AI Cosmos Team, Jinfeng Zhou, Yuxuan Chen, Fangming Chang, Mingyu Su, Yifan Zhu, Yuan Wang, Bo Zhang 等 35 位作者
Large language models (LLMs) can improve task performance during learning, yet these gains do not necessarily reflect learners' ability to thereafter apply the knowledge and skills independently. We present AI Cosmos, an agentic infrastructure for situated learning that organizes learning around two complementary principles: scenario-based participation and cumulative assets of practice. AI Cosmos organizes learners and role-specialized agents within learning scenarios that integrate task-relevant knowledge and, when needed, executable workbenches. An asset-memory loop preserves the processes and outcomes of practice as reusable assets with explicit provenance and dependencies. We evaluate scenario-based interaction through two preregistered randomized controlled studies examining the acquisition of legal reasoning and programming competence, holding the same underlying LLM across conditions. AI Cosmos achieves greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.
Large language models (LLMs) have become a powerful tool for listwise relevance ranking, which can improve web search by comparing retrieved passages jointly. However, processing every candidate verbatim and generating a permutation autoregressively can make inference costly, every candidate must be serialized verbatim into the prompt and the input therefore grows linearly with both the number and the length of the passages. While sliding-window ranking handles context limits through repeated re-encoding of overlapping blocks, hard compression techniques sacrifice the fine-grained details required for accurate discriminative judgment. To better understand and mitigate this trade-off, we conduct a comprehensive study of soft compression for LLM-based listwise ranking. Our findings reveal that a vanilla gist representation is ill-suited to ranking for two reasons: strictly causal encoding weakens how thoroughly each gist token absorbs its passage, and the representation is never trained to expose the relative-preference structure that separates closely competing candidates. To this end, we propose GistRank, a three-stage training framework that equips a large language model with gist-compressed passage representations by learning faithful gist tokens under a semi-bidirectional attention scheme, instilling discriminative comparison through pairwise training, and adapting the model to the listwise objective. GistRank collapses each passage into only $K{=}4$ gist tokens, this gives 14$\times$ speedups over one-shot full-text ranking while reaching 59.17 average nDCG@10 over TREC DL19/20 and five BEIR subsets within 0.89 of the strongest full-context baseline. Overall, GistRank provides an effective means to enhance efficiency for searches in real-world environments.
Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.
Chloe Su, Jiahan Li, Zhenting Qi, Chaoran Cheng, Haoyu Chen, Sejoon Jun, Weirui Ye, Yilun Du
Agents operating over long horizons need an effective memory that allows them to reuse their experience throughout long interactions. If the entire history of an interaction is retained, it becomes difficult to find the relevant experience, while compressing it into summaries may result in the loss of visual and spatial details that are required later on. We present \textbf{MFM}, a multimodal filesystem memory architecture that keeps complementary types of agent memory. \textbf{Episodic memory} stores the observations, actions, and outcomes, and \textbf{spatial memory} associates experiences with locations. \textbf{Semantic and procedural memory} build up general knowledge and reusable methods of acting, and \textbf{working memory} holds the current plans. MFM makes use of searchable textual records to connect these different representations to the underlying multimodal episodes. The agents are able to find previous experiences by means of contextual cues and to examine the original observations in order to recover details that are missing from their textual descriptions. Through consolidation, the agents convert experience into higher-level knowledge and procedures while still keeping references to the original episodes. In the case of long-horizon Minecraft and computer-use tasks, MFM leads to improved performance with three base models. With GPT-6-Astra, MFM completes \emph{Mine Diamond from Scratch} and reaches 16 out of 24 milestones on the \emph{Ender Dragon} task, as against 12.7 for the best baseline. On CUABench-KiCad it attains a mean reward of 0.440, as compared to 0.400 for the best baseline and 0.267 in the case of no persistent memory.
Hankun Wang, Bohan Li, Shi Lian, Xiaoyu Gu, Jing Peng, Da Zheng, Yiwei Guo, Colin Zhang 等 10 位作者
Speech editing for content creation requires precise control over both what an edit should do and where it should apply. Free-form natural language provides a flexible interface for expressing edit requests, but its ambiguity may leave the intended operation, parameters, or target region underspecified. We study a semi-formal transcript-grounded edit instruction that specifies operation categories and parameters and binds them to transcript spans or boundaries. This semantic timeline avoids explicit timestamp alignment and provides an externally inspectable contract for compositional edits. We instantiate the interface in dots$.$tts$.$edit, a speech editing model built on a continuous autoregressive speech generation architecture. Four representative speech-creation controls cover lexical content, affective expression, pitch and speaking-rate delivery, and temporal phrasing through text, emotion, prosody, and pause editing. Task-specific data pipelines construct operation- and scope-controlled pairs while retaining source-derived context outside each target region. We further introduce doteBench, a bilingual benchmark that measures precise instruction following, local preservation, and audio quality across the four controls and their composition. Experiments show leading overall instruction following and local preservation across its five editing categories, while audio quality remains comparable to existing open-source systems. Across three Seed-TTS-Eval shards, the model shows negligible differences from the base model in zero-shot TTS recognition error rate and speaker similarity.
Personalized recommendation depends on learning semantic and collaborative relations among items from online content and user interactions. Generative recommendation (GR) offers a promising paradigm by autoregressively generating discrete semantic IDs (SIDs) of items. However, randomly initialized SID token embeddings and GR models trained from scratch do not directly inherit the semantic relations inherent in items, while inter-item collaborative relations are learned largely implicitly through token-level attention and next-token supervision. To address this limitation, we propose REGER, a unified framework that explicitly incorporates inter-item relations into semantic pretraining, history encoding, and SID generation. We pretrain the GR model through generative semantic retrieval, enabling it to inherit inter-item semantic similarity by learning to generate semantic neighbors' SIDs. We combine global attention with local attention aligned with SID boundaries to capture historical context and directed transitions between adjacent items. We further develop a theoretical framework for hierarchical preference decomposition over the SID prefix tree, proving an exact factorization of collaborative preference distributions and a weighted decomposition of item-level KL divergence into prefix-conditional KL divergences. Building on this analysis, we distill a discriminative teacher's full conditional distributions over all legal next tokens at ground-truth prefixes along the target SID path, jointly optimizing this objective with cross-entropy. Extensive experiments on real-world datasets demonstrate consistent improvements over existing GR baselines and support the effectiveness of the proposed preference distillation.
Automatic speech recognition systems commonly rely on reference transcriptions for evaluation, while reference-free approaches often depend on internal confidence estimation or auxiliary language models. We propose READ (Reference-free Hypothesis Evaluation with Acoustic Discrepancy), a novel metric that evaluates ASR hypotheses directly from the speech signal. READ emphasizes the acoustic grounding of hypotheses. It uses a pretrained auto-regressive TTS model to compute the conditional likelihood of speech tokens given a text hypothesis, to measure fine-grained acoustic discrepancy between speech and text. Without additional training, READ can be applied for hypothesis refinement. Experiments show that READ correlates with specific recognition errors and improves ASR outputs, achieving up to 20\% relative error rate reduction, with particularly strong gains under noisy conditions.
Keepace. ai Team, Bowen Wu, Heng Wang, Jufeng Mi, Kang Liu, Kun Li, Ling Man, Long He 等 15 位作者
Large language models hold promise for delivering safe, scientifically grounded, and personalized exercise guidance at scale, yet the sports-and-health domain still lacks curated training corpora and rigorous evaluation benchmarks. Keepace.ai is a domain-specialized LLM built on Qwen3.5-27B through post-training, currently optimized for Chinese-language exercise guidance across three tasks: workout plan generation, sports-and-health knowledge QA, and personal sports data analysis. In internal rubric-based evaluations designed with domain experts, Keepace.ai outperforms its base model and larger general-purpose baselines, with the most pronounced improvements in exercise safety, boundary compliance, structured output completeness, and scientific accuracy. This report summarizes the current design, data pipeline, safety-first evaluation framework, and evaluation results of Keepace.ai.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Maojiang Su, Jiale Chen, Qijie Zhu, Shang Wu, Jiayi Wang, Jihai Zhao, Guo Ye, Yibo Wen 等 9 位作者
PairFormer-based all-atom models demonstrate strong capabilities in biomolecular binder design. However, existing methods often use all-atom representations for modality-specific binder design, such as protein and peptide design. In contrast, recent geometric latent diffusion models pursue unified binder design by representing molecular blocks from different modalities in a shared latent space. These complementary strengths motivate us to investigate how to generalize the use of pretrained all-atom representations to unified binder design. We introduce PairFuse, a parameter-efficient method that enables this generalization through representation fusion. Specifically, we use lightweight layerwise adapters to integrate pretrained all-atom representations into an independently pretrained geometric latent diffusion model. Our adapters integrate pretrained all-atom representations through two mechanisms: enriching the diffusion model's molecular block features and incorporating pairwise structural information into the geometric diffusion model's denoiser. Theoretically, we prove that the augmented denoiser preserves geometry and pairwise attention fusion provides additional attention expressivity beyond feature-only fusion. Experimentally, across antibody, peptide, and small molecule binder-design benchmarks, PairFuse achieves stronger performance than the pretrained diffusion model and existing representation-fusion baselines.
Spiking reservoir computing combines the training efficiency of fixed recurrent reservoirs with the energy efficiency of event-driven inference, where computation is triggered by sparse spikes rather than executed densely at every timestep. However, existing spiking reservoirs almost exclusively rely on Leaky Integrate-and-Fire (LIF) neurons, whose low-pass dynamics limit their ability to distinguish information encoded at different temporal frequencies. We introduce HRF-Res, a liquid state machine in which LIF neurons are replaced by a heterogeneous population of Harmonic Resonate-and-Fire (HRF) neurons. These neurons exhibit intrinsic oscillatory dynamics and selectively respond to input frequencies, enabling the reservoir to capture frequency-specific temporal patterns. The result is more discriminative representations and higher classification accuracy than LIF-based reservoirs, achieved without sacrificing the energy efficiency of spiking computation, with up to 19x lower energy consumption than an equivalent non-spiking oscillatory reservoir. Evaluated on ten benchmarks spanning time series, sequential images, audio spike trains, and neuromorphic vision, HRF-Res achieves state-of-the-art accuracy among spiking reservoir methods. These results demonstrate that introducing frequency-selective dynamics into spiking reservoirs improves representational quality and classification performance while preserving their inherent energy efficiency.
Diffusion video models produce striking clips, but they cannot reliably render an exact formula, a correct chart, a rule-consistent game or narration synchronized to on-screen events. A different route is emerging: multimodal agents that write code to render every frame and synthesize every sound. We introduce VibeVideoBench, a benchmark for evaluating agents on code-driven video generation. Its 239 self-contained tasks in English and Chinese span 12 categories, from data stories and music-driven pieces to classroom lessons and paper explainers; each asks an agent, running in its native harness inside a sandbox with no image or video generation model, to deliver a finished MP4 with sound within one hour. Evaluation is hierarchical and evidence-based: program checks of the delivery spec, about 13k task-specific rubric checks traced to the requirements they enforce, and five quality dimensions anchored per task, scored by three VLM judge panels. Across nine frontier agent systems and 2,151 scored rollouts, GPT-6 Astra and Claude Opus 5.5 rank first and second with both the judges and human annotators, and the strongest system satisfies every core requirement on 76.3% of tasks. Delivery is a second axis: counting missing videos as zero changes the ranking, and three quarters of the lost videos are final renders that started too late, ran too slowly or never started, a failure of how agents plan their hour. Agents build their own rendering pipelines and inspect their own frames, yet how much they look does not predict quality. The judges' ranking of the systems closely matches blind human preferences.
Zhiheng Fu, Yupeng Hu, Zixu Li, Zhiwei Chen, Qinlei Huang, Shaokun Wang, Han Liu, Fan Liu 等 10 位作者
The SIGIR 2026 eCommerce Challenge Query Reformulation Task evaluates the reformulation of task-oriented queries into expressions suitable for product retrieval, followed by BM25-based retrieval. These queries often imply product needs through activity-level semantics, creating a lexical gap with catalog descriptions. Systems must preserve task intent while adapting the number of reformulations to the Automatic and Interactive protocols. We propose TaskBridge, a catalog-anchored framework comprising Catalog Semantic Anchoring, Unified Five-Candidate Rewriting, and Protocol-Aware Selection. It grounds queries in catalog vocabulary, generates five complementary reformulations, and selects one query for Automatic retrieval or up to four for Interactive retrieval using BM25 scores, query-product reranking, and weighted Jaccard overlap. The selected queries are independently executed using BM25 to produce results in the required six-column format. Experiments demonstrate improved retrieval effectiveness, with multiple complementary queries providing additional gains in the Interactive setting. TaskBridge achieved first place in the SIGIR 2026 eCommerce Challenge.
Yu Zhao, Zekun Zhang, Fan Jiang, Bo Zeng, Xintong Wang, Chonghan Liu, Shimin Shan, Yichao Du 等 9 位作者
Long chain-of-thought (CoT) reasoning models such as DeepSeek-R1 keep lengthening their inference chains under test-time scaling, yet standard self-attention incurs quadratic cost in context length, making long-CoT deployment costly in production. We therefore introduce LISA (Linear-Indexed Sparse Attention), a plug-and-play replacement for self-attention that requires no full pretraining. LISA combines a linear-attention branch for long-range memory with a Lightning Indexer that dynamically selects the most relevant tokens for a sparse self-attention branch. The two branches are fused through a learned gate, so the model retains broad context while focusing computation on informative tokens. We train LISA in two stages. In Stage 1, a linear-attention branch is integrated with sliding-window attention and distilled from a frozen teacher's full-attention distribution. In Stage 2, the Lightning Indexer replaces the static window and is trained with a per-head KL-divergence objective to mimic teacher attention patterns. The resulting model reduces eager-mode self-attention computation from quadratic in sequence length to linear in the number of generated tokens and the selected context budget. Experiments on DeepSeek-distilled Qwen models show that LISA cuts measured eager-mode mean sample latency by 23.9% at 16K context without degrading long-context reasoning quality. Across nine mathematical and code reasoning benchmarks, the gated LISA-1.5B model improves the equal-weight benchmark average by 4.2 percentage points relative to the full-attention baseline.
Long-horizon agents performing multi-step Web tasks must reuse earlier page observations and interactions to support subsequent actions. However, preserving relevant history does not ensure that it exposes the relations needed for the current decision. We call this gap representation mismatch. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and renderer transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.
Modern image generators can produce complex visual artifacts, including diagrams, interfaces, and records that combine text, graphics, and layout to convey detailed information. They may refuse requests for overtly harmful imagery yet still generate artifacts that guide dangerous actions or support deception. We call the risk arising from what such images enable, rather than only what they depict, functional visual harm (FVH). Existing safety benchmarks focus largely on overtly harmful content, such as pornography and graphic violence, leaving image generators' safety alignment against FVH insufficiently evaluated. To address this gap, we introduce FVH-Bench, comprising 669 prompts across five harmful visual functions and ten risk domains. The 11 evaluated image generators rarely refuse these requests and frequently generate artifacts that fulfill the targeted harmful functions. High scores for target harm realization and harmful artifact usability further indicate that these outputs can support the intended harmful purposes. A further analysis finds that existing safety detectors also provide limited coverage of FVH, indicating gaps in their ability to recognize harmful visual functions. Together, these findings show that evaluating image-generator safety alignment requires considering not only what images depict, but also what they may enable.
Autonomous agents increasingly depend on harnesses that govern interaction, memory, tools, and execution around large language models (LLMs). While recent work has begun to evolve these harnesses from execution experience, the supervision used to guide evolution often remains relatively static, making it difficult to track changing behavioral weaknesses across rounds and potentially limiting generalization to unseen tasks. To address this issue, we propose RuHarness, a rubric--harness co-evolution framework for continual agent self-improvement. RuHarness continually discovers and reassesses behavioral criteria from execution trajectories, prioritizes the most relevant unresolved weaknesses, and uses them to guide targeted harness updates while keeping the task objective fixed. Extensive experiments across diverse agent tasks and backbone models show that RuHarness outperforms strong harness-evolution baselines, particularly on unseen held-out tasks. Further analyses show that adaptive rubric prioritization tracks shifting weaknesses and supports more sustained improvement and stable generalization across evolution rounds.
Learning controllable world dynamics from heterogeneous interaction videos is challenging because robot embodiments use incompatible action spaces, while human videos lack native control labels. We present RynnWorld-Latent, a cross-embodiment video world model that uses the 3D-grounded latent actions introduced by RynnLAM as a shared conditioning interface. This representation enables joint pretraining on human interaction videos, robot demonstrations, and simulation rollouts without requiring a common native action space. Built on a pretrained video transformer, RynnWorld-Latent combines global action tokens with an identity-initialized local pathway that aligns actions with future video latents. We pretrain on 23 datasets comprising approximately 553.5M frames and adapt the model to target robots through post-training. The framework supports two data-generation routes: digital teleoperation uses target robot commands to generate action-video demonstrations, while latent-action synthesis transfers source motion to an edited initial frame and pairs the resulting rollout with retargeted pseudo robot actions. Experiments show improved video prediction metrics on public benchmarks and two real-robot platforms, alongside qualitative action following when different action sequences share the same initial observation. Synthetic mid-training improves policy success rates on LIBERO and LIBERO-Plus, and both generation routes improve real-world policy success rates over real-demonstration-only training in the evaluated settings. These results support universal latent actions as a scalable interface for learning controllable visual dynamics across heterogeneous embodiments.
Suozhi Huang, Kelsey Fu, Binxu Li, Luna Lyu, Kaize Wu, Wenbo Lu, David Smerkous, Jiahan Li 等 15 位作者
Autonomous scientific agents must not only reason about experiments but also interact with the physical environments in which they are executed. Existing laboratory simulators typically rely on manually constructed assets and predefined procedures, while laboratory videos capture real equipment, layouts, and interactions but leave this information implicit in pixels. We introduce LabWorld, an agent-driven Real2Sim framework for reconstructing interactive 4D laboratory worlds from monocular RGB video. LabWorld recovers reusable 3D assets, a metric scene, time-varying object states, and procedure events. Multi-frame observations provide complementary evidence for asset and scene reconstruction, while motion reconstruction combines visual and geometric observations with event constraints encoding the spatial and temporal requirements of laboratory interactions. We evaluate LabWorld on real pipetting and vortexing videos collected in our laboratory against general-purpose agentic and dedicated reconstruction baselines. For pipetting, LabWorld achieves a mean Mask IoU of 0.501, exceeding the strongest baseline by 0.033. For vortexing, it achieves a mean Tube-Length-Normalized PCK-AUC of 0.638, exceeding the strongest baseline by 0.035. On five pipetting videos, event-conditioned reconstruction increases the fraction of modeled physical constraints satisfied from 14.3% (5/35) to 100% (35/35). The resulting worlds expose geometry, motion, articulation, and events as explicit, editable state for replay, inspection, and future embodied scientific systems.
Zeyu Wang, Yiheng Duan, Tao Chen, Donghyun Koo, Sihan Liu, Wenbo Zhang, Chuan He, Liang Liu
Low-precision training often relies on stochastic rounding (SR) to keep quantized gradients unbiased and prevent small updates from rounding to zero. Standard SR draws an independent random threshold for every gradient element. However, accelerators generate random numbers on vector units whose throughput lags far behind their matrix units. Drawing one threshold per element saturates the vector unit and can make INT8 training with SR slower than BF16 training. Sharing random thresholds across tensor elements removes this cost but forces those elements to round in the same direction and shifts their rounding errors by a common offset. We find that this offset degrades training primarily through the activation gradient, which propagates the error backward into earlier layers. Because earlier layers cannot cancel an incoming offset using their own local rounding errors, the offset must be removed at the layer where it originates. We propose Structured Subspace Stochastic Rounding (S3R), which quantizes each upstream gradient once for both the activation and weight gradient matrix multiplications and shares a single threshold per token across the hidden dimension. Its error correction, S3R-EC, subtracts the structured mean rounding error from the activation and weight gradients while leaving the optimizer unchanged. In LLM pre-training experiments, uncorrected threshold sharing degrades convergence, whereas S3R-EC eliminates this degradation and matches the validation loss of independent SR. Furthermore, S3R-EC accelerates full INT8 training steps of dense LLMs on TPU v5p by 1.24x to 1.32x over BF16 and by 1.11x over independent SR.
Guo Ye, Zexi Zhang, Xu Zhao, Shang Wu, Haoran Lu, Shihan Lu, Han Liu
Vision-Language-Action (VLA) models have shown remarkable generalization by mapping web-scale knowledge to robotic control, yet they remain blind to physical contact. Consequently, they struggle with contact-rich manipulation tasks that require reasoning about force, texture, and slip. While some approaches incorporate low-dimensional tactile signals, they fail to capture the high-resolution dynamics essential for such interactions. To address this limitation, we introduce DreamTacVLA, a framework that grounds VLA models in contact physics by learning to feel the future. Our model adopts a hierarchical perception scheme in which high-resolution tactile images serve as micro-vision inputs coupled with wrist-camera local vision and third-person macro vision. To reconcile these multi-scale sensory streams, we first train a unified policy with a Hierarchical Spatial Alignment (HSA) loss that aligns tactile tokens with their spatial counterparts in the wrist and third-person views. To further deepen the model's understanding of fine-grained contact dynamics, we finetune the system with a tactile world model that predicts future tactile signals. To mitigate tactile data scarcity and the wear-prone nature of tactile sensors, we construct a hybrid large-scale dataset sourced from both high-fidelity digital twin and real-world experiments. By anticipating upcoming tactile states, DreamTacVLA acquires a rich model of contact physics and conditions its actions on both real observations and imagined consequences. Across contact-rich manipulation tasks, it outperforms state-of-the-art VLA baselines, achieving up to 95% success, highlighting the importance of understanding physical contact for robust, touch-aware robotic agents.
A small number of GPU kernels often determines the efficiency of large language model training and inference. High-performance kernel development typically relies on expert knowledge through manual implementation, collaboration with general-purpose models, or specialized agents equipped with domain knowledge, evaluation methods, and search workflows. We investigate how much value this knowledge provides as foundation models become stronger. We introduce JustGoal, a goal-only baseline in which a general-purpose agent receives only the optimization objective and constraints, without kernel-specific infrastructure, knowledge bases, or predefined workflows, and builds its own evaluation loop. On NVIDIA B200 GPUs, JustGoal outperforms open-source libraries on 8 of 10 operators, achieving per-operator geometric mean speedups of 1.14x to 2.34x. It also surpasses the fastest publicly released specialized-agent kernel on 5 of 9 operators with available comparisons. In a controlled same-model comparison using Fable 5.1 on single shapes, the KDA-0.5 harness reaches 0.95x of JustGoal's performance, while single-mechanism ablations show no consistent benefit from domain knowledge or predefined workflows. These results indicate that specialized machinery should be justified by clear gains over a same-model, same-scope goal-only baseline and reevaluated whenever the foundation model changes.
Yueshan Li, Lorenzo Fiaschi, Andrea Censi, Emilio Frazzoli
Independently developed autonomous agents, such as autonomous vehicles, must guarantee joint compatibility, such as safety and task-completion requirements, when they operate in a shared environment. Existing methods require a centrally designed specification, or interaction mechanism before any agent is deployed. However, such centralized design is unrealistic when agent designs are confidential to their developers, and it is not flexible enough to accommodate future agent designs. We study a bi-level structure in which a regulator coordinates developers by publishing seed agents, and each developer independently tests its agent to ensure that it is compatible with every seed-compatible agent, a property we call future-proofness. Under this structure, joint compatibility is guaranteed for any agents that are designed and tested independently, without any coordination with each other. In this work, we first formulate compatibility for general multi-agent systems and show why it is not trivially implied by seed-compatibility. We then characterize the future-proof agents as the closure of the seed and show that this closure is monotone in the seed. Finally, we study a class of executable tests that involve interaction with the seed alone, and we ask to what extent such tests can distinguish the future-proof agents.
Structure predictors like ESMFold fold proteins using neural network computation and have transformed structural biology. However, the driving mechanism from amino-acid (AA) sequence to 3D structure is insufficiently understood, and biological meaning remains locked in the dense network. Sparse autoencoders (SAEs) have shown promising results in obtaining interpretable concept features over an overcomplete basis of intermediate protein sequence-model representations. Yet folding models, which predict physical shape from AA sequence semantics, have no such account. We hypothesize that the features carrying the folding computation evolve in the 3D structural regression process. In this paper, we interpret the folding trunk in three steps. First, a pilot analysis on the representation distribution across modules and recycles locates distinct spatial computation in the trunk. Second, the dictionaries expose a division of labor: the single representation carries inherited semantics, and the pair representation computes the pairwise geometry. Third, we interpret the dictionary at scale with a reasoning language model, trace how the concepts compose across blocks by gradient-based attribution, and follow the interpreted code to the model's folding output, verifying where the computation aligns with, or departs from, known biochemistry. Together, the sparse dictionaries we train capture the relational geometry the trunk computes beyond the sequence model, and their concept activations forecast the model's own fold quality, turning the black-box folding trunk into a sparse, interpretable, and scientifically verifiable circuit.
Franziska Herbert, Vignesh Prasad, Han Liu, Dorothea Koert, Georgia Chalvatzaki
Learning structured task representations from human demonstrations is essential for bimanual manipulation, where action ordering, object involvement, and interaction geometry vary significantly across executions. A key challenge lies in jointly capturing the discrete semantic task structure and the temporal evolution of object-centric geometric relations in a form that supports reasoning over task progression. We introduce a semantic-geometric graph-based task representation that jointly encodes object identities, inter-object semantic relations, and per-object motion histories, via a Message Passing Neural Network (MPNN) encoder and a Transformer-based decoder. The encoder operates solely on the temporal scene graph, producing structured representations decoupled from action labels. The decoder then conditions on action-context to forecast future actions, associated objects, and object motions. This decoupling learns task-agnostic representations, enabling encoder reuse across embodiments through decoder-only finetuning on a small robot dataset. Across eleven bimanual tasks from two datasets, we find that the benefit of structured semantic--geometric representations over simpler sequence-based models grows with task variability in action ordering and object involvement. At deployment, a planner couples the action and motion predictions with learned Probabilistic Movement Primitives, achieving full task success on two real-robot bimanual tasks and outperforming graph ablations, Transformer, decoder-only, and finetuned vision-language model baselines.
Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Feichen Song, Boyi Zeng, Yiqin Hao, Shixiang Song, He Li, Yizhou Yang, Zitong Wang, Yifan Liu 等 10 位作者
Looped Transformers increase computation per token by repeatedly applying a shared backbone, offering a way to improve language modeling and downstream task performance at a fixed backbone parameter count. However, each loop depends on the preceding loop, so increasing recurrent depth also increases decoding latency. Existing approaches include Parallel Loop Transformers, which reduce latency by overlapping loop computation across token positions, but yield rapidly diminishing marginal accuracy gains at larger depths. To address this challenge, we introduce LoopFold, which folds computation at multiple recurrent depths into parallel batches for the same token prefix. A lightweight predictor estimates inputs to later loops so they can start before earlier loops finish. A corrector then refines the outputs of these later loops using newly available results from earlier loops. All components are trained jointly using only the final next-token prediction loss. In experiments with 1.4B models, deeper LoopFold configurations achieve progressively higher average zero-shot accuracy across eight tasks, rising from 53.41% to 56.38%. Vanilla looped models and PLT show limited gains over the same depth range. Compared with a nine-loop vanilla model, LoopFold uses five sequential stages and additional parallel computation to reduce decoding latency by 46%.
Feichen Song, Boyi Zeng, He Li, Shixiang Song, Yiqin Hao, Zitong Wang, Yizhou Yang, Yifan Liu 等 10 位作者
Looped Transformers offer a parameter-efficient approach to language modeling by repeatedly applying a shared backbone. Vanilla looped Transformers use each backbone output directly as the input to the next iteration. These outputs may contain information useful several iterations later, but retaining it across the loop depends entirely on the single hidden state. We introduce StateLoop, which maintains one additional persistent state per token to retain information across iterations. At each iteration, a learned gate controls how much of the previous state is retained and how much information from the backbone output is incorporated. The updated state is projected back and added to the backbone output for subsequent iterations and final prediction. This design reduces Pile perplexity and improves average accuracy across eight zero-shot tasks over Vanilla Looped. Its perplexity advantage over Vanilla Looped grows with trained loop depth, consistent with persistent state becoming increasingly useful as recurrent computation deepens.
Mental health research on social media has been dominated by English-language datasets and DSM-5-centered resources, limiting cross-lingual and cross-cultural understanding of online mental health expression. We introduce MATCHA, the first Chinese multi-disorder mental health dataset that links disorder-specific Weibo Super-Topic communities with users' broader public timelines. MATCHA covers seven DSM-5 disorders and a control group, comprising 46,025 cohort-level users and 1.8 million posts. The dataset distinguishes users with explicit self-disclosure evidence from broader disorder-related participants through an LLM-assisted annotation pipeline, and is released in a rigorously de-identified, controlled-access format. Using MATCHA, we analyze behavioral, linguistic, symptom-related, and life-event-related patterns across disorders, and conduct cross-lingual comparisons with Twitter-STMHD. Our analyses show that Weibo Super-Topics function as concentrated spaces for affective and health-related expression, while cross-platform differences between Weibo and Twitter often exceed cross-disorder variation. MATCHA provides a new resource for community-aware and cross-lingual mental health NLP research.
Microkernels provide good features such as high security, safety, and flexibility, but their reliance on hardware-based isolation leads to heavy performance overhead, limiting their widespread adoption. Although constructing single-address-space microkernels based on safe language-based isolation is a promising approach to addressing the performance issue, this approach still faces challenges in maintaining the principle of minimality and enforcing capability-based, fine-grained access control over system resources. We propose safeL4, an approach to constructing microkernels based on safe languages. It guarantees minimality through low-level abstractions for unsafe operations, and implements a capability system based on the safe language's object capabilities to achieve fine-grained access control over all operations on system resources. We implement a prototype system in Rust, and our evaluation results show that the primitives provided by safeL4 are, on average, 1.9$\times$ faster than those of seL4.