgetting at two levels: representation-level forgetting via task-vector merging of...
[ 6, 3, 6, 5, 5 ]
5
5AJ8R4z5g0
Now I have all the evidence I need. Let me produce the final consolidated review. --- ## Summary This paper addresses CATE estimation under hidden confounders by leveraging outcome-only RCT data (no covariates from the RCT, only outcomes) to regularize an observational CATE estimator. Two regularization modules are ...
[ 3, 6, 1, 3 ]
3.25
5AtlfHYCPa
Now I have all the information needed. Let me produce the consolidated review. ## Summary This paper introduces HR-Extreme, a high-resolution (3km) dataset for evaluating extreme weather forecasting, built from NOAA's HRRR operational data and covering 17 types of extreme weather events across the US for the year 202...
[ 5, 8, 6, 8 ]
6.75
5BXAXOpaWu
Now let me write the final consolidated review. --- ## Summary This paper proposes ISA (Image2Sentence based Asymmetric zero-shot composed image retrieval), a framework that addresses two limitations of existing ZSCIR methods: (1) reliance on symmetric heavy-model deployments impractical for mobile/edge devices, and...
[ 6, 8, 8, 8 ]
7.5
5Dwqu5urzs
Now I have all the information needed. Here is the final consolidated review. --- ## Summary This paper proposes Phy-DRL, a framework combining three architectural innovations — residual action policy (model-based + data-driven DRL), a safety-embedded reward derived from LMI-based Lyapunov reasoning, and physics-know...
[ 6, 8, 8, 8 ]
7.5
5IkDAfabuo
Now I have all the information. Here is the final consolidated review. --- ## Summary This paper proposes Prioritized Generative Replay (PGR), a framework that replaces a standard replay buffer with a conditional diffusion model trained online. The generative model densifies experience by creating new transitions, a...
[ 8, 8, 6, 8 ]
7.5
5M2MjyNR2w
Now I have all the information needed. Let me compose the final consolidated review. ## Summary The paper proposes AdE (Adaptive Expansion), a hypergraph-to-graph expansion method that replaces fixed expansion strategies with learned, task-adaptive ones. AdE has two components: (1) GSi-Net, which learns a feature-imp...
[ 3, 8, 6, 5 ]
5.5
5XUlfPcQnG
Now I have all the verification needed. Let me compose the final consolidated review. ## Summary This paper proposes a lightweight 2D finite-difference thermal simulator for office buildings, augmented with a randomized air-shuffle approximation of convective mixing, and a calibration procedure that tunes physical pa...
[ 5, 1, 1, 3 ]
2.5
5jcav5RcKw
## Summary This paper proposes the Joint Autoregressive Mixture (JAM) framework for fusing two pretrained autoregressive decoder-only models — an LLM (7B) and a text-to-image model (CM3leon, 7B) — into a single architecture capable of interleaved text-and-image generation. Three fusion strategies are explored: weight ...
[ 5, 6, 5, 6 ]
5.5
5kMwiMnUip
All retrieved anchors are topically relevant to LLM jailbreaking or safety (the paper's area). Here are the final anchor paths:
[ 1, 3, 1, 1, 1 ]
1.4
5lUdTogEL3
Here is the final consolidated review: --- ## Summary This paper proposes a new task called "Cloth-Irrelevant Lifelong person Re-IDentification (CIL-ReID)" and an "Adaptive Discriminative Knowledge Consolidation (ADKC)" framework with Selective Knowledge Forgetting (SKF) and Selective Knowledge Retention (SKR) modul...
[ 1, 1, 1, 1 ]
1
5sixirvG0I
Here is the final consolidated review. --- ## Summary This paper extends Whittle index to a constrained MARL setting with multiple discrete actions and a state-space constraint (rather than the standard action-space constraint), yielding WIMS (Whittle Index with Multiple actions and State constraint). The key contri...
[ 8, 3, 5 ]
5.33
5z9GjHgerY
Now I have all the information needed. Let me construct the final consolidated review. ## Summary This paper presents DPLM-2, a multimodal diffusion protein language model that extends the discrete diffusion-based DPLM to jointly model, understand, and generate both protein sequences and 3D structures. Key technical ...
[ 3, 8, 8 ]
6.33
5zwrpqYIK5
Now I have all the information I need. Let me write the consolidated review. **Summary of verification against the paper:** 1. **Theorem 4 conditions (i)-(iii):** The paper text (lines 247-249) says "if the following conditions are satisfied:" followed by definitions of s, κ, c₂, c₃, then refers to "condition (i)" at...
[ 5, 3, 3, 3 ]
3.5
60rQpnbgmE
Now I have a thorough understanding of the paper. Let me write the final consolidated review. ## Summary This paper addresses the problem of variance in self-consistency (SC) confidence estimation for LLM reasoning when the number of samples is limited. The authors propose Perplexity Consistency (PC), which replaces ...
[ 6, 3, 3, 5 ]
4.25
Free AI Image Generator No sign-up. Instant results. Open Now