Papers - 2026-09-11
吾能观之数千而面色如故
Grounding-driven Visual Reasoning#
WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data
Recent advances in wearable sensing enable continuous monitoring of physiological and behavioral signals, yet existing benchmarks rarely evaluate whether AI systems can reason over a real user's longitudinal wearable record. We introduce WearableQA, a benchmark comprising 4,084 10-option multiple-choice questions constructed from the wearable time series, blood biomarkers, and demographics of 200 real users, each with up to 500 days of daily measurements. WearableQA preserves authentic wearable distributions that include device noise and inter-individual variability. To evaluate distinct reasoning capabilities, we introduce 16 question types organized along two complementary axes: data versus health reasoning, which distinguishes computation over longitudinal measurements from physiological interpretation; and single- versus cross-signal reasoning, which separates reasoning about individual signals from the integration of multiple signals. To construct reliable questions at scale, we adopt a dual-grounding framework that combines literature-grounded physiological findings with statistically validated population-grounded physiological patterns. This enables the capture of meaningful relationships observed in real-world wearable data. Evaluation of 14 proprietary and open-source LLMs demonstrates that WearableQA effectively differentiates model capabilities, with performance ranging from 19.6% to 72.9% against a 10% chance baseline. Moreover, WearableQA remains far from solved: most models achieve accuracies below 60%. Overall, WearableQA provides a realistic and diagnostic benchmark for evaluating LLM reasoning over real-world wearable data.
3D LLM#
Show-Harness: Just a VLM Agent Can Play Robots
Foundation vision-language models (VLMs) exhibit broad intelligence about the world, yet translating this intelligence into robot control remains challenging. We present Show-Harness, an Embodied Harness that enables VLMs to "play" robots through a compact semantic interface linking intent to action. Show-Harness exposes discrete semantic action units that VLMs can naturally reason over, while embodiment-specific interpreters deterministically ground them into local robot actions, keeping the VLM directly responsible for fine-grained physical decisions. Through the same interface, Show-Harness demonstrates the feasibility of (1) directly unlocking closed-source frontier VLMs for zero-shot robot control, and (2) adapting small-scale open-source VLMs for low-cost deployment with just a few GPU-hours of fine-tuning. We further develop GUMI (GUI Manipulation Interface), which extends the same semantic action space to GUI-based demonstration collection, allowing humans and agents to "play" robots across embodiments without specialized teleoperation hardware. Extensive experiments show that Show-Harness-equipped VLM agents generalize robustly across tasks, embodiments, and environments, outperforming representative agentic and VLA paradigms. These results suggest that the right interface can unlock substantial embodied capability from foundation VLMs, without requiring additional model capacity or costly embodiment-specific pretraining.
Programmable World Model
Recent video world models generate increasingly realistic and interactive visual experiences, yet lack reliable mechanisms for maintaining persistent world state and enforcing programmable rules over extended interactions. We introduce Programmable World Model, a framework that decouples world-state evolution from visual observation generation. An agent translates natural-language instructions into executable programs that specify entity states and state-transition rules, enabling direct control over individual entities and their interactions. A lightweight engine executes these programs to update and maintain an explicit, persistent global world state, including off-screen entities and non-visual attributes. To connect world state with visual generation, we introduce state-augmented 3D oriented bounding boxes (OBBs) as an intermediate representation. This representation, together with the target camera trajectory, is deterministically compiled into pixel-aligned spatiotemporal conditioning signals for a pretrained video model serving as the generative renderer. This design allows users to create playable games with predefined mechanics, direct control over individual entities, and persistent world state throughout gameplay. We further introduce CombatStateBench, a benchmark for evaluating programmable world models. On CombatStateBench, our method achieves 94% Count Accuracy and 98% State Accuracy, substantially outperforming existing interactive video world models while supporting coherent long-horizon generation. These results demonstrate the effectiveness of separating explicit state evolution from generative rendering for building persistent, programmable worlds.
AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a target prompt without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose \textbf{AgentGrad}, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention modifies the behavior of one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks and reduces wall-clock optimization time by $2.5\times$ on average compared to the next-fastest baseline.
Puppeteer: Object-Grounded Posture-Aware Co-Speech Gesture Generation
Generating co-speech gestures that are temporally coherent, semantically aligned with speech, and grounded with surrounding objects remains challenging. Prior speech-driven gesture models emphasize audio-gesture alignment but do not explicitly account for posture constraints or surrounding objects, failing to capture the inherent correlation between body gestures and the physical space. We present Puppeteer, a posture-aware, object-grounded co-speech gesture diffusion model operating in a causal latent space. We decompose long gestures into structured primitives and learn a causal variational autoencoder that encodes them into temporally ordered latent tokens, each depending only on the past. We then perform conditional diffusion directly in the causal latent space, conditioning on speech signals, motion history, an initial posture reference, and object geometry to synthesize physically consistent gestures. This temporally ordered latent formulation enables explicit temporal control and supports tasks such as gesture in-betweening and gesture completion. To better assess co-speech gesture synthesis beyond existing measures, we introduce new evaluation metrics tailored to this task. We also created SceneGes, the first curated synthetic 3D dataset of embodied co-speech gestures and corresponding 3D objects, enabling object-grounded gesture generation. Experiments show that Puppeteer generates more diverse and temporally synchronized gestures than prior methods, while enabling object-grounded gesture synthesis.
SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?
While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment. Mechanistic interpretability tools are essential to bridge this gap, among which Sparse Autoencoders (SAEs) serve as a cornerstone by isolating interpretable features for model inspection and steering. In this paper, we introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery. Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT to discover the optimal feature, evaluated against curated expert reference features anchored on Neuronpedia across activation rank, concept selectivity on contrastive texts, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capabilities and lead different evaluation dimensions, but remain well behind the expert baseline, approaching expert levels on separating target concepts from contrastive controls while lagging substantially in causal generation steering. Further analysis reveals that although agents can design contrasts to rule out spurious candidates, they frequently misinterpret experimental measurements. These results establish experimental model understanding as a measurable capability for closed-loop autonomous AI R&D. Our code is available at https://github.com/Trae1ounG/SAEScientist.
Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents
AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while withholding the target research history. Every valid method reaching the numerical target supplies a recovery witness and triggers the Core veto. DCP Core requires adequate controls, zero observed recoveries, and a finite-sample bound on recovery in one fresh registered episode. Optional Gate 3 measures the average effect of truthful feedback relative to a specified neutral policy from a shared checkpoint. DCP Evidence adds this effect after independent null calibration and a registered effect margin. Two controlled audits exercise the complete protocol in SQLite optimization and virtual catalyst control under different models. Each produced zero recoveries in 96 episodes, with an upper bound of 0.0468. Each paired study yielded 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies. Additional cases exercise Core, recovered, and audit-incomplete decisions. A deterministic, LLM-free verifier reproduces the decisions from frozen evidence. DCP provides a common evidence language for useful outcomes, alternative routes, and feedback effects across AI research.
DianShi-RxnDB: A Large-Scale, Fine-Grained Organic Reaction Data Platform Built via a Fully Automated Pipeline for Researchers and AI Agents
High-quality structured organic reaction data are essential for developing artificial intelligence for chemistry (AI4Chem), yet much of this knowledge remains dispersed across patent text, images, and reaction schemes. We present DianShi-RxnDB, a large-scale, fine-grained organic reaction data platform built via a fully automated extraction and normalization pipeline integrating patent text, images, and reaction schemes. Its corpus covers organic synthesis patents from the USPTO and EPO published between 1976 and 2025, yielding approximately 24 million reaction instances, of which approximately 14.8 million (61.7%) pass automated qualification checks. Each instance represents a specific single-step experiment recording participants, roles, quantities, temperatures, reaction times, yields, experimental procedures, and provenance links to source patents. In a manual evaluation of 1,300 sampled qualified instances, the micro-averaged field-level accuracy was 92.95%. A matched comparison with Pistachio further indicated advantages in deduplicated record counts, representation granularity, and field-level exact agreement. The platform provides a Web research workbench for searching, filtering, comparing, and source-verifying records, and a Model Context Protocol (MCP) service offering AI agents composable structured retrieval tools. DianShi-RxnDB is available at https://dianshi.opendatalab.org.cn/ .
Revisiting Complete Reasoning Traces for Post-Training
Large language models (LLMs) are often post-trained on pre-collected reasoning trajectories to improve their reasoning capability. Such trajectories tend to be long due to complex, interwoven paths, which often include detours on the path toward the answer. However, it has been underexplored whether LLMs indeed benefit from learning complete trajectories in post-training, such as supervised fine-tuning (SFT). Starting from our pilot study, we find that full trajectories provide only limited benefit, while partial trajectories are effective even under heavy truncation. We analyze redundancy in reasoning trajectories through attention-based analyses and controlled token-removal studies, both of which show that intermediate tokens contribute minimally to final reasoning quality. This suggests that avoiding redundant information may allow LLMs to internally infer coherent alternatives by inferring missing steps from their internal knowledge, given known trajectory endpoints. Furthermore, we show that training LLMs using endpoints leads to consistent changes in reasoning behavior, and that it also benefits post-training methods based on reinforcement learning or on-policy distillation, highlighting the need to revisit complete reasoning traces. Code is available at https://github.com/naver-ai/revisiting-trace.
SchemeArena: Factorized Stress Testing of Scheming in LLM Agents
We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how scheming arises from the interaction of key factors, such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. Prior work examines only a small number of scenarios, limiting the ability to isolate how these conditions shape an agent's propensity or capability to scheme. This limited scale and task diversity also restrict coverage of realistic deployment settings and the range of scheming strategies that can be observed. To this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms. To enable scalable and reliable monitoring, we further propose SCOUT, a scheming monitor that grounds multi-criteria judgments in evidence drawn from agents' reasoning and actions. Across controlled stress tests on five LLM agents, we find that explicit instrumental goals are the strongest driver of scheming propensity. Strategic hints play a distinct role by helping agents translate scheming reasoning into concrete covert behavior. Oversight has mixed effects: in several closed models, action-only monitoring increases scheming, suggesting that partial oversight can act as an optimization constraint rather than a deterrent. CoT is a useful but incomplete monitoring signal: it can reveal latent scheming before execution, yet action-only scheming shows that covert behavior may occur without explicit reasoning evidence. We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.
Multimodal World Model#
SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators
World models are increasingly used as policy-in-the-loop imagination environments, where reliable rollouts require fine-grained controllability with respect to low-level robot actions. A key obstacle to scaling such models in robotics is that actions are not a universal language in pixel space: changes in visual environment, camera view, robot placement, or embodiment alter how the same numerical action manifests visually, leading to conflicting supervision under mixed training and brittle generalization at deployment. We introduce SyncWorld, an action-conditioned world model that serves as a zero-shot simulator across unseen environments without any additional training. SyncWorld leverages a visual calibration episode---paired frames and actions that showcase all the controllable degrees of freedom---to specify the setup-specific Action--Visual Mapping in context. Training with visual calibration contexts teaches the model to interpret actions through visual evidence and to leverage interaction history when explicit calibration is unavailable. Experiments show that SyncWorld can accurately simulate action outcomes in previously unseen settings, and that its capability of simulating rollouts enables test-time policy improvement without training.
$Φ$-Bench: Can Large Language Models Engineer the Infrastructure That Powers Them?
Large language models (LLMs) have demonstrated remarkable capabilities in reasoning and code generation, raising the prospect that they could assist in developing and optimizing the very infrastructure that powers them. However, existing benchmarks mainly focus on isolated kernels, predefined operators, or pre-specified optimization targets, and therefore fail to evaluate the ability of LLMs to perform open-ended, long-horizon LLM infrastructure engineering. To address this gap, we present $Φ$-Bench, a benchmark for systematically evaluating LLMs on engineering the LLM infrastructure stack. Derived from optimization problems studied in frontier research and grounded in real-world code repositories, $Φ$-Bench provides broad coverage of the LLM infrastructure stack and spans tasks of varying complexity, ranging from localized kernel-level function completion to long-horizon implementation and end-to-end system optimization. Extensive experiments on frontier LLMs reveal their current capabilities and limitations in engineering complex LLM infrastructure, offering insights into the challenges that remain on the path toward autonomous optimization of future AI infrastructure.