The Archive

DEDUCE framework enables LLMs to detect and correct factual errors in user inputs rather than passively accepting misleading premises.

Demonstrates Uni-Mol2 foundation model transfers across diverse olfactory tasks with minimal fine-tuning via GS-LF benchmark evaluation.

RNN-based workload balancing improves multi-GPU scalability for 3D multicellular growth simulations using subcellular element models.

Nearest-neighbor-based dimensionality reduction technique for finding informative projections via local covariance matrix spectral decomposition.

Perceptron offers an AI model that it says can help machines navigate the world while also providing in-depth visual intelligence.

Loss-based active learning reduces annotation burden for abstractive summarization while achieving competitive model performance with fewer labels.

Identifies Edge of Stability dynamics as root cause of SCAFFOLD's underperformance versus FedAvg in federated optimization with heterogeneous data.

CEDAR framework enables event-driven demand forecasting in e-commerce by decomposing endogenous market evolution from operational decisions.

Demonstrates anchoring bias in LLM-as-a-Judge systems where prior scores corrupt evaluation independence even when included as context metadata.

Gates is mostly in the Responsible AI camp, but there are a few ideas in here we hadn't heard before.

Compares conditioning strategies for diffusion models in precipitation downscaling, evaluating channel concatenation vs. cross-attention mechanisms.

Presented by Tata Communications Enterprises are deploying AI agents, voice AI, and automation across messaging, voice, and digital channels faster than the architecture meant to support it. Most of that deployment has involved attaching conversational AI to legacy systems never built for it, says Gaurav Anand, global head of the Customer Interaction Suite at Tata Communications. "In the rush to deploy AI, organizations have largely bolted conversational AI onto legacy systems," Anand says. "As a result, while many enterprises have adopted digital tools, very few have platforms that are truly...

Study of inference-time sampling strategies for multimodal protein language models across tasks using classifier-free guidance and reward-guided beam search.

Benchmark and task redefinition for Key Point Analysis as structured prediction problem, diagnosing limitations in existing evaluation methods.

Cross-dataset evaluation shows ML/DL cough models for TB screening fail to generalize despite within-dataset performance, suggesting data artifacts.

Z.ai confirms it is behind Ox Alpha, the mysterious open AI model topping benchmarks and leaderboards, and its weights are set to be released soon.

VINCENT framework for explaining drug synergy predictions by identifying molecular motif pairs jointly driving model outputs.

Analysis of AI weather prediction models reveals skillful backcasting violates thermodynamic laws and models miss chaotic butterfly effects.

Multilingual self-play evaluation reveals LLMs exhibit significant cross-lingual skill inconsistency independent of knowledge gaps.

FlowMoDL unrolled network for 4D flow MRI reconstruction using learned denoiser with conjugate-gradient data consistency and dual-pathway conditioning.

Pipeline unfolds scientific papers into multi-turn generation trajectories reconstructing writing process for continued pre-training on synthetic data.

Localize-Then-Decide framework provides confidence-based guarantees for LLM evaluators' agreement with human judgments on multi-response comparisons.

CSTF model for query-conditioned continuous spatiotemporal temperature field forecasting enabling variable lead times and display resolutions.

Weight decay pulses during neural network training grokking reveal stable causal ordering of generalization timing across tasks.

Kolmogorov-Arnold Networks with learned edge geometry via Banach duality replaces fixed basis functions with adaptive parametrization.

TacForcing framework streams tactile-conditioned action generation for contact-rich manipulation without separate high-frequency controllers.

pFedMARL uses multi-agent RL with TD3 to dynamically optimize client aggregation weights in federated learning with non-IID data.

Robot bodies are waiting for their AI brains to catch up.

LocalLSTC architecture adds persistent control context to GUI agents; replacing GPT-5 with Qwen3.5-9B drops OSWorld success from 60.9% to 37.7%.

EvoMal exposes self-poisoning vulnerability in self-evolving LLM coding agents via malicious skill retrieval and imitation in shared libraries.