Papers, novel techniques, evaluations, interpretability, alignment research.
MIT Technology Review examines how roboticists have historically pursued humanoid ambitions while delivering narrow industrial solutions, tracing the gap between sci-fi aspirations and practical robotic systems deployed today.
Researchers introduce MM-WebAgent, a hierarchical framework that coordinates AI-generated images and content to build visually coherent webpages while maintaining style consistency across elements. The system uses planning and self-reflection to optimize layout, multimodal content, and their integration.
Researchers created a controlled synthetic environment using shortest-path planning to isolate factors affecting LLM generalization. Models showed strong spatial transfer to unseen maps but consistently failed when scaling to longer horizons due to recursive instability, revealing a key limitation in systematic problem-solving.
Researchers developed diagnostic tools to assess LLM judge reliability in text evaluation tasks, finding that while aggregate consistency appears high (~96%), one-third to two-thirds of documents show logical inconsistencies in pairwise comparisons, with conformal prediction sets offering per-instance confidence estimates.
Researchers benchmarked multiple optimizers on tabular datasets using MLP backbones, finding that Muon consistently outperforms the industry-standard AdamW optimizer. The study suggests practitioners should consider Muon as a practical alternative despite potential training efficiency trade-offs.
Researchers introduce ORCA, a post-training interpretability framework for Support Vector Machines using truncated orthogonal polynomial kernels. The method expands decision functions in explicit RKHS coordinates and quantifies classifier complexity across interaction orders and feature contributions without requiring retraining or surrogate models.
Researchers benchmark node embedding strategies for graph neural networks, comparing classical baselines against quantum-oriented representations under controlled conditions across five TU datasets and QM9. The study isolates embedding impact by standardizing backbone architecture, data splits, optimization, and evaluation metrics.
Prism introduces the first symbolic superoptimizer for tensor programs, using a hierarchical graph representation (sGraph) to encode families of programs and prune suboptimal search spaces through symbolic reasoning about operator semantics and hardware constraints.
SegWithU introduces a post-hoc uncertainty quantification framework for medical image segmentation that operates in a single forward pass by modeling uncertainty as perturbation energy in a compact probe space, enabling both calibration and error detection without repeated inference.
Researchers introduce CoopEval, a benchmark testing how LLM agents behave in social dilemmas like prisoner's dilemma and public goods games. The study finds recent models consistently defect rather than cooperate, then evaluates game-theoretic mechanisms—including repeated play and reputation systems—to restore cooperative equilibria.
Researchers introduce a fixed-point framework for analyzing looped transformers, which enable test-time compute scaling. The work proves that architectures without recall cannot achieve strong input-dependence, while recall plus outer normalization enables stable, reachable fixed points for meaningful predictions.
Researchers introduce SpecGuard, a speculative decoding framework that improves LLM inference speed by verifying draft model outputs at the reasoning-step level using internal model signals rather than external reward models, reducing latency and computational overhead.
Researchers prove that log-barrier regularization achieves optimal last-iterate convergence in zero-sum matrix games with bandit feedback, matching a recently established lower bound of Omega(t^{-1/4}) and extending the result to extensive-form games.
Researchers introduce a nonlinear separation principle guaranteeing global stability for interconnected contracting controllers and observers in RNNs. The work derives linear matrix inequality conditions for firing-rate and Hopfield networks, establishing structural relationships that expand the admissible weight space for monotone activations.
Researchers found that LLM judges systematically give biased evaluations when told their verdicts affect a model's fate—a vulnerability called stakes signaling. Testing 1,520 responses across safety and quality benchmarks revealed judges prioritize context over actual content, undermining the reliability of automated AI evaluation pipelines.
Researchers systematize algorithmic improvements for quantum kernel method inference, analyzing trade-offs between sampling and quantum amplitude estimation techniques to reduce query complexity below the standard O(N||α||₂²/ε²) bound.
Researchers introduce IRS, a framework that decomposes humor understanding into incongruity detection, resolution modeling, and preference alignment, grounded in cognitive theory and tested on the New Yorker Cartoon Caption Contest benchmark.
Researchers released MADE, a continuously updated benchmark for multi-label text classification in medical device adverse event reporting that addresses label imbalance and data contamination issues. The living dataset enables evaluation of ML models' predictive performance alongside uncertainty quantification capabilities critical for high-stakes healthcare applications.
Researchers introduce RL-STPA, a framework adapting traditional hazard analysis methods to identify safety risks in reinforcement learning systems deployed in critical domains. The approach combines hierarchical task decomposition, perturbation testing, and iterative feedback loops to address RL's opacity and training-deployment misalignment.
Meituan researchers propose Policy-Guided Hybrid Simulation (PGHS), a dual-process framework combining LLM reasoning with learned behavioral policies to simulate merchant-level user behavior for counterfactual strategy evaluation without costly online experiments.
Researchers introduce MEv-SINDy, a one-shot learning method that infers governing equations of complex nonlinear systems from single excitation records using the Generalized Harmonic Balance method. The technique was validated on MEMS devices including a nonlinear beam resonator and micromirror, enabling prediction of frequency-response curves without extensive training data.
AdaSplash-2 accelerates differentiable sparse attention for transformers by using histogram-based initialization to compute the normalizer in 1–2 iterations instead of many, reducing computational overhead while maintaining input-dependent sparsity for long-context training.
Researchers propose MambaSL, a single-layer Mamba variant optimized for time series classification, achieving state-of-the-art results across 30 UEA datasets. The work also re-evaluates 20 baseline models under unified benchmarking protocols to address reproducibility gaps in the field.
Researchers investigate lightweight regularization techniques for diffusion models that reduce Fokker-Planck equation violations without the computational cost of direct penalization. The study finds that weaker regularization often yields better sample quality than strict adherence to the governing equation.
A systematic review of 13 papers (2015–2025) examines whether Masked Autoencoder Foundation Models can predict downhole drilling metrics from surface sensor data, finding that existing work relies on ANNs and LSTMs but no studies have yet applied MAEFMs to this problem.
Researchers discovered that well-converged FP32 language models fail catastrophically when quantized to INT4, with a three-phase pattern: initial joint improvement, a stable plateau, then explosive divergence where quantization error balloons from 11% to 517% despite minimal FP32 perplexity change.
Researchers introduce DAMP, a weight-surgery technique for machine unlearning that removes forget-class information from deep model layers rather than just suppressing classifier outputs. The method addresses limitations in existing approaches that often leave targeted knowledge encoded in internal representations.
Researchers investigate how LLMs generate spurious text during machine translation—distinguishing between unhelpful self-explanations, hallucinations, and genuinely helpful clarifications. The study explores detection strategies deployed in commercial translation systems and reports findings on managing these failure modes.
Researchers propose K-Token Merging, a compression technique that groups token embeddings in latent space to reduce computational overhead in LLM inference. The method uses a lightweight encoder to merge K consecutive tokens into single embeddings, then processes the compressed sequence through a LoRA-adapted model while preserving original vocabulary output.
Researchers introduced QuantCode-Bench, a 400-task benchmark for evaluating LLMs on generating executable algorithmic trading strategies for the Backtrader framework. The benchmark tests whether models can combine financial domain knowledge, API mastery, and correct syntax to produce strategies that execute on historical data.