This paper presents a solution combining Graphs of Convex Sets (GCS) for trajectory optimization and the Consensus-Based Bundle Algorithm (CBBA) for distributed task allocation in multi-agent systems. GCS finds optimal trajectories in a 3D+time configuration space, while CBBA coordinates task assignments, enabling collision avoidance and accurate time estimates. Effectiveness is demonstrated in simulated cluttered environments with static and dynamic tasks.

This paper introduces N(CO)$^2$, a reinforcement learning-based neural combinatorial optimization framework that solves the Stochastic Orienteering Problem without hand-crafted heuristics. It achieves competitive performance with state-of-the-art mixed-integer linear programming across diverse instances, reducing human design effort.

VEGA trains navigation Vision-Language-Action (VLA) models from unlabeled egocentric videos by reconstructing scene geometry and generating obstacle-aware trajectories. It introduces VEGA-Bench with 250k scenes and 5M navigation goals, achieving significant improvements in collision reduction and obstacle clearance.

PAIWorld introduces a diffusion-transformer framework with geometry-aware cross-view attention, geometric rotary position embedding, and latent 3D-REPA distillation to achieve multi-view 3D consistency for robotic manipulation, ranking 1st on WorldArena and 2nd on AgiBot-Challenge2026.

Guava is a harness framework for embodied manipulation, identifying three key ingredients: iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations. It distills embodied capabilities into a 4B open-source model using fewer than 2K simulation trajectories, achieving performance comparable to frontier proprietary models with strong generalization.

This paper introduces ReSYNC, a method that progressively discovers and refines state abstractions from failure-recovery experience to support abstract planning, jointly learning skills and concepts through a dual process. It outperforms baselines by over 50% in simulation and transfers to real-world tasks.

Large language models (LLMs) are increasingly used for clinical text tasks, but their ability to preserve diagnostic uncertainty is underexplored. This paper introduces a benchmark of 1,200 clinical documents with 9,184 uncertainty annotations across five levels, and evaluates three LLMs. Results show that LLMs preserve original uncertainty cues less than half the time and struggle with nuanced distinctions between adjacent levels, highlighting a failure mode not captured by standard metrics.

The Montreal Forced Aligner (MFA) was released in 2016 and has since become the most widely used tool for forced alignment in research and industry. MFA 3.0 achieves state-of-the-art or near state-of-the-art performance across four benchmark datasets with mean boundary errors below 15 ms. Adaptation and cross-language remapping are effective for languages outside MFA's training distribution, and pronunciation probability modeling and phonological rules provide gains in specific conditions.

A cross-lingual mechanistic analysis reveals that math-related parameters in multilingual LLMs partially overlap across languages, with English having the largest set.

VISUALSKILL is a hierarchical multimodal skill library that incorporates visual figures into skill artifacts, significantly improving computer-use agents' performance on long-horizon tasks and unseen software. On CUA-World and OSExpert-Eval benchmarks, a Claude Code CLI agent using VISUALSKILL achieved an average score of 0.456, a +15.3 point absolute lift over the no-skill baseline and +8.3 points over a text-only skill.

JetFlow is a head-based speculative decoding framework that combines one-forward drafting efficiency with branch-wise causal conditioning, enabling larger draft budgets to yield longer accepted prefixes and higher speedups. It achieves up to 9.64x speedup on MATH-500 and 4.58x on conversational tasks with Qwen3 models.

This paper proposes activation steering as an alternative for synthetic data generation in low-resource languages, using Language Steering and Quality Steering strategies. Experiments on four open-source LLMs and 11 languages show that steering on early layers improves data diversity and downstream task performance.

SproutRAG is a novel hierarchical RAG framework that uses learned inter-sentence attention to build a binary chunking tree, enabling multi-granularity retrieval without costly LLM calls or lossy summarization. It achieves a 6.1% average improvement in information efficiency across four benchmarks.

Educational dialogue is a valuable but sensitive resource. Existing approaches force a tradeoff between governance and accuracy. This paper proposes a fully local cascade framework that reframes de-identification from open-ended entity recognition to constrained privacy triage. A recall-first union proposer over-generates candidate spans, and a context-aware reviewer makes binary Redact/Keep decisions. On math tutoring transcripts, the strongest local configuration achieves 0.958 macro F1, outperforming LLM-only baselines and a commercial API, while running entirely on a single laptop.

CoAT introduces a continuous latent workspace for LALMs to preserve acoustic information via expert distillation, improving performance without extra decoding cost.

SAGE is a post-hoc sanitization method that repairs the retention damage caused by LLM unlearning. By extracting dominant activation geometry from a retain proxy and solving a closed-form optimization, SAGE suppresses update components aligned with high-energy retained directions while preserving the forgetting carrier, alleviating the retain-forget trade-off. Experiments across multiple unlearning methods and model scales demonstrate consistent retention improvement.

TRIDENT, a novel MARL framework, addresses the three-way coupling of hybrid discrete-continuous actions, safety constraints, and physics dynamics. It introduces Richardson-Romberg gradient correction, Lyapunov-constrained trust-region updates, and a physics-informed residual critic, achieving an O~(1/√K) convergence rate and reducing training violations by up to 95.5% over baselines.

DRIFT proposes a novel method for refining supervised fine-tuning data by using on-policy influence functions, which address key limitations of standard influence functions. It improves the data distribution to enhance the capability upper bound of LLMs, and achieves consistent improvements on 7B models.

This paper proposes a structural pruning framework for Mixture-of-Experts models by reformulating prune-ratio allocation as a channel-score coverage maximization problem, solved efficiently via attribution-based approximation. Experiments on DeepSeek and Qwen MoE models show accuracy preservation under 50% or 25% structured pruning with 4-bit quantization, achieving 5.27× memory reduction on Qwen3-30B-A3B and outperforming baselines.

This paper establishes a mathematically explicit link between shock-wave theory and the symmetry-quotiented learning dynamics of stochastic gradient descent, using differential geometry, Lie group theory, and fluid mechanics. After quotienting parameter symmetries and applying local-entropy coarse-graining, the effective dynamics satisfy a viscous Hamilton-Jacobi equation on the quotient manifold. Under the assumption that raw parameter dynamics can be summarized by a gradient field on the quotiented space, the gradient of the coarse-grained loss function obeys a Burgers-type equation, and shock formation can be rigorously established. The theory is applied to multilayer perceptrons, CNNs, Transformers, and mean-field networks, which all obey the Hamilton-Jacobi or Burgers-type equations.…

Artemis is a region-level causal framework that eliminates confounding from demographic factors in multimodal neuroimaging by independently intervening on each brain region, learning region-specific confounder representations, and improving GNN robustness and interpretability in disease diagnosis and classification tasks.

CODEBLOCK is a structure-aware sparse supervision framework for fine-tuning code LLMs. It selects structure-complete code blocks rather than isolated tokens, achieving stronger average pass@1 on six benchmarks while using only 1.9% of supervised tokens.

As RL agents improve, the supply of frontier tasks—valid, solvable tasks at the right difficulty—becomes the limiting factor. The PROPEL framework trains task generators at a targeted solve rate using a lightweight activation probe that predicts solver pass rate, avoiding costly solver rollouts. Experiments show significant gains in generating learnable frontier tasks across coding and software engineering benchmarks.

Gaussian Mixture Attention (GMA) is a novel attention mechanism that replaces pairwise query-key comparisons with routing through K learned Gaussian mixture components, achieving O(NK) memory scaling instead of O(N²). Empirically, GMA is competitive on long-context classification, and causal GMA outperforms linear/random-feature attention on WikiText-103 but lags behind optimized causal SDPA and Mamba. GMA offers a probabilistic, interpretable, fixed-K linear-time attention alternative.

ProfiLLM is an agentic LLM data pipeline that creates utility-aligned user profiles for ride-hailing dispatch. It uses tool-augmented global knowledge mining and utility-aligned profile exploration to overcome scaling challenges. Deployed on DiDi, it achieves significant improvements in prediction AUC, GMV, and completion rates.

R2D-RL bridges RCSS2D and Python-based MARL via shared memory, supporting full-field/scenario training, configurable opponents, hybrid action spaces, EPV-shaped rewards, and parallel execution, with baseline results for 11v11 matches and front-goal scenarios.

A new paper develops a formal account of memory requirements for generalist AI agents, proving that near-optimal performance across multiple environments requires storing domain-specific information beyond current observations. The research shows that memory serves as a substrate for domain disambiguation and transition model reconstruction.

Researchers introduce ForecastBench-Sim, a simulated-world forecasting benchmark built on Freeciv, a turn-based strategy game modeled on Civilization. It overcomes real-world benchmark limitations by providing quickly resolvable tasks, rare event examples, and counterfactual scenarios through game rollouts.

DeFAb is a dataset and pipeline converting four decades of public knowledge bases into defeasible abduction benchmarks. A symbolic solver achieves 100% accuracy in under 50 microseconds, while the best frontier LLM reaches 65% and drops to 23.5% under robust evaluation. The benchmark uses polynomial-time verifiable checks to measure logical rigor and creativity. It releases 372,648+ instances, along with harder variants and a Lean 4/Mathlib creative variant.

CEO-Bench is a new benchmark that evaluates language model agents on long-horizon, uncertain tasks by simulating running a startup for 500 days. Even the most advanced models like Claude Opus 4.8 and GPT-5.5 barely finish above the $1M starting balance and fail to consistently turn a profit.