Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
Enterprise data analysis requires agents to connect analytical requests with knowledge distributed across databases, business documents, and existing example tasks. Without an explicit representation...
OmicsBench is a new long‑horizon, multi‑turn bioinformatics benchmark for AI agents, comprising 97 tasks that link 662 subtasks into directed acyclic graphs. Each subtask is a user turn scored deterministically, with optional in‑context feedback. Across 12 models, agents achieve mean scores between 0.43 and 0.67, revealing lower performance on dependent subtasks and highlighting qualitative failure modes through an unsupervised discovery pipeline.
The paper introduces PhysClosureBench, a benchmark of 50 expert‑verified physics research tasks that test whether AI agents can organize investigations and produce evidence‑supported conclusions. Each task specifies a research goal and constraints, while the agent must design the investigation and validate evidence. Seven leading LLMs were evaluated, with Claude‑Opus‑5 scoring highest on local scientific accomplishment but achieving full research closure (RC4) only 32.7% of the time, revealing a gap between local success and complete closure.
The study examines whether the CodeHealth metric can predict the token cost of agentic code review and fix for pull requests. Using a pre‑registered observational design, 465 agent‑authored PRs were processed by a five‑agent pipeline across three LLMs, and mixed models showed that each CodeHealth point corresponded to a 3.5%–6.1% reduction in token usage, independent of PR footprint. The effect was strongest during the review stage and appears mediated by file size rather than PR size.
The paper introduces Market‑Based Post‑Training (MPT), a framework that jointly learns action policies and self‑orchestration in decentralized agent swarms. Agents bid for participation, and these bids determine who acts and how local training credit is assigned, allowing both individual capabilities and participation patterns to adapt together without a central coordinator. Experiments on embodied crafting, robot navigation, and deep research show that MPT outperforms monolithic post‑training and alternative coordination strategies, and that learned self‑orchestration improves scalability, robustness to agent dropout, and adaptation to related tasks.
The paper introduces SYNTRA, a self‑improving framework that transforms a small set of retrieval examples into reusable synthesis instructions for generating training data. By inferring implicit information needs and relevance criteria from an unlabelled corpus and refining instructions through feedback, SYNTRA produces effective training data across 24 datasets and 15 domains, improving retrieval performance even for strong embedding models. Experiments show that SYNTRA’s self‑improvement yields better results than human calibration on four tasks and scales effectively, with continued gains on MS MARCO across a 32‑fold range of synthetic data sizes.
This paper introduces an autonomous optical wireless receiver that uses on-device TinyML to diagnose faults in real time. Built on an ESP32‑S3 microcontroller, it extracts key channel metrics—BER, signal voltage, SNR, and ambient light—and classifies four operational conditions with a 74.68% accurate INT8 TensorFlow Lite Micro model. The approach eliminates external processing, enabling deterministic, low‑latency fault detection for optical IoT nodes.
Intern-Decision is a family of multimodal decision models (0.8B, 2B, and 4B parameters) that output structured choices and candidate probabilities directly from text and optional images, avoiding the latency of free‑form responses. The models use an autoregressive masked language modeling objective, placing one decision placeholder per field in a complete assistant JSON skeleton, and evaluate all fields in a single causal forward pass. The 4B variant achieves 90.02 % unweighted average accuracy across seven evaluation suites, with a mean latency of 44.16 ms on an RTX 4090 for a 289‑token request, and shows improved calibration and accuracy when combined with temperature scaling and confidence‑based handoff.
The paper introduces On-Policy Reminder Distillation (OPRD), a method that enhances long-horizon coding agents by combining multi-scope reminder supervision with outcome-based routing. OPRD uses successful rollouts as autoregressive self-imitation targets and failed rollouts for reminder-conditioned reverse-KL supervision from a teacher that re-evaluates student prefixes. The approach improves task resolution rates by 11.3 points and cuts error events by 50.2% on several SWE-bench benchmarks.
SkillNEO introduces a non‑elitist evolutionary framework for refining agent skills, allowing lower‑reward versions to receive further edits and preserve useful partial repairs. By separating the selection of a version to use from the selection of versions to improve, the method identifies which steps to keep and which failures to fix, even revisiting earlier versions with new evidence. Experiments on 86 SkillsBench tasks across eight domains show significant reward gains for GPT‑5.4 Mini and substantial improvements when the evolved skills are reused by stronger models such as GPT‑5.5 and GPT‑5.6 Luna.
WEAVE is a framework that converts verified pull‑request histories into executable, dependency‑structured tasks for code agents. It organizes related PRs into bounded directed acyclic graphs (DAGs) using repository‑calibrated typed change loci, creating unified tasks that reflect real‑world code evolution. The framework also introduces Graph Credit, a deterministic, test‑grounded metric that measures partial progress consistent with dependencies, providing dense, verifiable rewards for reinforcement learning.
AlphaDiana is a measurement system designed to evaluate agents in verifiable reasoning tasks by providing modular interfaces for datasets, agents, environments, and scorers. It maintains harness-native control loops while separately managing file reuse and memory retention, enabling both macro-level analyses of entire agents and micro-level interventions on components such as tools, skills, and memory. Experiments on open-source models and harnesses across multiple benchmarks reveal that harnesses and component interventions do not consistently improve reasoning accuracy, and that token usage metrics must be interpreted alongside correctness.
IneqAgent is a new framework that combines large language model reasoning with symbolic computation to prove inequalities in Lean4. It uses lemma-level Monte Carlo Tree Search to coordinate lemma construction, goal reformulation, and backend selection, employing sum‑of‑squares, Bernstein, and successive difference‑substitution backends to generate exact certificates. In tests on four benchmarks, IneqAgent with DeepSeek‑V4‑Flash solved 175 of 186 problems (94.09%) and produced Lean‑verified strengthenings of Dittert’s conjecture and a certification of the Holens–Djokovic $(4,4)$ inequality.
EvolvingBrain introduces an adaptive skill system that enables multi-modal agents to self‑evolve in open worlds. The system comprises Stratified Skill Self‑Evolution, which extracts procedural and problematic patterns from trajectories to build a multi‑granularity skill architecture, and State‑Aware Adaptive Execution, which uses structured memory to guide skill fusion based on the agent’s current state. Experiments on VAB‑Minecraft and DiscoveryWorld show improved success rates and generalization, with code available for public release.
The paper examines the vulnerability of skill-policy co‑evolution in self‑improving agents, showing that policies become fragile when the skill bank changes after training. It introduces CODA, a deployment‑oriented evaluation protocol that simulates skill drift and measures worst‑case success, consistency under meaning‑preserving edits, and resistance to corrupted skills. Across multiple methods and benchmarks, CODA reveals that worst‑case performance can drop 5–13 points, erasing a large portion of the gains achieved during co‑evolution.
arXiv:2610. 21055v1 Announce Type: new Abstract: Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence.
Sciπvot is a 248‑case benchmark that evaluates how well large language model (LLM) agents perform scientific judgment during autonomous research. The benchmark is built from six months of real‑world experiments where senior researchers ran LLM agents on 26 projects across six scientific fields and repeatedly corrected the agents’ decisions. In each case, the agent must choose between its original plan and a human‑corrected alternative using only the project record available before the correction, allowing a direct comparison of LLM and human performance under varying reasoning effort and record access.
The paper studies how large language models (LLMs) plan under partial observability, comparing aggressive planners that commit to long plans based on unobserved assumptions with conservative planners that only use observed information. Both strategies, whether the LLM outputs a plan directly or generates a PDDL formulation, achieve similar accuracy across three LLMs on two domains. Conservative planners reduce failures but increase API calls, while aggressive planners fail more often but replan efficiently when priors are informative. The study also finds that LLMs treat priors almost as facts, risking failure if priors are unreliable, and proposes hybrid strategies with tunable hyperparameters to balance safety and efficiency.
MemScope introduces a framework for evolving large language model (LLM) agent memory architectures by using localized code repair guided by detailed execution traces and runtime memory states. The Patch Agent identifies precise repair targets and applies bounded, in-place edits while respecting provider identity and memory compatibility constraints. Experiments on GAIA, xBench, and ALFWorld show that MemScope improves accuracy and correct answer rates, achieving higher success rates with a modest median code‑change ratio of 4.4%.
The paper introduces MINE-SAMPLE-RENDER (MSR), a test-time method that enhances diversity in generative models by externalizing it into a corpus-derived sampling process. MSR first mines an interpretable categorical schema from a corpus of real artifacts and uses the resulting statistics to sample attribute configurations. These configurations are then fed into a generative model, which acts purely as a conditional renderer, producing outputs that better align with the corpus distribution and exhibit greater diversity across several application domains.