Searcharxiv⌕ Search

arXiv subjects

Ziyang Luo

Publications and source records attributed to Ziyang Luo.

At least 19 recordsLinked to original sources

When the World Is the Test Suite: Agentic Intelligence for the Physical World

Coding agents built to repair software now write, test, and repair programs for robots, laboratories, and industrial controllers. Unlike in software engineering, no test suite comes with the code. Whether a program works must be established by running it on a robot, in a physics simulator, or through the policies it helps to train, and such evidence is costly to obtain and can itself be wrong. As these agents take on more physical tasks, how their programs are judged matters as much as how they are written. This paper examines this emerging area, agentic intelligence for the physical world, by treating the world as the program's test suite. We organize the discussion around three questions: how agents write programs for the physical world, how these programs are tested, and how agents learn from the results. We first trace how language models for robots and software engineering agents came together, ask which software testing practices carry over to physical tests, and describe each program by where it sits relative to the actuator and by what judges it. For writing, we cover programs that act as robot policies and programs that support policy learning by producing rewards, tasks, and training worlds. For testing, we examine what physical trials cost, what traces, learned judges, simulators, and formal checks can establish in their place, and what safety requires beyond task success. For learning, we discuss how tested programs are retained as experience, in skill libraries and in model weights. Finally, we turn to benchmarks and evaluation practice and outline open challenges. Writing programs is becoming cheap, while trustworthy evidence about what they do remains costly. Progress will therefore depend on a physical counterpart of continuous integration: verification cheap enough to run often, reliable enough to trust under search, and maintained alongside the experience it validates.

cs.RO↗

You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model

A frozen language model under-uses the evidence in its hidden states and answers confidently when the context contains no answer. We fix both on the residual stream without changing any weight. For accuracy, a small trained network (the probe) adds an input-dependent vector to the residual at several middle layers; it raises Qwen2.5-1.5B on $α$NLI from 51% to 73% accuracy. For abstention, a fixed vector (the direction), the mean hidden state of contexts that contain the answer minus that of contexts that do not, scores the hidden state; a low score abstains. On four QA datasets with human answerability labels (Qwen2.5-7B), moving the direction to a dataset it was not fit on costs under 3 points of decision accuracy (from 0.83 in domain) on average and still beats asking the model whether its context suffices; the same judgment trained as a classifier into the probe's activation edit or into LoRA weights loses 16 and 22 points and ends below asking. We set out to put both functions on a single forward pass, and found that the probe's edit changes the very hidden state the direction scores, degrading it. A small map -- trained without labels, only to reconstruct the unedited hidden state from the edited one -- removes 63% of that interference on 1.5B. Refit on the hidden-state pairs of LoFiT or LoReFT, the same recipe repairs those methods too: the steering module is replaceable. The resulting system, YOPO, answers and abstains in one forward pass. On $α$NLI it raises three-way accuracy from 0.38 to about 0.75 on 1.5B, and stays within 0.02 of a two-pass reference on 1.5B, 3B and 7B, at half the cost. We also observe improvements on the four QA datasets, where answerability is human-labeled and answers are generated as free-form text. Near-parity with the two-pass reference holds across nine backbones spanning six model families without backbone-specific hyperparameter tuning.

cs.CL↗

Toward Recursive Self-Improvement in Physical AI

Recursive self-improvement (RSI) describes a system that improves not only its behavior but also the process by which it improves. Recent digital agents test this idea by rewriting their own code and checking each version with executable tests. Bringing these mechanisms into the physical world requires learning from partial observations through costly trials, while accounting for changing operating conditions and potentially irreversible consequences. We position RSI within robot learning and trace pathways from improving robot capabilities to improving the processes that generate them. We examine how physical experience can strengthen these improvement mechanisms and when the resulting gains can support further self-improvement of the mechanisms themselves. Gains carry forward only when feedback addresses remaining learning needs, the receiving component can use the change, its value survives changes in tasks or hardware, and continued practice remains safe and affordable. Existing studies, concentrated in manipulation, support behavioral improvement, benefits to later learning, and procedure improvement. Some systems also change the improver's knowledge or tools, but the physical studies reviewed here have yet to isolate gains attributable to recursion. We synthesize these findings into evaluation principles and research directions for machines that, trial by trial, safely improve the way they improve.

cs.RO↗

From Evidence to Action: How Tool-Using Agents Fail

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

cs.CL↗

Chain of Modality: From Static Fusion to Dynamic Orchestration in Omni-MLLMs

Omni-modal Large Language Models (Omni-MLLMs) are designed to reason over diverse sensory streams within a unified model. However, we find that multimodal reasoning is not solely determined by sensory evidence, but also by how modality information is organized. Through systematic analysis across diverse reasoning scenarios, we show that different modality organizations exhibit distinct advantages and limitations, and no single topology is universally optimal. Rather than proposing a general-purpose improvement to Omni-MLLM accuracy, we ask a narrower question: can these topology-induced failures be systematically identified and corrected? Motivated by this, we propose Chain of Modality (CoM), a framework that dynamically reorganizes multimodal topology during inference and learns adaptive organization strategies. Experiments across five benchmarks, diverse architectures, and model scales show that CoM reliably recovers topology-sensitive failures, through both training-free planning and lightweight Planner-SFT, while preserving performance on the full benchmark.

cs.CV↗

Decision Readouts for Text-Mediated Video Anomaly Detection: An Exploratory Evaluation of Jev and Qwen

How much does the decision readout matter when video-derived textual evidence is held fixed? We evaluate Jev typed decisions and three Qwen readouts on a sparse development sample of 40 videos and 400 target anchors from UCF-Crime and XD-Violence, each presented as a summary and ordered captions. Each dataset contributes 20 source groups and 200 anchors, including only 10 and 37 positives, respectively. The original five-backend pilot requested 4,000 predictions; Jev Choice returned 776 valid responses out of 800 under the study's strict numerical policy, blocking its full-coverage quality comparison. On XD captions, Jev Noul achieved 75.99% average precision versus 48.47% for Qwen generated probability and 57.81% for the stronger local ordinal-likelihood expectation. The latter paired difference was 18.18 percentage points (95% source-group bootstrap interval 5.53-31.50). UCF did not show a corresponding advantage: caption ROC-AUC was 52.26% for Noul and 65.95% for ordinal likelihood. Both probability readouts had higher, hence worse, UCF Brier scores than the evaluation-prevalence reference of 0.0475. We additionally audit historical LAVAD scores at exactly matched anchors and distinguish response structure from numerical consistency. A binary-likelihood control is missing. These exploratory offline results characterize ranking, probability quality and interface failures; they establish neither a causal typed-interface benefit nor general superiority, calibration or end-to-end acceleration.

cs.AI↗

Calibration, Not Answer Selection: Distilling Internal Confidence in Reasoning Models

Reinforcement learning with binary correctness rewards trains correctness, not calibrated confidence. The confidence that reasoning models verbalize is systematically overconfident, and the problem is not merely one of scale: verbalized confidence tracks how willing a model is to commit to an answer, not how likely the answer is to be right. Post-hoc rescaling therefore fits one distribution but rarely transfers. We look inside the model instead. On factual question answering, a linear probe on the hidden state between the chain of thought and the answer is substantially better calibrated: its expected calibration error is 5 to 38 times lower than that of the verbalized score across four benchmarks and two model families. However, when used to pick among N sampled answers, that same probe nearly ties majority voting yet falls far short of the oracle. Internal states answer "how certain am I" well and "which answer is right" poorly, so the signal should be reported as a confidence rather than used to select answers. As a result, we introduce probe-guided self-distillation (Probe-SD): score a model's own sampled traces with the probe, overwrite the confidence each trace states, and finetune the base checkpoint of the same family, so nothing but the model itself remains at test time. On Qwen3-14B, Probe-SD cuts ECE from 0.178 to 0.024 in-domain and from 0.542 to 0.113 out-of-domain, where it also beats post-hoc recalibration and self-consistency distillation. The resulting confidence is well-calibrated and useful for weighted voting, behaviors previously attributed to online RL, here obtained with supervised finetuning alone.

cs.CL↗

MCP-Universe RL: A Framework for Training MCP Tool-Use Agents via Reinforcement Learning

Reinforcement learning (RL) has become an effective way to improve the tool-use ability of large language models (LLMs), but most existing RL frameworks stop at the policy update. For every new domain, the user is left with two hard systems problems: standing up an isolated environment for each of hundreds of concurrent trajectories and connecting it to training, and scheduling the rollout so that the GPU stays busy across long, multi-turn episodes that spend much of their time stalled on slow tool calls. We present MCP-Universe RL (MCP-U RL), an open-source framework that takes over both. It uses the Model Context Protocol (MCP) as the interface to the environment, so any tool already exposed as an MCP server plugs into training with no RL-specific integration code. It builds the two missing layers once and reuses them across domains: an environment-orchestration layer that provisions, isolates, and recycles the MCP environments over a pluggable container backend, and a rollout-orchestration layer whose staged pipeline overlaps trajectories to keep the GPU busy while episodes wait on tools. A backend-agnostic training layer then applies the update through an existing RL backend, with veRL and slime integrations. With one configuration, changing only the task specification, we train software-engineering, deep-research, and general tool-use agents on gpt-oss-20b and improve task reward in all three.

cs.AI↗

StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents

Computer-use agents are usually improved by strengthening perception: better models for reading a screenshot and choosing where to click. Yet a screenshot is only a lossy rendering of the underlying program state, e.g., the files, application backends, and DOM that hold the task data. Different states can produce the same pixels, while code can inspect and modify that state directly. StateAct is a code-first, multi-agent harness built around this distinction. Its main agent works directly with program state by using code, while a dedicated GUI subagent handles screenshot-and-click interaction on the few subgoals that need it, just 28 of 108 tasks and 1.1% of main-agent steps. The same direct access to program state also supports verification: an independent finish gate double-checks the saved result for structural failures, e.g., output that is missing, unsaved, or written to the wrong path. To stay on track over hundreds of steps, the main agent hands subgoals to fresh subagents, keeping its own context focused. On OSWorld 2.0, StateAct lifts Claude Opus 4.8 from 20.6% to 26.9% on binary success, and from 54.8% to 61.6% on partial success, at ~ 9x lower cost per task than the same model driven by screenshots alone; a code-only variant with no GUI subagent reaches only 45.9% partial, below that screenshot-based baseline's 54.8%. In general, grounding action, verification, and memory in state, what we call state-grounding, shifts the main bottleneck from perception toward reasoning: failures depend more on what the agent thinks than on what it sees.

cs.SE↗

From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

Large Language Model (LLM)-based agents have fundamentally reshaped artificial intelligence by integrating external tools and planning capabilities. While memory mechanisms have emerged as the architectural cornerstone of these systems, current research remains fragmented, oscillating between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis and a coherent evolutionary perspective. To bridge this gap, this survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing the development process into three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). We first formally define these three stages before analyzing the three core drivers of this evolution: the necessity for long-range consistency, the challenges in dynamic environments, and the ultimate goal of continual learning. Furthermore, we specifically explore two transformative mechanisms in the frontier Experience stage: proactive exploration and cross-trajectory abstraction. By synthesizing these disparate views, this work offers robust design principles and a clear roadmap for the development of next-generation LLM agents.

cs.AI↗

DiffCoT: Diffusion-styled Chain-of-Thought Reasoning in LLMs

Chain-of-Thought (CoT) reasoning improves multi-step mathematical problem solving in large language models but remains vulnerable to exposure bias and error accumulation, as early mistakes propagate irreversibly through autoregressive decoding. In this work, we propose DiffCoT, a diffusion-styled CoT framework that reformulates CoT reasoning as an iterative denoising process. DiffCoT integrates diffusion principles at the reasoning-step level via a sliding-window mechanism, enabling unified generation and retrospective correction of intermediate steps while preserving token-level autoregression. To maintain causal consistency, we further introduce a causal diffusion noise schedule that respects the temporal structure of reasoning chains. Extensive experiments on three multi-step CoT reasoning benchmarks across diverse model backbones demonstrate that DiffCoT consistently outperforms existing CoT preference optimization methods, yielding improved robustness and error-correction capability in CoT reasoning.

cs.CL↗

GPA: Learning GUI Process Automation from Demonstrations

GUI Process Automation (GPA) is a lightweight but general vision-based Robotic Process Automation (RPA), which enables fast and stable process replay with only a single demo. Addressing the fragility of traditional RPA and the non-deterministic risks of current vision language model-based GUI agents, GPA introduces three core benefits: (1) Robustness via Sequential Monte Carlo-based localization to handle rescaling and detection uncertainty; (2) Deterministic and Reliability safeguarded by readiness calibration; and (3) Privacy through fast, fully local execution. This approach delivers the adaptability, robustness, and security required for enterprise workflows. It can also be used as an MCP/CLI tool by other agents with coding capabilities so that the agent only reasons and orchestrates while GPA handles the GUI execution. We conducted a pilot experiment to compare GPA with Gemini 3 Pro (with CUA tools) and found that GPA achieves higher success rate with 10 times faster execution speed in finishing long-horizon GUI tasks.

cs.CV↗

Probabilistic Concept Graph Reasoning for Multimodal Misinformation Detection

Multimodal misinformation poses an escalating challenge that often evades traditional detectors, which are opaque black boxes and fragile against new manipulation tactics. We present Probabilistic Concept Graph Reasoning (PCGR), an interpretable and evolvable framework that reframes multimodal misinformation detection (MMD) as structured and concept-based reasoning. PCGR follows a build-then-infer paradigm, which first constructs a graph of human-understandable concept nodes, including novel high-level concepts automatically discovered and validated by multimodal large language models (MLLMs), and then applies hierarchical attention over this concept graph to infer claim veracity. This design produces interpretable reasoning chains linking evidence to conclusions. Experiments demonstrate that PCGR achieves state-of-the-art MMD accuracy and robustness to emerging manipulation types, outperforming prior methods in both coarse detection and fine-grained manipulation recognition.

cs.CV↗

Towards Comprehensive Stage-wise Benchmarking of Large Language Models in Fact-Checking

Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs' factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs' factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

cs.CL↗

TAViS: Text-bridged Audio-Visual Segmentation with Foundation Models

Audio-Visual Segmentation (AVS) faces a fundamental challenge of effectively aligning audio and visual modalities. While recent approaches leverage foundation models to address data scarcity, they often rely on single-modality knowledge or combine foundation models in an off-the-shelf manner, failing to address the cross-modal alignment challenge. In this paper, we present TAViS, a novel framework that \textbf{couples} the knowledge of multimodal foundation models (ImageBind) for cross-modal alignment and a segmentation foundation model (SAM2) for precise segmentation. However, effectively combining these models poses two key challenges: the difficulty in transferring the knowledge between SAM2 and ImageBind due to their different feature spaces, and the insufficiency of using only segmentation loss for supervision. To address these challenges, we introduce a text-bridged design with two key components: (1) a text-bridged hybrid prompting mechanism where pseudo text provides class prototype information while retaining modality-specific details from both audio and visual inputs, and (2) an alignment supervision strategy that leverages text as a bridge to align shared semantic concepts within audio-visual modalities. Our approach achieves superior performance on single-source, multi-source, semantic datasets, and excels in zero-shot settings.

cs.CV↗

AURORA:Augmented Understanding via Structured Reasoning and Reinforcement Learning for Reference Audio-Visual Segmentation

Reference Audio-Visual Segmentation (Ref-AVS) tasks challenge models to precisely locate sounding objects by integrating visual, auditory, and textual cues. Existing methods often lack genuine semantic understanding, tending to memorize fixed reasoning patterns. Furthermore, jointly training for reasoning and segmentation can compromise pixel-level precision. To address these issues, we introduce AURORA, a novel framework designed to enhance genuine reasoning and language comprehension in reference audio-visual segmentation. We employ a structured Chain-of-Thought (CoT) prompting mechanism to guide the model through a step-by-step reasoning process and introduce a novel segmentation feature distillation loss to effectively integrate these reasoning abilities without sacrificing segmentation performance. To further cultivate the model's genuine reasoning capabilities, we devise a further two-stage training strategy: first, a ``corrective reflective-style training" stage utilizes self-correction to enhance the quality of reasoning paths, followed by reinforcement learning via Group Reward Policy Optimization (GRPO) to bolster robustness in challenging scenarios. Experiments demonstrate that AURORA achieves state-of-the-art performance on Ref-AVS benchmarks and generalizes effectively to unreferenced segmentation.

cs.CV↗

Semantic Temporal Single-photon LiDAR

Temporal single-photon (TSP-) LiDAR presents a promising solution for imaging-free target recognition over long distances with reduced size, cost, and power consumption. However, existing TSP-LiDAR approaches are ineffective in handling open-set scenarios where unknown targets emerge, and they suffer significant performance degradation under low signal-to-noise ratio (SNR) and short acquisition times (fewer photons). Here, inspired by semantic communication, we propose a semantic TSP-LiDAR based on a self-updating semantic knowledge base (SKB), in which the target recognition processing of TSP-LiDAR is formulated as a semantic communication. The results, both simulation and experiment, demonstrate that our approach surpasses conventional methods, particularly under challenging conditions of low SNR and limited acquisition time. More importantly, our self-updating SKB mechanism can dynamically update the semantic features of newly encountered targets in the SKB, enabling continuous adaptation without the need for extensive retraining of the neural network. In fact, a recognition accuracy of 89% is achieved on nine types of unknown targets in real-world experiments, compared to 66% without the updating mechanism. These findings highlight the potential of our framework for adaptive and robust target recognition in complex and dynamic environments.

eess.IV↗

Saliency-R1: Incentivizing Unified Saliency Reasoning Capability in MLLM with Confidence-Guided Reinforcement Learning

Although multimodal large language models (MLLMs) excel in high-level vision-language reasoning, they lack inherent awareness of visual saliency, making it difficult to identify key visual elements. To bridge this gap, we propose Saliency-R1, the first unified MLLM framework that jointly tackles three representative and heterogeneous saliency tasks: Salient Object Detection (SOD), Salient Instance Segmentation (SIS), and Co-salient Object Detection (CoSOD), enhancing the model's capacity for saliency reasoning. We introduce a textual interface with structured tags ( , ) to encode region- and instance-level referring expressions, enabling a single referring segmenter to produce task-appropriate masks. To train the MLLM efficiently, we propose Confidence-Guided Policy Optimization (CGPO), a novel single-sample reinforcement learning algorithm. CGPO improves on GRPO by replacing group-normalized advantages with a per-sample signal based on reward-confidence discrepancy, thereby reducing computational waste, mitigating signal dilution, and lowering training overhead. Our model exceeds or matches the performance of robust open/closed-source MLLMs and specialized state-of-the-art methods across all three tasks, demonstrating the efficacy of our framework in saliency reasoning.

cs.CV↗