All Updates
Page 1306 of 1311
February 12, 2026
CRL Steers SAE Features Token-by-Token
CRL uses reinforcement learning to select sparse autoencoder (SAE) features for steering language models at each token, revealing which features impact outputs. It includes adaptive masking for diverse features and enables analysis like branch point tracking and layer-wise comparisons. Tested on Gemma-2 2B, it improves benchmarks while providing interpretable logs.
Confounds Limit FM CT Specificity
Foundation models match task-specific discrimination in abdominal trauma CT but suffer specificity drops from negative-class heterogeneity like solid organ injuries. Task-specific models handle confounds better. Adaptation via labeled training reduces susceptibility.
CLI-Gym Scales CLI Task Generation
CLI-Gym generates 1,655 CLI tasks via agentic environment inversion from Dockerfiles. It simulates histories to create buggy states and derives tasks with error messages. Fine-tuned LiberCoder boosts Terminal-Bench scores by 21.1%.
C^2ROPE Advances 3D Multimodal Reasoning
C^2ROPE enhances Rotary Position Embedding for 3D Large Multimodal Models by addressing spatial locality loss and long-term attention decay. It introduces spatio-temporal continuous positional embeddings using triplet hybrid indices and Chebyshev Causal Masking. Evaluations show superior performance on 3D scene reasoning and VQA benchmarks.
BNRM Prevents Reward Hacking in RLHF
BNRM introduces Bayesian non-negative reward modeling to combat reward hacking in RLHF. It uses sparse latent factors for disentangled, debiased rewards. Scalable amortized VI enables end-to-end training on LLMs.
Blockwise Advantages for Multi-Objective RL
Introduces Blockwise Advantage Estimation for GRPO in structured generations, assigning per-objective advantages to avoid interference. Uses Outcome-Conditioned Baseline to estimate advantages without nested rollouts. Competitive on math tasks with uncertainty estimation.
Benchmark Tests TSFMs on Energy Loads
Multi-dimensional zero-shot benchmark evaluates four TSFMs (Chronos, Moirai, TinyTimeMixer) vs. baselines on ERCOT data. Tests context sensitivity, calibration, robustness to shifts like COVID/Winter Storm. Top models hit MASE 0.31; Chronos-2 best calibrated.
Benchmark for Self-Evolving Coding LLMs
EvoCodeBench evaluates LLM-driven coding systems on self-evolution, efficiency, and human-comparable performance across languages. Tracks dynamics like solving time and improvements over iterations. Enables cross-language robustness analysis.
Auto-Shaping Rewards for Robust Control
Proposes causal reward shaping from offline data for continuous RL under confounders. Derives tight value bounds via causal Bellman equation for PBRS. Outperforms SAC on benchmarks.
Authenticated Workflows Secure Agentic AI
Introduces authenticated workflows as a complete trust layer for enterprise agentic AI, protecting prompts, tools, data, and context. Enforces intent and integrity via cryptography and MAPL policy language. Integrates with nine AI frameworks for deterministic security.
AugVLA-3D Boosts VLA with Depth Augmentation
AugVLA-3D integrates depth estimation from RGB inputs via VGGT to enrich 3D features in vision-language-action models. An action assistant module ensures consistency with control tasks. It enhances generalization and robustness in complex 3D robotic environments.
AudioRouter Boosts LALMs via RL Tool Use
AudioRouter applies RL to teach large audio language models (LALMs) when to use external audio tools, improving fine-grained perception without heavy training. It optimizes a lightweight routing policy while freezing the base model. Achieves big gains on benchmarks with 600x less data than traditional methods.
Aletheia Powers Autonomous Math Research
Aletheia is a math research agent that generates, verifies, and revises solutions using advanced Gemini Deep Think. It achieves milestones like fully AI-generated papers, human-AI collaborations, and solving four open Erdos problems. The work proposes standards for quantifying AI autonomy in math.
AI-PACE Framework Boosts Medical AI Education
AI-PACE synthesizes literature to propose a framework for integrating AI into medical education across the learning continuum. It identifies key competencies, curricular approaches, and strategies emphasizing longitudinal integration and interdisciplinary collaboration. The framework balances technical fundamentals with clinical applications to prepare physicians for AI-enhanced healthcare.
AI Fails Basic Arithmetic Despite Advanced Math Wins
Frontier AI models excel in advanced math but consistently fail at multi-digit integer addition. Errors primarily stem from operand misalignment or carry failures, explaining most mistakes in top models like Claude, GPT, and Gemini. These issues link to tokenization and random carrying failures.
AgentTrace Enables AI Agent Observability
AgentTrace instruments LLM agents for structured logging across operational, cognitive, and contextual traces. Provides runtime transparency for security and monitoring in high-stakes settings. Minimal overhead supports accountability and risk analysis.
Affordances Build Partial LLM World Models
Proves LLMs possess predictive partial-world models via task-agnostic affordances for intents. Introduces distribution-robust affordances for multi-task efficiency. Reduces search branching in robotics, outperforming full world models.
Adversarial Threat Detection in Autonomous Driving
ADยฒ analyzes vulnerabilities in end-to-end driving agents like Transfuser to physics, EMI, and digital attacks in CARLA. Driving scores drop up to 99% under threats. Proposes lightweight attention-based detector for spatial-temporal consistency.
Adapters Unlock Reliable Self-Interpretation
Lightweight adapters trained on interpretability artifacts enable reliable self-interpretation in frozen LMs. A simple scalar affine adapter outperforms baselines in feature labeling, topic identification, and implicit reasoning decoding. Gains scale with model size, driven mostly by learned bias.
ADAlign Auto-Adapts Graph Domains
ADAlign tackles graph domain adaptation by adaptively aligning discrepancies via Neural Spectral Discrepancy (NSD). Uses neural characteristic functions and minimax sampling without heuristics. Outperforms SOTA on 10 datasets with efficiency gains.