arXiv subjects

Jiankai Sun

Publications and source records attributed to Jiankai Sun.

At least 19 recordsLinked to original sources

arXiv2026

EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics

We study coding agents for long-horizon, dexterous robotics and ask whether their solutions can provide scalable supervision for learning general robot policies. To test this, we develop EMBODIEDSWE-BENCH, a simulation benchmark for coding agents spanning contact-rich manipulation, deformable objects, and long-horizon tasks requiring up to half an hour of continuous interaction. We find that frontier coding agents can solve complex long-horizon tasks and transfer prior solutions across both tasks and embodiments. We also design supporting tools that help agents more effectively solve these tasks. However, the resulting solutions require substantial iterative interaction and are typically specialized to individual task instances. We therefore introduce EMBODIEDSWE-GEN, which expands a single solution from coding agent into large diverse trajectories for training a VLA. VLA performance improves with more generated demonstrations, and agent-aided diversification improves generalization to held-out task variations. We also show that a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on real robot. Together, our framework uses coding agents to solve complex robotics tasks and turn verified solutions into scalable supervision for robot policies.

cs.RO↗

arXiv2026

ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience-a challenge we call the scientific experience bottleneck. We introduce ScienceIDE, infrastructure for turning the world's scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, we train PhAI-IDE-72B, PhAI-IDE-9B, and PhAI-IDE-4B. The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity's scientific software a shared substrate for developing scientific intelligence. Code: https://github.com/aitofound/ScienceIDE

cs.CL↗

arXiv2026

Harbor Adapters and Harbor-Index: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation

Evaluating agents on the growing number of agentic benchmarks is challenging because they often require complex environments and agent integrations. We introduce Harbor Adapters, a unified evaluation infrastructure for agentic benchmarks. Our work makes three contributions. First, we develop benchmark adapters that port more than 80 benchmarks to evaluate arbitrary agents, and validate them through rigorous code review and parity experiments. Second, we conduct a large-scale evaluation of 8 models spanning capability tiers across 54 benchmarks; every model is run with Terminus-2 and with one of 3 native harnesses. This enables a broader analysis of agent capabilities and failure modes than was previously possible. Third, we introduce Harbor-Index, a curated set of 82 difficult, diverse, and high-quality tasks spanning 29 benchmarks, refined from the adapted suite through difficulty filtering, AI and human audit, and an audit-and-fix loop. Harbor-Index preserves the challenge and breadth of large-scale agentic evaluations while being affordable to run; no evaluated model-harness configuration exceeds 30% pass rate, and the strongest (GPT-5.5 with Codex) reaches 28.0%. We release the adapters, evaluation results, in-depth analysis, and Harbor-Index as open-source artifacts to support more reliable and comprehensive evaluation of language-model agents.

cs.AI↗

arXiv2026

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

LM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task. Existing defenses rely largely on task-specific patches, prompt instructions, or post-hoc detectors. They do not provide reusable evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation's reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims. We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23-94% to 77-100%, same-vector coverage from 16-56% to 43-78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

cs.CR↗

arXiv2026

Scaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence

Recent advances in large reasoning models (LRMs) have shown that reinforcement learning with verifiable rewards (RLVR) can substantially improve reasoning in mathematics and code, where outcomes can be checked automatically. Extending this progress to open-ended and agentic tasks remains difficult because reliable rewards are harder to obtain and direct human supervision cannot keep pace with the scale and complexity of model-generated experience. This paper studies how LRMs can continue to improve as human supervision gradually recedes from the learning loop. We examine two connected dimensions of this problem. The reward axis traces the development from per-instance human judgments to reusable verifiers and rewards that operate even without human feedback. The experience axis examines how learning can progress from human-curated tasks and environments toward self-generated curricula, constructed environments, and autonomous co-evolution. We connect these dimensions through a five-level ladder from L0 to L4 that identifies which parts of the learning process remain under continued human control. Our analysis further highlights the risks introduced by increasingly autonomous rewards and experience generation, including reward hacking, feedback drift, curriculum collapse, and environment errors. Consequently, we also provide the evaluation around three complementary objects: policy capability, feedback fidelity, and experience quality. This analysis provides a structured account of current approaches to scaling LRMs beyond human supervision and the open problems involved in developing self-sustaining learning systems toward superintelligence. Furthermore, we maintain a continuously updated \href{https://github.com/visitworld123/Awesome-Scaling-LRM-Beyond-Human-Supervision}{GitHub repository} to track the latest advances.

cs.AI↗

arXiv2026

Cubit: Token Mixer with Kernel Ridge Regression

Since its introduction in 2017, the Transformer has become one of the most widely adopted architectures in modern deep learning. Despite extensive efforts to improve positional encoding, attention mechanisms, and feed-forward networks, the core token-mixing mechanism in Transformers remains attention. In this work, we show that the attention module in Transformers can be interpreted as performing Nadaraya-Watson regression, where it computes similarities between tokens and aggregates the corresponding values accordingly. Motivated by this perspective, we propose Cubit, a potential next-generation architecture that leverages Kernel Ridge Regression (KRR), while the vanilla Transformer relies on Nadaraya-Watson regression. Specifically, Cubit modifies the classical attention computation by incorporating the closed-form solution of KRR, combining value aggregation through kernel similarities with normalization via the inverse of the kernel matrix. To improve the training stability, we further propose the Limited-Range Rescale (LRR), which rescales the value layer within a controlled range. We argue that Cubit, as a KRR-based architecture, provides a stronger mathematical foundation than the vanilla Transformer, whose attention mechanism corresponds to Nadaraya-Watson regression. We validate this claim through comprehensive experiments. The experimental results suggest that Cubit may exhibit stronger long-sequence modeling capability. In particular, its performance gain over the Transformer appears to increase as the training sequence length grows.

cs.LG↗

arXiv2026

PRIMAL3: Pathfinding via Reinforcement and Imitation Multi-Agent Learning - Leveraging LaCAM3

We present PRIMAL3, an ultra-large-scale learning-based framework for multi-agent pathfinding (MAPF) that integrates reinforcement learning, topology-aware communication, LaCAM3-guided training, and PIBT-based action refinement. PRIMAL3 targets failures at topologically critical states, where agents must coordinate decisively around bottlenecks, dead ends, and persistent conflicts. Each agent is represented using features derived from cut vertices, dead-end regions, shortest-path distances, and blocking estimates. Two complementary graphs capture agent interactions: a same-direction following graph propagates multihop context along compatible paths, while a different-direction conflict graph differentiates agents competing for shared space through masked attention and relative features. During training, we propose to let policy entropy identify uncertain agents, for which LaCAM3 provides confidence-triggered action interventions and label-smoothed imitation targets. During execution, a priority-aware PIBT module refines the proposed joint actions using persistent, learned, and distance-aware priorities together with policy-aware fallback preferences while maintaining collision-free execution. The resulting framework combines learned exploration with structured expert guidance without requiring LaCAM3 at inference. Experiments demonstrate that PRIMAL3 substantially outperforms state-of-the-art learning-based baselines and scales to ultra-large instances with up to city-level 100,000 agents. Real-world experiments further demonstrate the feasibility of deploying PRIMAL3 on physical robotic systems and ablation studies validate the individual contributions the components we proposed. Project page: https://marmotlab.github.io/PRIMAL3/

cs.RO↗

arXiv2026

FARS: A Fully Automated Research System Deployed at Scale

Recent automated research systems show that language-model agents can generate hypotheses, run experiments, and write complete manuscripts, but most evidence still comes from selected examples, human-framed topics, or a few pre-defined research tasks. We present FARS (Fully Automated Research System), a fully automated AI-for-AI research system designed to operate across research topics at scale. FARS autonomously generates and advances projects through ideation, planning, experimentation, and writing, using stage-specific agents coordinated through a shared workspace that records proposals, code, logs, results, and manuscripts. In its first public deployment, FARS produced 166 complete research papers spanning 67 fine-grained AI/ML topics while preserving intermediate artifacts as an auditable corpus rather than a curated set of successes. We evaluate this corpus with 282 structured reviews from volunteer reviewers covering 140 papers, including overall ratings, sub-scores, integrity checks, and LLM-use disclosure. The reviews indicate that FARS can produce review-worthy and occasionally strong AI/ML research artifacts in a large-scale public deployment, while also exposing recurring failure modes in narrow experimental scope, methodological limitations, and integrity issues.

cs.AI↗

arXiv2026

PERRY: Policy Evaluation with Confidence Intervals using Auxiliary Data

Off-policy evaluation (OPE) methods estimate the value of a new reinforcement learning (RL) policy prior to deployment. Recent advances have shown that leveraging auxiliary datasets, such as those synthesized by generative models, can improve the accuracy of OPE methods. Unfortunately, such auxiliary datasets may also be biased, and existing methods for using data augmentation within OPE lack principled uncertainty quantification. In high stakes domains like healthcare, reliable uncertainty estimates are important for ensuring safe and informed deployment of RL policies. In this work, we propose two methods to construct valid confidence intervals for OPE with data augmentation. The first provides a confidence interval over $V^π(s)$, the policy value conditioned on an initial state $s$. To do so we introduce a new conformal prediction method suitable for Markov Decision Processes (MDPs) with continuous state spaces, extending prior work to higher-dimensional settings. Second, we consider the more common task of estimating the average policy performance over all initial states, $V^π$; we introduce a method that draws on ideas from doubly robust estimation and prediction powered inference. Across simulators spanning inventory management, robotics, healthcare, and a real healthcare dataset from MIMIC-IV, we find that our methods can effectively leverage auxiliary data and consistently produce confidence intervals that cover the ground truth policy values, unlike previously proposed methods. Our work enables a future in which OPE can provide rigorous uncertainty estimates for high-stakes domains.

cs.LG↗

arXiv2026

Pose-Agnostic Robotic Functional Grasping via Observation-Action Canonicalization

Functional robotic grasping requires a policy that generalizes across diverse object geometries and poses while maintaining task-specific contact precision. We study this challenge through mug-handle grasping, where thin handles, instance variation, and upright or inverted placements make both perception and control sensitive to object configuration. Grasp pose detection methods operate open-loop and are sensitive to estimation errors on thin handle structures. Learned visuomotor policies must implicitly learn to handle the coupled variation in visual appearance and action direction induced by different object placements, limiting generalization. We propose AnyMug, a canonicalized visuomotor reinforcement learning framework for functional grasping that trains a single closed-loop policy entirely in simulation and deploys it zero-shot on a real robot. AnyMug introduces observation-action canonicalization, which transforms both the depth observation and the predicted end-effector action into a shared object-centric frame. The policy therefore sees a consistent mug-centered view and emits actions in a canonical direction regardless of mug placement, allowing the same grasping behavior to be reused across configurations. A handle-aware reward further encourages precise approach, gripper alignment, and opposing-finger placement, while a pose curriculum and domain randomization improve training stability and sim-to-real transfer. In simulation, AnyMug achieves over 93% success rate on both unseen upright and inverted mugs and transfers zero-shot to a real Franka Panda, reaching 80% success rate on 5 held-out physical mugs across both pose categories.

cs.RO↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2026

SARM2: Multi-Task Stage Aware Reward Modeling for Self Improving Robotic Manipulation

Fine-tuning vision-language-action (VLA) policies for long-horizon manipulation still relies heavily on behavior cloning, which requires costly high-quality demonstrations and keeps policies near the demonstration distribution. Reward models can reduce this dependence by reweighting demonstrations and providing dense supervision for on-robot reinforcement learning (RL), but they must be dense, accurate, and general. Existing methods fall short: task-specific stage-aware models are accurate but require per-task annotations, while general vision-language-model (VLM) reward models are broadly applicable but too coarse for fine-grained long-horizon progress. We introduce RM, a multi-task stage-aware reward model that combines an action-primitive-based stage estimator with a multi-gate Mixture-of-Experts (MMoE) value head to produce dense per-step rewards across manipulation tasks. Building on RM, we further propose SPIRAL (Self-Policy Improvement via Reward-Aligned Learning), an on-policy reward-guided framework that improves VLA policies from cheap autonomous rollouts. On a 10-task benchmark, RM reduces value-estimation MSE by 80% over the strongest baselines; when used in SPIRAL, it improves task success from around 50% to near-perfect performance on Folding Shorts (58% to 100%) and Cleaning Whiteboard (50% to 90%), showing that high-quality dense rewards are key to a stable robot data flywheel. Project website: https://qianzhong-chen.github.io/sarm2.github.io/.

cs.RO↗

arXiv2026

SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?

AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments. Yet current agent benchmarks largely evaluate short-form tasks, such as single pull requests, small tickets, or 5-10 minute exercises, limiting our ability to measure agents' capabilities in planning, long-context understanding, and memory use. We introduce SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains. Each task consists of a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Logged agent attempts average 27.2M total tokens, making SWE-Marathon substantially longer-horizon than existing SWE and command-line agent benchmarks. Current frontier coding agents solve fewer than 30% of tasks. Failures often arise from poor self-verification, self-reported infeasibility, and premature termination. We also observe reward-hacking behavior in 13.8% of rollouts, where agents attempt to exploit the environment or verifier to bypass the intended workflow. SWE-Marathon includes adversarial review of test suites and execution environments, as well as multi-layer checks designed to prevent shortcut solutions. We release SWE-Marathon, evaluation code, and agent trajectories at https://swe-marathon.org/.

cs.SE↗

arXiv2026

Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning

This paper investigates the entropy dynamics of Chain-of-Thought (CoT) and uncovers a consistent two-phase structure: an Uncertainty Region of exploration transitioning sharply to a Confidence Region of convergence. We demonstrate that the Confidence Region possesses two critical properties: 1) High Reliability -- answers in the confidence region become highly accurate and stable, and 2) High Redundancy -- models generate unnecessary tokens long after reaching the correct answer. These properties unlock more efficient and reliable inference strategies: 1) Early Exit leverages reliability and redundancy to terminate computation safely when returns diminish, and 2)Test-Time Scaling uses the Confidence Region signal to prioritize converged trajectories. To operationalize these insights, we formulate Confidence Region detection as a sequential change-point detection problem, being the first to apply classical change-point methods to monitor CoT reasoning. Using the Cumulative Sum (CUSUM) algorithm, a statistically optimal change-point detector, we develop a training-free framework for real-time inference control. Experiments show our approach establishes a superior Pareto-frontier for early exit. CUSUM achieves 63.06% accuracy with 11.1% token reduction, outperforming DEER and Dynasor by 3.28% and 4.36% in accuracy respectively. For test-time scaling, CUSUM-weighted voting consistently outperforms self-consistency.

cs.CL↗

arXiv2026

LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World

Training vision-language-action (VLA) policies for humanoid loco-manipulation is constrained by the high cost and complexity of collecting human teleoperation demonstrations. VLA policies fine-tuned in simulators have, until now, failed to transfer effectively in humanoid loco-manipulation tasks. We present LEGS (Loco-manipulation via Embodied Gaussian Splatting), a hybrid simulator that composites a mesh foreground (robot, objects, props) over a photorealistic 3D Gaussian Splatting (3DGS) background reconstructed from a handheld scene capture. LEGS uses a procedural motion-primitive generator to synthesize labeled demonstrations at scale without human teleoperation, and a deterministic two-stage color calibration to align the rendered 3DGS image to the robot's deployment camera. On a Unitree G1 humanoid robot, across three pick-and-place tasks of increasing whole-body difficulty and three VLA backbones (psi_0, pi_0.5, GR00T N1.6), a policy trained purely on LEGS data matches or exceeds one trained on human teleoperation demos on every experiment. It also outperforms a mesh-only simulation baseline that ablates the effect of the 3DGS background, showing that photorealistic rendering is a key enabler for synthetic data transfer. Humanoid motion is recorded independently of scene appearance in LEGS, allowing the same auto-generated demonstrations to be re-rendered under new backgrounds and object meshes--covering a new scene at more than 15x lower cost than teleoperation--to augment training data for robustness to scene variations. Under combined object-and-scene appearance shift, the policy trained on re-rendered LEGS-AUG data maintains task success while the baseline trained on teleoperation data fails entirely. Our project page is located at https://legsvla.github.io/.

cs.RO↗

arXiv2026

GRaD-Nav++: Vision-Language Model Enabled Visual Drone Navigation with Gaussian Radiance Fields and Differentiable Dynamics

Autonomous drones capable of interpreting and executing high-level language instructions in unstructured environments remain a long-standing goal. Yet existing approaches are constrained by their dependence on hand-crafted skills, extensive parameter tuning, or computationally intensive models unsuitable for onboard use. We introduce GRaD-Nav++, a lightweight Vision-Language-Action (VLA) framework that runs fully onboard and follows natural-language commands in real time. Our policy is trained in a photorealistic 3D Gaussian Splatting (3DGS) simulator via Differentiable Reinforcement Learning (DiffRL), enabling efficient learning of low-level control from visual and linguistic inputs. At its core is a Mixture-of-Experts (MoE) action head, which adaptively routes computation to improve generalization while mitigating forgetting. In multi-task generalization experiments, GRaD-Nav++ achieves a success rate of 83% on trained tasks and 75% on unseen tasks in simulation. When deployed on real hardware, it attains 67% success on trained tasks and 50% on unseen ones. In multi-environment adaptation experiments, GRaD-Nav++ achieves an average success rate of 81% across diverse simulated environments and 67% across varied real-world settings. These results establish a new benchmark for fully onboard Vision-Language-Action (VLA) flight and demonstrate that compact, efficient models can enable reliable, language-guided navigation without relying on external infrastructure.

cs.RO↗

arXiv2026

Breaking Lock-In: Preserving Steerability under Low-Data VLA Post-Training

Have you ever post-trained a generalist vision-language-action (VLA) policy on a small demonstration dataset, only to find that it stops responding to new instructions and is limited to behaviors observed during post-training? We identify this phenomenon as lock-in: after low-data, supervised fine-tuning (SFT), the policy becomes overly specialized to the post-training data and fails to generalize to novel instructions, manifesting as concept lock-in (fixation on training objects/attributes) and spatial lock-in (fixation on training spatial targets). Many existing remedies introduce additional supervision signals, such as those derived from foundation models or auxiliary objectives, or rely on augmented datasets to recover generalization. In this paper, we show that the policy's internal pre-trained knowledge is sufficient: DeLock mitigates lock-in by preserving visual grounding during post-training and applying test-time contrastive prompt guidance to steer the policy's denoising dynamics according to novel instructions. Across eight simulation and real-world evaluations, DeLock consistently outperforms strong baselines and matches or exceeds the performance of a state-of-the-art generalist policy post-trained with substantially more curated demonstrations.

cs.RO↗

arXiv2026

ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive disclosure, and a meta prompt that coordinates behavior across services) and vary both to measure their separate and combined effects. Experiments across 6 models, 4 agent harnesses, and 33 conditions show that with full scaffolding, agents achieve task success rates of 39-64% but exhibit unsafe action rates of 7-33%. On OpenClaw, the top five models fall within a 10 percentage-point band on task success (53-63%), with unsafe action rates from 7% to 23% and no consistent ordering between the two metrics. We identify eight recurring patterns of unsafe behavior, including multi-step sandbox escalation and silent contract modification. We release the trajectories and future dataset at https://clawsbench.com.

cs.AI↗