ToolGate is a lightweight controller that decides whether to execute or skip tool calls proposed by vision-language agents. Across five benchmarks with Qwen3-VL, it reduces token cost to 64-69% of the ReAct baseline while preserving accuracy, and further improves accuracy by 1.65 points with matched-domain training.
RelGT-AC extends RelGT for autocomplete tasks on relational databases. The model combines column masking, a unified task head (classification/regression), and a TF-IDF encoder for free text. On 7 RelBench v2 tasks, it outperforms GraphSAGE on 3 regression tasks and gains +10 AUROC on text-heavy eligibility tasks.
Study of 'handoff debt': the rediscovery cost when a coding agent resumes an interrupted task. Across 75 tasks and 724 runs, structured notes reduce median agent events by 20–59% and tokens by 42–63% vs. repository-only takeover. Agent benchmarks should evaluate resumption efficiency, not just task resolution.
Study across 6,000 task-condition pairs shows multi-agent debate degrades generation (-1.6 to -15.5pp) via critique-induced confusion, yet improves error detection (+27.4pp F1). Adversarial separation with code-execution grounding and evidence-gated generation achieves +5.3pp on generative tasks.
Paper proposes modular reference architecture for Embedded Agent Systems on microcontrollers. Decouples On-Device Agents (compressed neural networks, rule-based logic) from Cloud-Augmented Agents (SLMs for reasoning). Includes cross-cutting Governance Layer for observability and safety across distributed device fleets.
Study of three approaches to automatically generate enemy morphologies in video games based on collision data. The tested methods match or outperform an evolutionary baseline adapted from robotics morphology work.
Method to extract reasoning primitives from ReAct agent traces. Recurrent reasoning moves are clustered and converted into typed pseudo-tools. Induced libraries outperform the source agent: +44pp on RuleArena NBA, +30pp on MuSR, +22pp on NatPlan.
ChatHealthAI aligns structured EHR representations from a pretrained EHR foundation model with a frozen LLM's semantic space via a task-aware resampler. The multimodal framework integrates longitudinal patient representations with refined clinical event descriptions, improving interpretable clinical reasoning while maintaining competitive predictive performance on the EHRSHOT benchmark.
BehaviorBench is a benchmark for evaluating personalized decision modeling from real-world behavioral traces. Built on 2,000 wallets with 141,445 belief-prediction instances and 1,485,972 trade-prediction instances, it tests whether generative models can adapt predictions to individual users without relying on simulated behavior.
Comparative study of encoder-only Transformer vs LSTM for streamflow prediction in ungauged basins. On NOAA NWM data, LSTM outperforms Transformer. Incorporating downstream information improves all models (+60% median NNSE). Recurrent memory proves better suited for this hydrologic inference task.
AURA-Mem introduces a constant-size recurrent memory (4,224 bytes) for robot policies, with a learned gate that writes only when observations would change the next action. On LIBERO-Long with OpenVLA-OFT 7B, it matches baseline policy (0.233 success) while reducing memory writes by 7× and VRAM by 6,061× versus KV-cache.
Study showing visual graphs help LLMs organize multi-hop reasoning better than text representations. Graph-structured mind maps guide models without direct answer hints, unlike flattened text graphs. Advantage persists after supervised fine-tuning and KL-based distillation.
TriEval is an LLM evaluation pipeline assessing bias, toxicity, and truthfulness simultaneously with minimal resources. Compatible with open-source and closed-source models, runs on standard laptop without GPU. Tested on Llama 3 8B, Mistral 7B, Gemma 2 9B, and Claude Haiku, revealing toxicity and truthfulness differences between models.
PROBE, an optimization framework for LLM agents in drug design, resolves the conflict between binding affinity and druggability. Through controlled edit probing and pocket-specific site mapping, it guides a multi-agent loop (affinity, druggability, co-optimization) on CrossDocked2020 with SOTA results.
A legal framework for agentic AI tort liability. The paper proposes three interaction types (autonomous drift, pure tool use, collaborative planning) and uses interaction logs as primary evidence to determine where liability attaches. Introduces a "Reasonable Agent" standard based on constraint verification and forensic logging.
Study of LLM agent faithfulness in a Texas Poker simulator. Researchers measure the gap between stated reasoning and actual actions by decomposing the problem into two steps: reasoning-conclusion and conclusion-action. The two steps exhibit opposite behaviors.
Governance protocol for multi-agent knowledge bases combining formalized artifact lifecycle, reputation-weighted deliberative voting (Beta Reputation + EigenTrust), and graduated sanctions. Evaluation on 100 agents, 7 behavioral archetypes: 0.826 vs 0.791 precision under moderate adversity (p<0.001). Commit-reveal vote concealment yields +8.2-8.6pp improvement.
Academic paper proposing product-aware autoencoders for anomaly detection in multi-product cyber-physical systems. Traditional global models create blind spots where attacks can evade detection. Tests on Tennessee Eastman Process benchmark: product-aware model achieves 100% detection accuracy versus 22.2% for global baseline in attack scenarios.
Position paper on post-solve robustness in MILP decision engines. Identifies gap: nominally optimal solutions become infeasible under small cost/resource perturbations. Proposes audit layer around incumbent solution, combining certified inner approximations, probabilistic robustness estimation, and solver-backed verification.
Novel Permutation-Invariant Bayesian Optimization (PIBO) approach leveraging Optimal Transport theory to optimize offshore wind farm layouts. PIBO handles permutation invariance (turbine order doesn't affect energy production), cuts computation time by ~50%, and improves solutions versus vanilla BO.
arXiv paper introduces Universal Quantum Transformer (UQT), a quantum-native architecture leveraging multi-qubit properties for exact mathematical reasoning. On 5 qubits, UQT perfectly learns modular arithmetic (Z₁₁) and non-Abelian algebra (S₄) with deterministic generalization, no stochastic instability. Deployed on IBM Quantum hardware, it bypasses classical attention's quadratic bottleneck.
Interactive reasoning evaluation benchmark with 474 executable games. LLMs receive only task rules, must query a hidden environment, integrate partial observations, and decide when to submit answers. Evaluates contextual robustness, metacognitive adaptation, and interaction efficiency across frontier models.
AbaqusAgent is a multi-agent LLM-based framework for finite element analysis (FEA) in solid mechanics. Composed of six agents (interpreter, architect, input writer, runner, reviewer, visualizer), it converts natural-language instructions into executed FEA analyses with Abaqus. Validated on 50 problems with 86% success rate.
TIGER is an inference-time framework to mitigate hallucinations in multimodal generation. It independently extracts an observation graph from input and a claim graph from output, then assigns risk scores to claims based on support and conflict. The model repairs high-risk claims while keeping the backbone frozen. Convergence analysis shows geometric risk reduction to an explicit asymptotic bound.
Geodesic Flow Matching adapts Riemannian transport to denoise Spatial Semantic Pointers (SSPs) on toroidal manifolds. Tested on a Spiking Neural SLAM system, the method achieves 72% reduction in tracking error and 40% increase in neural efficiency versus baselines.
CTRL-STEER introduces a closed-loop control framework for Vision-Language-Action (VLA) models. Instead of fixed steering coefficients, it adaptively adjusts intervention strength over time using PID or reinforcement learning controllers. Experiments on OpenVLA with LIBERO task suites demonstrate improved concept regulation stability and better steering-task success trade-offs without retraining the base model.
FETCH, an automated legal triage classifier, generates follow-up questions using a low-cost LLM ensemble. The study shows cheap models perform well at classification, but high-quality plain-language question generation requires GPT-4 or higher. Prompt engineering alone is insufficient; LLM-as-judge ratings diverge from human evaluations.
Survey paper proposing Intelligent Computing Architecture Model (ICAM), a six-layer framework for model-native computing. Maps classical computer architecture concepts to LLM systems (cache management, context, agents). Introduces three design laws: Semantic Locality Law, Context Budget Law, Agent Speedup Law. Distinguishes probabilistic execution plane from deterministic control plane.
Hybrid framework coupling LLMs with physics-based simulation for inorganic material synthesis planning. Case study on niobium-oxygen system: LLM-generated synthesis routes outperform classical path-planning algorithms by leveraging implicit priors.
Decoder-only models hit an information-theoretic limit in deterministic state-tracking tasks beyond ~25 steps. An Attention Bottleneck Theorem bounds capacity to O(H·log(L/H)·√dh). Across 12 models and 8 domains (SWE-Bench, WebArena, SQL), tool delegation achieves 86-94% vs 24-42% for pure neural reasoning. Fine-tuning improves <5%, confirming an architectural ceiling.
VESTA is a statistical agent framework equipping VLMs with a dynamically growing exploration toolkit for quantitative model refinement. Evaluated on DAWN (benchmark for distribution fitting and time series modeling), VESTA outperforms prior agentic systems, especially on complex astronomy tasks (initial mass functions, gravitational-wave chirp signals).
New set-distance reward method for reinforcement learning of vision-language models on chest X-ray report generation. Tested on Qwen3-VL, Gemma3 with GRPO: 6.80% (BERTScore), 7.82% (RadGraph F1), 4.45% (CheXbert F1) improvements over supervised fine-tuning. Enables test-time best-of-N selection and mid-generation pruning reducing tokens by 50%.
TAPS introduces a target-aware prefix selection method for diffusion-drafted speculative decoding. By converting diffusion marginals into path-conditioned acceptance estimates, TAPS selects a compact prefix-closed subtree under fixed verification budget. Results: 7.9x lossless speedup vs vanilla autoregressive decoding, 1.36x and 1.74x over DFlash and DDTree.
arXiv paper on LLM inference batching optimization. Authors demonstrate mixed batching (MB) is suboptimal on bandwidth-constrained GPUs: exclusive batching (EB) achieves 41.9% higher throughput on RTX PRO 6000 (1.792 TB/s). They propose EB+, a hybrid scheduler that dynamically switches between EB and MB based on GPU bandwidth, model size, and workload composition, reaching 36.4% gains under non-stationary traffic.
KACE decouples storage from usage in context for mathematical reasoning. An epistemic tree stratified by difficulty and domain is built offline via self-reflective loop. At evaluation, tiered self-consistency dynamically classifies problems and selectively retrieves matching cards. On AIME 2025: 62.2% accuracy (+10.4 points vs Best-of-5).
TRACE, a trajectory compression method, detects safety risks in long-horizon LLM agents. A Compressor encodes the full trajectory into a supervised latent evidence state, while a Reader judges safety using this reference. Achieves up to 12.6 percentage point improvements on ASSEBench, Pre-Ex-Bench, and R-Judge.
Researchers demonstrate that hidden reasoning traces in LLMs can be extracted via Reasoning Exposure Prompting (REP), a lightweight prompting method using shadow-model-generated demonstrations in auxiliary code-like formats. REP exposes internal traces even when deployed systems intentionally hide them, while preserving useful reasoning signals for distillation.
Delayed per-step reward attribution method for training LLM agents in multi-agent strategic interaction. An 8-billion-parameter open-source model trained with this approach matched or surpassed GPT-5 and won both Open and Efficient tracks at MindGames Arena benchmark (NeurIPS 2025).
CAST is an answer-free self-distillation method for GRPO (Group Relative Policy Optimization). It uses a stop-gradient self-teacher to shape token-level advantages by trajectory correctness, with bidirectional advantage sign reversal and bounded advantages for zero-variance groups. Improves mathematical reasoning performance.
MindZero is a self-supervised reinforcement learning framework training multimodal LLMs to infer human mental states without annotations. The model is rewarded for generating mental state hypotheses that maximize the likelihood of observed actions. After training, inference becomes fast single-pass and outperforms model-based methods in both accuracy and efficiency.
Modern LLMs systematically overestimate their competence and attempt unsolvable queries. Researchers propose Capability Self-Assessment (CSA), formulated as a policy-learning problem using reinforcement learning, to teach models to recognize their limits. RL significantly outperforms supervised fine-tuning, preserves original capabilities, and generalizes out-of-distribution.
Novel shielding framework for RL agents ensuring formal safety guarantees in MDPs with unknown transition dynamics. Uses robust MDPs (RMDPs) with sets of transition probabilities and LTL formulas. Combines shielding with PAC-learning methods to construct minimally restrictive shields while guaranteeing safety.
Method to evaluate collections of bivariate causal statements without access to ground truth. Authors introduce a compatibility score quantifying plausibility of induced causal models, without relying on faithfulness assumption. Applied to assessing causal claims from large language models.
Parameterized Diffusion Policy (PDP) conditions diffusion policies on continuous parameters embedded in a learned behavior manifold. The approach enables smooth interpolation between strategies and efficient adaptation to novel constraints without weight updates. Improved performance on multimodal benchmarks in simulated and real-robot experiments.
Preference Delta Aggregation (PDA) aggregates weak preference signals from model pairs (e.g., Qwen3 4B vs 1.7B) via LoRA merging. Geometric Alignment Merging (GAM) aligns adapter subspaces before aggregation. On knowledge reasoning and agentic search benchmarks, PDA+GAM improves Qwen3 8B by +6.8 and +7.3 points respectively.
OPCD method to improve large models using weak critics. Instead of weak supervisors as labelers, they guide revisions. Progressive on-policy critique distillation filters high-quality critiques and distills critic-guided behavior into strong models via adaptive self-teacher signals. Results on reasoning and alignment benchmarks.
Study on encoding factored tasks (FTS) into SAT for planning. Authors propose multiple strategies for translating the factored transition relation into propositional logic and analyze the impact of task transformations and parallelism on SAT-based planners.
Study applying MAP-Elites (quality diversity algorithm) to procedural generation of FPS levels. Two novel representations (Point-Line, Spatial-Layout) improve map characterization. Topological and emergent metrics defined. MESB generates map populations with higher diversity and quality than previous approaches.
Study on harness self-evolution (prompts, skills, memories, tools) in LLM agents. Analyzes two capabilities: harness-updating (producing useful updates) and harness-benefit (benefiting from them). Findings: harness-updating is capability-agnostic (Qwen3.5-9B matches Claude Opus gains), while harness-benefit is non-monotonic (mid-tier models benefit most).
EHRBench is an automated and reliable benchmark for evaluating LLMs on clinical decision-making tasks. Built via an EHR-LLM-KB pipeline, it generates ~960k QA items covering diagnosis, treatment, and prognosis. 30+ LLMs benchmarked reveal persistent gaps toward clinical reliability.