New model releases and updates, with the benchmark numbers traced to who ran them. If a claim is the company's own and unverified, we say so — not after the fact.
68 published signals in MODELS — page 1 of 3
Neural networks can have redundant parametrizations, making their evolution dynamics ill-conditioned. Dirac-Frenkel dynamics with inertia can help. It's tested in simulation only, so we don't know if it holds up in real-world problems.
Deep transformers form hierarchical representations, but their expressivity is not well understood. This analysis uses bounded-depth grammars to study how they capture abstract features. The findings could impact language modeling and beyond.
Predicting pedestrian movement from ego-centric videos is tough due to complex scene interactions and intentions. This model tries to tackle that. We don't know yet how well it generalizes to real-world scenes.
Evidential Deep Learning models uncertainty with Dirichlet distributions, but its foundations are shaky. This update uses density-informed pseudo-counts to improve calibration. It's a step towards more reliable uncertainty-aware classification, but we don't know yet if it holds up outside benchmarks.
Multimodal AI agents are vulnerable to black-box visual attacks on their long-term memory. Lucid exploits this by manipulating visual data. This raises concerns about trusting AI memories.
Error-prone channels need smarter encoding. This method switches between two codes to mitigate burst errors, potentially reducing decoding delay. It's tested in simulation, but real-world impact is unclear.
Household electricity demand is hard to predict because people's habits vary wildly. This paper embeds inferred behavioral patterns into a neural process model to forecast short-term load. Most models fail to capture the diversity of household routines, but this one does better. If your smart grid relies on a model that can't handle real people, you're in trouble.
Linear models and single-qubit mixed-state models are compared for binary classification. Qubit models offer different interpretability, but it's unclear if that's an advantage. This comparison is still theoretical, not tested on real-world data.
Complex systems are hard to predict, and current models often fail to account for spatial and temporal patterns. This new framework combines multiple techniques to improve prognostics. It's tested in simulation, but we don't know yet whether it holds up in real-world systems.
Physicists still manually classify high-energy collisions. A new model uses a graph neural network to classify events from 12 physics processes, trained on 120 million simulated collisions. This could speed up physics discoveries, but we don't know yet how well it generalizes to real data.
Recursive reasoning models lack a principled inference mechanism for test-time scaling. This model uses energy guidance to improve recursive reasoning, but its real-world impact is unclear. If it works, it could change how we scale neural networks for complex tasks.
Machine learning models need to be optimized for deployment across different environments. A new framework helps select the right compression and acceleration techniques. This could make models more efficient and widely adoptable.
Transformer inference is bottlenecked by key-value cache memory costs, which grow with batch size and context length. A new method combines Tucker and JL-Residual allocation to compress the cache with minimal loss. This could significantly improve throughput for long-context models.
Traditional hub capacity planning models fail to account for qualitative business context. A new framework uses a large language model to propose hub capacity plans based on textual business inputs. This approach may improve planning accuracy, but its real-world effectiveness is untested.
Sampling from complex densities is hard. This method stops the sampler early when a classifier says it's good enough, which can speed up Markov chain Monte Carlo methods. We don't know yet if this holds up outside the benchmark.
Evaluating large language models in conversations is expensive and key events like jailbreaks emerge late. Dynamic budget allocation can help. This method could make LLM evaluation more efficient, but we don't know yet if it holds up outside the lab.
Attention heads define transport operators, and diagnostics read model behavior from their spectrum. This work explores the limits of such diagnostics. It affects how we understand model hallucination and behavior.
Conformal prediction guarantees fail under distribution shift, but pseudo-calibration can help. This method offers a way to maintain coverage guarantees even when the data distribution changes. It's a step towards more reliable predictions in changing environments.
A single transformer-based generative model can capture Standard Model structure from sub-GeV to TeV, a range no single Monte Carlo sample covers. This changes how we model complex particle interactions. We still don't know if this holds up outside the LHC data
Remote sensing vision-language models struggle to support open-ended reasoning over Earth Observation data. A new model uses a simple recipe to achieve large-scale results. Its real-world impact is still unclear.
Vision-language models struggle with test-time transduction. This model uses dynamic shrinkage to improve performance. Realistic evaluations are still a challenge
Automated essay scoring is held back by transformer limitations. This study proposes a generative AI solution to summarize long essays efficiently. It's unclear if this will actually improve scoring accuracy outside the lab.
Backdoor attacks can manipulate speech recognition models. SpeechGuard defends against these attacks online. Its impact on voice interaction security is significant.
Diffusion language models can generate text in parallel, but their quality lags. Adaptive multi-step lookahead decoding improves this by refining masked tokens more efficiently. This could make diffusion models more viable for real-world text generation tasks.
Hypergraphs model complex interactions, and a new analysis shows semi-supervised learning on them can be consistent with large datasets. This matters for understanding multiway relationships in data. The real-world impact is still unclear, but it could improve predictions in social networks or biology.