LegalEvolver is a query optimization framework for legal judgment prediction that improves retrieval by generating legally informative queries. It uses heuristic seed queries and rubric-guided reinforcement learning to iteratively self-evolve, surpassing initial performance ceilings. Benchmarks show effectiveness over direct SFT+RL and fixed rubrics.
A new process-based hydrological model, CaRT+, built on physical laws and refined with AI assistance, achieves accuracy comparable to LSTM neural networks on catchments in Britain, the US, and Australia. With 35 parameters per catchment and no daily neural network, it reaches median Nash-Sutcliffe efficiencies of 0.863 and 0.765 in held-out years on British and US data, and outperforms LSTMs on Australian catchments. The results suggest that catchment-scale laws exist as laws of form, with shared constitutive families and only parameters learned.
Why it matters: This is a concrete, recent research result showing a process-based model matching deep learning accuracy, with implications for hydrology and AI-assisted model refinement.
MA-Mesh is a framework that adds a role-based self-healing layer to multi-agent systems, generating resilience-aware specifications and executable policy code. It introduces cross-agent reflection and role-based failover to improve robustness under uncertainty. Experiments across software development, tool use, and scientific reasoning show significant gains in performance and robustness.
Why it matters: Provides a concrete, novel framework with explicit resilience mechanisms and experimental validation, valuable for AI system design.
This paper introduces GIST, a two-level data selection method for offline reinforcement learning with verifiable rewards (RLVR). It derives a factorization of the KL divergence between the current policy and the Gibbs-optimal target, identifying which responses and tokens are most informative. GIST selects correct responses on-policy and incorrect responses near the policy's mode, and supervises only the tokens where the mismatch lies. Experiments show accuracy gains across GSM8K, MATH-500, and MBPP for three models, with 66.6% fewer supervised tokens in a controlled comparison.
This paper investigates zero-shot coordination (ZSC) in LLM agents, finding that even frontier models struggle to coordinate in simple, controlled environments due to limited understanding of the coordination problem and weak reasoning about partners' beliefs. The authors propose Coordination-friendly definitions (CFDs) as a principled approach to enable robust coordination)Skip
Why it matters: Provides concrete experimental findings and a novel method (CFDs) for improving LLM coordination, relevant to AI research and deployment.
This technical report presents three practical defense tricks against data poisoning-based backdoor attacks in multi-stage face recognition systems, assuming in-house training but compromised third-party data. The defenses include a training-time method to detect poisoned identities and two inference-time filters for suspicious inputscars. Each is evaluated on realistic workloads, aiming to provide actionable hardening techniques for system providers.
This paper frames LLM steering as a feedback-control problem, using the network's internal state to condition interventions, inspired by closed-loop neurostimulation. It introduces an H-infinity-based controller that explicitly handles model mismatch and disturbances, improving performance over baselines, especially in out-of-distribution settings. The authors also propose a pre-deployment measure, robust steerability, to quantify tolerance to dynamics mismatch.
Why it matters: Offers a novel cross-disciplinary framework with concrete robustness gains and a new evaluation metric, valuable for AI control and safety.
D3D is a standalone diffusion drafter that removes inter-round dependency in speculative decoding, allowing drafting of successor blocks during verification in a packed pass. It combines blockwise denoising with auxiliary autoregressive supervision and prunes layers to reduce cost. Across seven benchmarks, D3D achieves 5.00x and 5.71x speedups with Qwen3-8B and Qwen3-32B targets, outperforming DFlash.
Why it matters: This paper presents a concrete, recent AI development with measurable speedup improvements and open-source code, offering practical value for efficient LLM inference.
EviSeek is a training-free framework that makes question-derived evidence demands explicit acquisition objectives for embodied active question answering. It links observation-conditioned actions and selectively retained evidence to the information the agent seeks, avoiding undirected interaction. On 3,436 ESI-Bench questions, EviSeek improves accuracy over a Single-model agent by 7.1 percentage points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro, while also improving visual-support-weighted accuracy and reducing token usage.
SumOne is a standalone summarizer that compresses long-horizon agent histories via masked reconstruction, using a frozen reconstructor and GRPO to optimize without task-specific rewards. It reduces inference cost by ~78% on DeepSearchQA and ~75% on SWE-bench Verified with minimal performance loss, and transfers across models and harnesses. The method enables plug-and-play context compression for agents.
Why it matters: Provides a concrete, novel AI development with measurable efficiency gains and cross-task transfer, valuable for agent deployment.
AmbiHTMLBench is a new benchmark for evaluating document parsing systems that output HTML instead of Markdown, aiming to preserve fine-grained spatial structure and visual fidelity. It includes 1,115 diverse pages with annotations and uses both objective and subjective evaluation protocols. Tests on ten multimodal systems show that the best objective score is 0.8092, while the subjective winner differs, indicating complementary aspects of parsing quality. Fine-tuning a small expert model shows task-specific adaptation improves document-to-HTML reconstruction.
Why it matters: Introduces a novel benchmark and evaluation methodology for HTML-based document parsing, providing concrete results and insights for researchers and practitioners.
This study evaluates eight ImageNet-pretrained EfficientNet variants (B0-B7) for four-class retinal disease classification from SD-OCT images at fixed 224x224 resolution. Results show non-monotonic accuracy scaling, with B1 achieving the highest accuracy (96.95%) and Drusen F1, while B7 shows a secondary accuracy recovery. The OCT accuracy ranking differs from ImageNet hierarchy, indicating task-dependent scaling effects.
Researchers propose a unified methodology combining adaptive surrogate enhancement with simulation-driven inverse design for microfluidic droplet generators. Neural network ensembles predict flow regimes and droplet diameters, with uncertainty quantification guiding targeted simulations to improve accuracy near regime transition boundaries. The framework, validated on independent datasets, nearly doubles predictions within 0-3% error and reduces high-error cases, while inverse design reliably finds physically realizable configurations.
ArcRAG introduces a query-time interface induction method that makes retrieval over knowledge graphs portable across different graph constructions. It builds a local graph interface from entities mentioned in a question)Skip planning multi-hop traversals and executing them as Cypher, without needing construction-specific schema or retuning. Experiments on MuSiQue, HotpotQA, and TriviaQA show that this construction-agnostic retriever remains effective across schema-guided and open-information-extraction graphs, though absolute accuracy still depends on construction.
Why it matters: This paper addresses a critical gap in GraphRAG—cross-construction portability—and offers a practical solution with strong experimental validation, valuable for researchers and practitioners.
This paper challenges the standard practice of using uniform LoRA ranks across experts in model merging. Experiments with 36 experts across three domains and multiple scales show that assigning per-domain best ranks can improve multi-domain accuracy by up to 1.8 percentage points with averaging and utility-based methods, while magnitude-based conflict-resolution methods suffer. The authors identify that magnitude-based component selection discards useful spectral components and can drop entire experts, causing large performance drops.
Why it matters: Provides concrete evidence and a practical insight that can improve model merging efficiency and accuracy, relevant to AI practitioners.
FPTR is a new scheduling method for wireless systems that always maintains a feasible allocation (incumbent) and refines it in stages, only replacing it if a candidate is complete, validated, and improves the objective. On synthetic benchmarks, the full pipeline improves traffic by 3.83% over a lightweight reference and outperforms six search baselines by 6.68-8.35%, with all runs meeting the 100 ms deadline. The approach separates mandatory output construction from optional refinement, enabling better use of spare computation time.
Coca-Cola compresses tool observations into latent tokens embedded in the agent's input space, reducing token consumption by 70-89% and runtime by 24-56% while improving task performance on QA benchmarks. Action-Preserving Training aligns next-action distributions to retain decision-critical information.
Why it matters: Offers a concrete, novel method for agent context compression with significant efficiency gains and open-source code.
NaijaFacts is a new benchmark of factual and statistical questions about Nigeria, authored by native speakers and available in five languages. Testing eight models shows a dominant knowledge effect: English-query accuracy is only 20% for Nigeria versus 49-50% for US/UK, while language effects are smaller. The gap persists across all 54 African countries, with models scoring 96% on culture and 98% on politics but only 26% on statistics.
Why it matters: Provides concrete evidence of a major AI knowledge gap for African facts, with actionable insights for multilingual and knowledge-aware model development.
