Multi-modal semantic segmentation benefits from complementary sensors, but this benefit is fragile when some modalities are degraded or missing at test time. Many existing methods rely on dedicated fusion modules or multi-branch cross-modal encoders, and the fusion often becomes unreliable when modalities are absent, leading to substantial performance drops. We propose FuseAdapt, a parameter-efficient framework that performs fusion within the adaptation space of a frozen vision foundation model (VFM), exploring whether VFM priors can serve as an effective shared backbone for multi-modal segmentation. Instead of attaching a standalone fusion network, FuseAdapt assigns each modality a lightweight adaptation branch and performs adaptation-space fusion by keeping Q/V low-rank adaptation private to each modality and routing only FFN low-rank residual activations across modalities inside the frozen backbone. The pretrained VFM remains fixed while modality-specific corrections and explicit cross-modal interaction are learned through the adaptation paths. To handle missing modalities, we further introduce Predictive Modality Modeling (PMM), a direct representation-level objective for cross-modal redundancy. During training, PMM masks a target modality and predicts its embedding from the visible ones, supervised by a teacher target computed under full-modality input. PMM is used only during training and adds no inference-time overhead. Across three multi-modal segmentation benchmarks (MCubeS, DELIVER, MUSES), FuseAdapt consistently improves both modality-complete and modality-incomplete segmentation with a small trainable budget, yielding up to +3.01 mIoU under full modalities and +11.25 mIoU under missing modalities over prior state-of-the-art. Project page: https://devinxzhang.github.io/FuseAdapt-Page/
Memory enables large language models (LLMs) to use information beyond their parameters, such as past user conversations, long documents, or an agent's earlier actions. For a frozen LLM, access to this accumulated knowledge is limited to what a memory system places in its context window. However, gathering and combining relevant information needed for each query while keeping the model's context short remains challenging for existing systems. Here we introduce Memorilla, a memory system that gathers information from an entire document collection into a fixed number of query-dependent 'memory tokens' for a frozen LLM. Memorilla encodes documents (such as user chats, book chapters) once into reusable embeddings, then learns to retrieve and place query-specific information in the LLM's input embedding space. Using only 16 memory tokens, Memorilla achieves the best performance against baselines including Mem0 and MemoRAG on five of eight benchmarks spanning personalization, long-document and knowledge-intensive question answering. Notably, on FactKG, it reaches 84.33% accuracy versus 80.43% with full-context inference, while using 108x fewer tokens to represent the source documents. Furthermore, Memorilla can facilitate continual learning in agentic systems where the document count (the agent's past turns) increases throughout the trajectory. In a TextWorld environment, we show that Memorilla reduces the agent's tendency to wander between rooms it has already explored, and improves its average success rate from 18.6% with full-history in-context learning to 24.2%. Overall, Memorilla aids frozen LLMs in accumulating and using knowledge effectively across long-context interactions and takes a step towards building continually learning AI systems.
One-shot Font Generation (OFG) aims to synthesize a target glyph by transferring the stylistic characteristics of a single reference glyph while preserving the structure of the target character. However, accurately capturing font style from a single reference image remains challenging due to the limited stylistic information available in a single glyph, and this challenge is further amplified for complex scripts. We propose CoDeFont, a script-agnostic OFG framework that addresses this challenge through multi-level style-content factorization and progressive style-content reconciliation. CoDeFont independently aggregates complementary style and structural information across multiple spatial levels to obtain factor-specific representations, while a contrastive objective explicitly encourages separation between style and content embeddings. During reconstruction, the disentangled representations are progressively injected through multi-level style conditioning, enabling stylistic information to be incorporated at different decoding resolutions while preserving the underlying glyph structure. We evaluate CoDeFont on structurally distinct Bengali and Latin scripts across seen and unseen font styles. Extensive quantitative and qualitative analyses demonstrate consistent improvements over existing state-of-the-art methods, highlighting the effectiveness of the proposed framework for OFG across diverse writing systems.
Must the dimension and accuracy costs of zeroth-order optimization multiply? Global upper bounds for smooth stochastic nonconvex optimization scale as $O(d\epsilon^{-4})$, with both factors separately necessary. We show that their product is unavoidable. We establish an $\Omega(d\epsilon^{-4})$ lower bound for smooth stochastic nonconvex optimization, yielding an $\Omega(d\delta^{-1}\epsilon^{-3})$ bound for nonsmooth Goldstein stationarity. Both match the best-known full-space upper bounds, establishing joint minimax optimality in dimension and accuracy. These bounds hold against fully adaptive randomized algorithms with unrestricted queries and outputs, even under arbitrary sample reuse. As a byproduct, we establish noiseless exact-value lower bounds for both smooth and Goldstein stationarity, even against randomized algorithms.
Time-series language models typically align numerical signals with language using labels, captions, or question-answer pairs, without explicitly requiring their representations to preserve how the underlying signals evolve. We introduce forecast grounding, a pretraining principle that requires language-aligned time-series representations to remain predictive of the future evolution before they are adapted to downstream tasks. We instantiate this principle in Lotus, which divides each time-series channel into non-overlapping patches and maps them into the LLM embedding space. It then predicts the next patch from the LLM hidden states of the time-series tokens, encouraging the learned representation to retain temporal information. To capture both temporal changes and periodic patterns, we also propose a temporal-spectral patch representation that combines each raw time-series patch with its spectral representation before mapping it into the LLM embedding space. Across three language model backbones and six downstream evaluation datasets, Lotus achieves the best average performance with minimal trainable parameters, reaching 59.91% accuracy and outperforming the strongest matched baseline by 3.36%. Our representation analyses further show that Lotus uses the numerical input and preserves temporal structure in its learned patch representations. These results establish forecast grounding as a promising pretraining principle, enabling compact time-series language models to outperform models 25-50 times their size on TSRBench. Our code is publicly available at https://github.com/DA2I2-SLM/Lotus
Long-term time series forecasting requires learning temporal dependencies that remain stable and generalizable under temporal distribution shifts. However, real-world time series are often noisy, non-stationary, and heterogeneous, making deep learning models prone to overfitting incidental temporal interactions in the training data. This raises a question: when observations are difficult to align, could temporal relationships be easier to reuse? To investigate this possibility, we analyze patch-level temporal relationship matrices across commonly used time series datasets, revealing dominant patterns shared across different temporal windows and variables. Comparisons with patch-shuffled and IID Gaussian controls further support the presence of reusable temporal relationship structures. Motivated by these findings, we propose \textbf{T}emporal \textbf{Struct}ure Network (\textbf{TStruct}), a simple yet effective network that learns reusable temporal dependency structures and adaptively combines them for each sample and variable. Specifically, TStruct directly parameterizes a set of shared temporal dependency matrices and generates normalized combination weights using dynamic sample information and static variable priors. These weights represent a soft allocation of each sample-variable pair over shared temporal dependency structures, enabling the model to generate adaptive temporal dependency matrices under shared structural constraints, rather than fitting incidental interactions in a fully unconstrained manner. Notably, TStruct consistently outperforms several strong baselines across all evaluated datasets, achieving superior accuracy while maintaining high computational efficiency. Our code is available at https://github.com/yangzeli/TStruct.
AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck. A learned verifier could check proofs cheaply, but training one requires a correctness label on every step, and such labels are hard to obtain for open-ended proofs. Existing methods take them from human annotators, which is costly, or derive them from final answers, which proofs do not have, and formalizing each proof in Lean requires expert effort. We introduce LeanPRM, a framework that uses the Lean proof assistant to create step-level labels for natural-language proofs. A model writes a natural-language proof together with a paired formal proof, without feedback from Lean. Lean then labels every step, and a second model discards failures caused only by the formal syntax, so that the negative labels mark errors in the mathematics. LeanPRM contains 10,634 proofs with 152K labeled steps and no human annotation. A Qwen3-4B verifier trained on LeanPRM raises exact step-label accuracy on a held-out set of 200 proofs (LeanPRM-Bench) from 48.2 to 74.7, as opposed to 60.3 when the same base model is trained on PRM800K. It also selects better proofs in best-of-N selection on IMO-level problems, and outperforms the PRM800K-trained verifier on three of four external verification benchmarks. Training on LeanPRM improves verifiers at every size from 0.6B to 8B parameters on LeanPRM-Bench, Hard2Verify and IMO-GradingBench. Given the labels need no human annotation, LeanPRM can grow with the supply of formal statements, offering a path to step-level feedback on AI-generated proofs that are hard for humans to check.
Large language model (LLM)-based multi-agent systems (MAS) can solve complex problems beyond the reach of a single agent by decomposing tasks and coordinating complementary capabilities. Despite extensive research on MAS, existing work largely focuses on specific aspects or individual phases of their development and operation, resulting in fragmented perspectives with objectives and signals that are difficult to reconcile across these different settings. To bridge these disconnected perspectives, we conceptualize MAS development and operation as a three-phase lifecycle spanning design-time orchestration before execution, runtime intervention during execution, and post-hoc attribution after execution, and introduce MAS-Gate, a unified node-centric quantification applicable across these phases. Modeling each execution as a directed acyclic graph, MAS-Gate combines within-node utility, which estimates a node's contribution through changes in reference-answer likelihood, with inter-node strength, which approximates semantic uptake and potential downstream reach. We adapt this shared quantification to three objectives: providing a structure-aware dense reward for Orchestrator training, training an Editor to act on execution prefixes using paired rewards from completed edited and unedited trajectories, and supplying risk-ranked nodes as auxiliary evidence for LLM-based post-hoc attribution. Experiments spanning interaction environments, knowledge-intensive multi-hop question answering, general reasoning, and post-hoc attribution demonstrate that MAS-Gate improves workflow construction, targeted runtime editing, and LLM-assisted error localization over corresponding phase-specific baselines. Together, these results establish MAS-Gate as a unified structural interface for improving MAS generation, dynamic editing, and error attribution throughout the lifecycle.
Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals under a benign context. We identify a distinct challenge: ensuring the safety of the \emph{multi-step reasoning process} itself, where locally coherent steps may still lead to globally harmful outcomes. To explore this vulnerability, we propose PDRA (Paradox-driven Reasoning-based Attack), a novel attack method that programs a harmful reasoning trajectory. PDRA first constructs a paradoxical narrative core by pairing a malicious actor (e.g., a terrorist) with a benign task (e.g., writing a safety bulletin). It then embeds this core into a structured analytical framework that forces the model through a mandatory three-stage analysis -- surface purpose, latent intent, and operational reconstruction -- guiding it to produce detailed harmful content while maintaining narrative coherence. Mechanistically, we show PDRA steers the model's internal representations away from refusal-related patterns, suppressing safety-triggering lexicons while activating planning-oriented terms. Empirically, PDRA achieves state-of-the-art attack success rates across diverse LLMs with high efficiency. \textcolor{red}{\textbf{Warning:} This paper may contain sensitive content.}
This manuscript presents DeepThink, a knowledge-addressable framework for recurrent language modeling. A Knowledge Bank stores typed Knowledge Units, a Knowledge Activator selects units from the current execution state, Knowledge Assembly produces an executable configuration, and a Shared Thinking Operator updates that state. An output stage then organizes generation. This organization exposes knowledge capacity, activation width, and recurrent depth as separate design variables, supporting both complete parameter packages and fine-grained expert or low-rank modules. We formalize the interfaces and their execution semantics, including state-dependent selection, parameter-space assembly, and thinking followed by generation. For compatible templates and complete runtime-state configurations, we establish an exact Fixed-Schedule Realization of an existing fixed-depth model. Exact shared-base reparameterization and pretrained output-tail reuse provide complementary initialization constructions. Together, these results define a route from pretrained models to dynamically selected and assembled recurrent computation. A synthesis of conditional computation, recurrent models, modular parameters, and knowledge interfaces positions DeepThink within the research landscape and identifies the design questions arising from their integration.
Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model's answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is anchoring: the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator.
Depression is a prevalent mental disorder that affects millions worldwide, with early detection crucial for effective intervention. While exist?ing methods have achieved promising results in automated depression detection, they face two key limitations: (1) noise in post screening: symptom-based semantic screening is easily misled by superficial lexical patterns, thereby introducing false-positive posts; (2) weak op?timization of the post encoder: relying solely on coarse user-level supervision signals pre?vents models from capturing depressive cues within individual posts. To address these limita?tions, we propose PAT, a Personality-driven Augmentation and Transfer learning frame?work. In the first phase, PAT optimizes a post?level encoder via a fine-grained contrastive learning objective, utilizing LLM-based data augmentation driven by classic personality the?ories to alleviate data scarcity. In the second phase, the trained post-level scorer is deployed for high-quality post filtering, while its en?coder parameters are seamlessly transferred for user-level tasks. Extensive experiments demon?strate that PAT significantly outperforms exist?ing methods. Furthermore, PAT delivers multi?granular interpretability by tracing post-level mental-state trajectories and highlighting key symptoms, thereby offering reliable diagnostic evidence for clinical practice.
Clinical speech reveals identity through both spoken identifiers and voice, yet its pauses, delivery, and speaker roles can also carry useful clinical evidence. Protecting that speech therefore requires more than preserving its words. We present AegisVox, a factorized framework that assigns consistent surrogate voices and edits identifier-bearing turns using document-consistent replacements. Across 390 public simulated or non-patient conversations totaling 63.6 hours, a matched comparison of two speech renderers separates replacement realization from retained identifiers, speaker linkage, and clinical preservation. Text-to-speech rendering realizes 535/601 planned replacements versus 453/601 for the codec-based speech editor. However, similar aggregate clinical preservation conceals more model-labeled contradictions, and conversation-wide speaker scores conceal greater edited-speech linkage under a score-direction sensitivity analysis. Fresh transcription of original audio provides a reference for interpreting the clinical measurement chain; an independent judge tests sensitivity to model-based assessment. Better replacement realization can therefore coexist with worse clinical or speaker-linkage measurements. AegisVox supports use-specific evaluation of those tradeoffs, not a guarantee of identifier removal or clinical safety.
Speech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present O\'ido, an inference engine and a family of models that run open-vocabulary English speech recognition entirely on an ESP32-S3 (240 MHz Xtensa LX7, 8 MB of PSRAM, 16 MB of flash, no neural accelerator) faster than real time on one core. Version 1 of this work ran a 13 M-parameter Conformer-CTC at 1.97x the duration of the audio. We measured on the board that the bound was memory traffic and per-frame cost, not arithmetic: external memory reads at 89 MB/s (PSRAM) and 32 MB/s (flash), and a streaming recognizer re-reads every weight for every chunk. Version 2 combines a model built for that memory system (80 ms frame stacking, a 64-channel front end, 12 or 16 blocks, int4 weights that mostly fit in PSRAM, fine-tuned from the version 1 model) with engine changes that overlap the weight stream with compute (an asynchronous data-cache preload), keep work buffers in SRAM and make the decoder cheap. On a physical board the streaming real-time factor falls from 1.77 to 0.54 (variant B) and 0.72 (variant A), the final text appears 259 ms (B) or 431 ms (A) after the end of speech with audio fed at real-time speed, and the host build and the firmware produce identical transcripts. Variant A reaches 7.1%/16.4% word error rate on LibriSpeech test-clean/test-other in streaming mode with 0.64 s chunks and the on-chip language model (4.6%/10.7% with full context). Code, models and the raw board measurements are released.
Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.
Wood identification supports timber inspection and species conservation. However, benchmarks can overestimate generalization when photographs of the same tree occur in training and test sets. We present WoodBR-Cross: 4,875 transverse-surface microscope images, 38 operational taxonomic classes and 303 catalogue-linked trees. A species subset contains 3,406 images, 20 classes and 214 trees. Task A evaluates all classes using duplicate-aware image partitions. B1 partitions images of the species subset, allowing different photographs of one tree across partitions. B2 keeps each tree's photographs together, testing recognition of unseen trees on the same subset. We benchmark three backbones and baselines across three folds and five training seeds. Validation-selected ConvNeXt-Tiny achieves B2 tree-level macro-F1 of 0.8621 +/- 0.0317; frozen DINOv2 features yield 0.9000 +/- 0.0722. Values are means +/- SD across three folds; fine-tuned scores are seed-averaged within folds. This resource supports reproducible, group-aware evaluation.
Bayesian optimisation is the natural tool for shape design when objectives are expensive and non-differentiable, but it needs a compact yet expressive parameterisation of the search space. Hand-crafting one is a complex endeavour requiring domain expertise, and often yields implicit infeasible regions, artificial bounds, and coupled, unordered coordinates. We instead learn the parameterisation from a collection of existing designs, applying principal component analysis to the deformations between shapes. The result is a linear, interpretable search space in which the number of components explicitly trades expressivity against dimensionality. Across aerofoils, wings, and radio-frequency cavities, spanning 2D geometry to 3D aerodynamics and electromagnetics, we show improved sample efficiency and the ability to explore beyond the confines of hand-crafted baselines.
Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.
LLM agents increasingly improve at inference time by accumulating experience in context rather than by updating parameters. This process is often described as in-context reinforcement learning (ICRL). Whether in-context learning (ICL) can actually play the role of RL, however, has not been tested. We study this question in its simplest form, direct ICRL, where the model conditions directly on raw trajectory-reward pairs, and ask whether the reward acts as a learning signal. Through controlled experiments on four benchmarks across six models, we find that the reward is read, but its effect is small: flipping, randomizing, or removing the reward leaves the improvement curve almost unchanged, and this holds even under meta-prompts that explicitly instruct the model to explore, exploit, or reason over rewards. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for agent memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.
Reinforcement learning environments are now a primary lever for improving large language model (LLM) capabilities in post-training, yet most agentic benchmarks remain static: the world moves only when the agent acts, the reward is a terminal verdict, and the pass bar is set arbitrarily. We introduce StoreBench, a live-commerce environment in which an agent runs a mid-size online apparel store on a production-grade commerce backend, testing long-horizon planning and economic judgment under uncertainty. Customers order around the clock, suppliers reprice and fail, and market shocks arrive with partial or no warning. The agent acts through the same 29 merchant tools a human operator would use, under a windowed operation budget that makes simulated time a function of actions taken, so model latency cannot influence simulated time. Pass thresholds are calibrated against scripted anchor policies, the reward is hardened against a catalogue of reward hacks, and every episode replays identically given a sequence of actions. We evaluate seven frontier LLMs on 11 scenarios of 30 to 45 days and a full simulated year, over three world seeds at matched reasoning effort. No model matches the scripted smart-triage policy on average: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. Human experts working through the same tools and budgets outscore every model (mean composite 0.708 vs. 0.700). Over a full simulated year under the Claude Code harness, most models show dramatic performance improvement. In a GRPO post-training run, Qwen3.5-27B trained on only five disjoint tasks raises its mean composite on the held-out evaluation tasks from 0.136 to 0.373. We release five example training-split tasks, ten sample trajectories, and the scoring and verification tooling; the full environment and evaluation suite are withheld to keep the benchmark uncontaminated.