Long-horizon language agents accumulate tool results and replies until the history no longer fits the context window. Current agents shrink this history by having a model summarize it, so each compaction rereads and rewrites the history it replaces, and cache-aware triggers decide when to compact while taking the size of the result as given. We present Folio, a context manager that records, folds, and recovers history through one set of handles. While the agent runs, Folio records three linked levels: a short turn note the model writes as part of its ordinary reply (L1), an event manifest of every call written by the runtime (L2), and the captured source records (L3). Because these views already exist, the runtime folds the context programmatically, selecting coarser views for older turns. The fold prints the handles that index L2 and L3, so recovery tools that read bounded pages, compute over full text in files, and keep findings in a persistent dossier reach the evidence the agent needs with little added input. The fold trigger prices an explicit landing point that includes the retained budget. On 75 LOCA-bench tasks with DeepSeek-V4-Pro behind every system, Folio solves 66 tasks against 54 for the same runtime with model-written compaction and 49-59 for Claude Code, Codex, and SoL-Pi, in about half their working time, at 2.4 times lower cost than that runtime; each fold takes a median of 0.6 seconds.
https://arxiv.org/abs/2611.16627
Most sign languages have no annotated corpus, only small dictionaries with one or a few example videos per sign. This study examines how well a pose encoder trained on a single sign language serves such languages without any target-language training. An encoder trained with metric learning on American Sign Language (ASL-Citizen) is evaluated frozen on six datasets of Turkish, Russian, Indian, Argentinian, Pakistani and Sri Lankan sign languages, with one or four clips per sign enrolled and, where signer identities are available, the query's signer excluded from the gallery. With one enrolled clip, the encoder retrieves the correct sign first in 36.4-80.7 % of queries, against 16.0-69.6 % for the best learning-free method (angle-based dynamic time warping) and a six-target mean of 39.3 % for an ASL-trained OpenHands graph network. In a crossed ablation of the encoder (embedding geometry, aggregation of body and hand embeddings, training objective, and attention between articulators), two choices matter: a per-pair InfoNCE objective outperforms supervised contrastive and cross-entropy training, and blocking attention between body, left hand and right hand inside the encoder raises accuracy by up to 5.1 points. The best configuration reaches 68.0 % one-shot accuracy (59.0 % for the reference encoder). On CISLR, an Indian Sign Language dictionary of 4,765 words with one video per word, the encoders find the correct word first for 17.3-19.0 % of queries, above the 16.8 % of the original RGB baseline. Over eight seeds with the joint encoder, spherical, Euclidean and hyperbolic embedding spaces differ by at most 1.6 points, except the hyperbolic token-attention head trained with InfoNCE, which is 1.4-2.0 points worse, so the geometry is a minor factor.
https://arxiv.org/abs/2611.16621
Decoding upper-limb movements from scalp electroencephalography (EEG) is an important problem for brain-computer interfaces (BCIs), but many deep learning models provide limited insight into the brain regions underlying their predictions. We propose a two-branch architecture that combines an LGGNet-inspired local-global graph branch with an EEGNet-style convolutional branch for movement-versus-rest classification. The graph branch groups 61 EEG electrodes into anatomically informed regions and learns their spatial relationships, providing a direct representation of the regions emphasized by the model. We evaluate the proposed architecture using leave-one-subject-out cross-validation on the Ofner upper-limb movement dataset. The model achieves a mean F1-score of 0.74 +/- 0.06 across subjects. Beyond classification performance, we analyze the learned graph structure to identify the regions with the strongest connectivity. Across all held-out subjects, the highest connectivity is consistently observed in the Parietal Left Low region, corresponding to electrodes P1 and P3. This region is opposite to the performed right-arm movements and lies within the posterior parietal area previously associated with upper-limb movement processing. These results show that incorporating a region-based graph into the decoder can provide an anatomically interpretable representation while maintaining effective movement classification, offering a practical approach toward more transparent EEG-based BCIs.
https://arxiv.org/abs/2611.16609
Individualized dosing requires predicting how a patient's drug exposure would change under a different dose or schedule. However, a model that accurately predicts exposure at the current dose may not reliably predict what happens when treatment changes. We developed two neural models that learn pharmacokinetic dynamics from data without predefined pharmacokinetic equations. The first predicts exposure after a dose change, while the second also allows changes in dosing schedules. In simulations, the second model predicted exposure under new dosing schedules more accurately than competing neural models. On real data of kidney transplant recipients treated with tacrolimus, both models showed lower prediction errors than Bayesian estimation when predicting exposure at a later visit. These findings suggest that incorporating pharmacokinetic principles into neural models can improve predictions needed for individualized dosing.
https://arxiv.org/abs/2602.03215
Frontier AI agents increasingly succeed on realistic, long-horizon tasks and are beginning to improve the systems through which they operate. Yet existing benchmarks often restrict agents to narrow action spaces, use simplified environments, or rely on verification methods that are unreliable for evaluating arbitrary agent-generated systems. We introduce Singularity-Bench, an open-ended benchmark for autonomous AI development with six end-to-end tasks spanning pre-training, mid-training, post-training, inference, the agent harness, and cluster orchestration. Each task requires agents to implement core systems from a minimal starting point and evaluates them against mature human-engineered baselines on identical hardware. Additionally, through extensive red-teaming covering 9 of the 10 categories in the OWASP Top 10 for Agentic Applications, we expose reward-hacking vulnerabilities in the conventional Harbor framework and develop Breakwater, an extension of Harbor that blocks all attacks tested in our study. Across 11 agents built from 7 frontier models and 2 harnesses, Opus 5.5 with Claude Code matches or outperforms the human baseline on all six tasks with a Singularity-Index of 1.25, suggesting it may have the skills needed to reach the singularity. The remaining agents average only 0.58. Our vision is for Singularity-Bench to track how much of AI development agents can already automate and when they may transition from assisting to taking end-to-end responsibility for building stronger successors.
https://arxiv.org/abs/2611.16591
Long-term spacecraft autonomy requires embodied agents to use operational history under communication delays, time-critical operations, environmental uncertainty, complex physical interactions, power and computational constraints, and safety and reliability requirements. We introduce AstroMemBench and a Scenario-Challenge-Capability design method that maps these requirements to executable memory tests. The benchmark covers 16 memory subchallenges through 140 templates and 2240 instances across nine tasks in planetary surface operations, on-orbit servicing, and deep-space exploration. A deterministic environment scores actions for task completion, state changes, safety, and resource use. Ablations under nine conditions show that the benchmark is sensitive to how accurately agents recover and use historical information and to differences in memory-system design. Main experiments compare six configurations. Mem0, Chain-of-Memory, LightMem, and retrieval-augmented generation (RAG) improve same-session mean scores over native Pi Agent, with no clear advantage of the three memory systems over RAG. After a session restart, Mem0 outperforms RAG by 12.25 points. Benefits favor instruction retention and result reuse over ongoing state tracking and energy reservation. AstroMemBench supports the design, development, and selection of memory systems for embodied spacecraft agents. AstroMemBench 1.0 and its evaluation code are publicly available: this https URL.
https://arxiv.org/abs/2611.16590
This paper investigates whether Large Language Models can answer a personality questionnaire as a specific individual would. Given a user's social-media posts, the model completes the BFI-2-60 Big Five questionnaire as that user. The predicted responses are converted into trait scores and compared with the user's actual scores. We evaluate GPT-4, Gemma, Llama 3.1, and Mistral on the MyPersonality dataset. To determine whether predictions reflect information from the user's writing, we compare them with answers produced without user text and with values derived from general population tendencies. We measure prediction error across all users and separately for individuals with low, intermediate, and high trait scores. We also test whether an LLM can express an explicit personality profile when given only the target Big Five scores, without user text, and examine how predictions change as more text is provided. The results show that user text does not consistently improve performance. Models perform better near the middle of the trait distribution, while at the lower and upper ends they often remain close to central values or perform well at one end but poorly at the other. When given an explicit profile, GPT-4 most clearly translates trait information into consistent questionnaire responses, particularly for Extraversion and Neuroticism. Increasing context does not systematically reduce error: open-source models fluctuate more, whereas GPT-4 remains comparatively stable. Overall, the evaluated models do not consistently reproduce the questionnaire behavior of specific individuals.
https://arxiv.org/abs/2611.16586
We present Jais 3, a fleet of six open, Arabic-centric agentic language models ranging from 0.9B to 375B parameters. Jais 3 is built to complete tasks for Arabic speakers, not only to answer questions in plain text: it is trained to plan multi-step tasks, call tools and Model Context Protocol (MCP) servers, operate a terminal, search the web, and read, edit, and produce PDF, spreadsheet, and presentation files, while communicating with the user in Arabic. Instead of pretraining from scratch, we adapt each model of the fully open K2 Horizon fleet to Arabic through continual pretraining, followed by supervised fine-tuning (SFT), final token preference optimization (FTPO), and reinforcement learning with verifiable rewards (RLVR). We combine natural Arabic text with translated, transliterated, and synthetic Arabic, and replay the base model's own training data to keep its reasoning, agentic, and English abilities. Jais 3 covers Modern Standard Arabic and the major regional dialects, follows instructions with Arabic-specific constraints, and inherits the 512K-token context window of K2 Horizon (all models except the 0.9B). To our knowledge, its 375B model is the largest open Arabic-centric language model to date. For evaluation, we revisit how Arabic models are assessed, moving to fully generative evaluation on audited Arabic benchmarks. Across model sizes, Jais 3 matches or outperforms open-weight models of comparable size on Arabic and agentic benchmarks, with particularly strong agentic capabilities and Arabic instruction following, and far fewer hallucinations. We release the weights and the checkpoints from every training stage of all six models, together with samples of the training data, the data recipes and training configurations, the training logs, the audited benchmarks, and the evaluation code.
https://arxiv.org/abs/2611.10224
LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), an independent operational benchmark that evaluates 33 systems from 17 developers: specialized guardrails (e.g., Llama Guard 4, ShieldGemma-2B, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok 4.1 Fast), jointly on detection rate, false-positive rate, latency, and cost. We cover input/output moderation across 11 harm categories, jailbreak detection across 13 attack techniques, and prompt-injection detection, using in-house datasets built from handcrafted prompts, curated external samples, and quality-controlled synthetic generation, with synthetic moderation samples paraphrased to suppress generator fingerprints. Content-moderation runs are repeated five times and all rates reported with 95% intervals. The Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally ahead: top systems match the best generalist LLMs on detection (93-96%, no significant difference) at low false-positive rates while running ~5x faster; cost depends more on serving setup than on class. On jailbreak detection, generalist LLMs achieve higher median detection on three of four datasets and lower false-positive rates, but at higher latency and, for closed frontier models, 10-100x higher cost. On prompt injection, the two best dedicated detectors lead. We release the framework, an interactive leaderboard, and four datasets as a vendor-neutral basis for selecting safeguards under deployment constraints.
https://arxiv.org/abs/2606.20668
Photonic neural network accelerators have emerged as a promising solution for reducing the power consumption of deep neural network (DNN) inference by executing matrix multiplication on analog photonic hardware. However, their inference accuracy is limited by low-bit quantization required for photonic processing and by analog photonic noise. This paper proposes an approach for improving inference accuracy by exploiting the high similarity among activation vectors in DNNs. Specifically, our approach constructs a representative vector for each group of similar activation vectors, computes its inner product digitally with high precision, and offloads the residual inner products to the photonic accelerator. Since the activation vectors within each group are highly similar, the residual vectors have much smaller norms, thereby reducing the effects of quantization errors and analog photonic noise. Furthermore, our approach preserves the photonic inner-product computation pipeline, allowing seamless integration into existing and future photonic accelerators without hardware modification. Experimental results on six representative DNN models demonstrate that the proposed approach can improve inference accuracy across diverse network architectures and application domains.
https://arxiv.org/abs/2611.16561
We present a suite of 50 unpublished research-level theoretical physics problems with verifiable answers. Without AI assistance, a typical problem in our collection could be solved by a world expert in about a day or less but would take considerably longer for experts in adjacent specialties. We test the ability of off-the-shelf AI agents to tackle these problems autonomously, without human hints. When given access to tools available in a realistic setting, including the ability to search the literature and write scripts for numerical and symbolic work, the best-performing agent, GPT-6-Astra harnessed in Codex, obtains the final answers with a near-perfect accuracy of 98% at an average cost under USD 3 per problem. On average, agents show only small changes in accuracy when denied web search and access to background research papers, suggesting significant memorization of the literature. These results suggest that state-of-the-art agents can accurately perform concrete calculations required across theoretical physics research. We comment on the broader implications of these findings and discuss possible explanations for the current lack of agent-led breakthroughs in theoretical physics despite their outstanding problem-solving abilities.
https://arxiv.org/abs/2610.16733
As large language models increasingly generate mathematical proofs, whether they develop a proving style distinct from that of human mathematicians remains an open question. This question is particularly pressing as mathematicians increasingly voice concerns that proofs generated by large language models are difficult to interpret. Here we ask three large language models to independently reprove all 465 propositions of Euclid's \textit{Elements} and compare their citation choices against the human-proof baseline through Jaccard similarity and complex network analysis. We find that AI-generated proofs differ from human proofs, and this difference is stable and reproducible: mean pairwise Jaccard similarity among AI runs (~0.68) is nearly twice that between AI and human proofs (~0.37), and cross-run classification consistency is also high. By examining citation structure in detail, we identify the concrete signature of this difference: approximately 40% of AI citations point to foundational elements (definitions, common notions, and postulates), compared with only 13\% in human proofs; even among theorem-to-theorem citations, AI tends to cite more distant results. Roughly half of AI's extra citations trace to the upstream dependencies of the human proof. Together, these findings suggest that the AI proving style is a tendency to construct arguments on its own from foundational elements rather than directly citing later, more established results, which we term "building from the ground up". Our findings offer a new perspective on how AI systems organise and invoke mathematical knowledge, with implications for understanding the interpretability of AI-generated proofs.
https://arxiv.org/abs/2611.16527
LLMs have quickly evolved from models used in isolation to agentic systems that delegate tasks to external tools and other LLMs to solve complex assignments. This agentic augmentation has led to impressive capability gains. However, when and how delegation improves performance remains poorly understood. For example, it is unclear whether delegation helps when tasks involve post-cutoff information that the orchestrating LLM cannot reliably resolve from its own knowledge. In this work, we use purpose-built datasets with pre- and post-cutoff splits to compare various delegation modes. We find that delegation provides no consistent accuracy improvement over non-delegated setups, with effects varying across models and tasks. When tasks involve knowledge unavailable to the orchestrator, delegation can substantially degrade accuracy, while direct tool use with the same LLM performs much better. Our analysis identifies distorted delegated queries, inefficient tool use, and failures to preserve and use information between agents as recurring failure modes. These findings suggest that effective orchestration requires both well-calibrated delegation decisions and effective exchange of information between agents.
https://arxiv.org/abs/2611.16526
ZONOS3-Mini is an ultra-compact, ultra-fast, zero-shot voice-cloning TTS model. The model can run in real time on a laptop or smartphone. For its latency and size, it achieves state-of-the-art accuracy and cloning fidelity. To achieve this we propose a latent MoE backbone with only 886M parameters (100M active per token) and a new audio codec, Zyphra Audio Codec 1 (ZAC-1), which represents 44.1 kHz audio at a 15 Hz frame rate. On the new ZTTS2-Eval benchmark, ZONOS3-Mini reaches a 6.71% macro-average WER over 55 language sets (35 clean and 20 in the wild), which is competitive with larger and slower state-of-the-art models. We release the model weights, inference, and training code.
https://arxiv.org/abs/2611.16525
Victim search in wilderness areas is time-critical. Drones enable rapid area search, but scheduling them is challenging because victims may continue moving, detection results are uncertain, and effective decisions must combine structured mission information with unstructured contextual clues and domain knowledge. To address this, we propose an agent-aided drone scheduling system. The system employs an agent at the ground control station (GCS) to coordinate multiple drones over successive search cycles. In each cycle, the agent applies a scheduling policy that combines a large language model (LLM) for reasoning about where to search with an optimizer for determining how to dispatch the drones under energy constraints. Simulation results show that the proposed system consistently completes the victim-search mission, whereas none of the six non-LLM baselines does so and direct LLM scheduling is considerably less reliable. The source code is publicly available at this https URL.
https://arxiv.org/abs/2611.16524
Predicting every observation in a high-rate stream requires a model to finish before the next arrival. A large model may offer better predictions but miss that deadline. We propose RESYNC: the large model runs asynchronously, while a fast path updates its latest exact but stale latent state using the observations received since that state was computed and applies its unchanged prediction head. The updater learns from the frozen model's representations without task labels. Across one real and two synthetic streams and five architecture families, RESYNC beats the best compact or distilled alternative with a paired 95% interval above zero in all twelve dataset--architecture pairs with a material teacher advantage at anchor ages of 4--12 observations, and in eleven pairs at older tested ages up to 32. The share of the teacher's gain preserved remains complete on the real stream but is lower and generally declines with anchor age on the synthetic streams. In live replay with 1 ms arrivals, the synchronous large models cannot deliver any predictions on time while RESYNC serves at least 99.4% of arrivals within the interval, with 0.35--0.62 ms median latency on one CPU thread: around the same speed as or faster than the compact model, and 7.9--17.7$\times$ faster than synchronous large-model inference. RESYNC targets settings with background compute available but a strict per-observation response deadline.
https://arxiv.org/abs/2611.16523
Photonic neural accelerators offer ultra-low latency linear algebra, yet their usable numerical precision, sustainable throughput, and application accuracy remain tightly bounded by surrounding electro-optical interfaces and physical non-idealities. Pre-fabrication design-space exploration (DSE) requires a cross-layer electronic design automation (EDA) framework capable of evaluating disparate optical topologies under identical neural workloads and physical assumptions. We present \textbf{PISim}, a unified Photonic Inference Simulator that couples pretrained deep neural network execution with physics-based models of six accelerator architecture families, six electro-optic modulator technologies, four optical source organizations, and direct or coherent readout schemes. PISim embeds stage-resolved impairment operators---covering control and analog-to-digital quantization, finite extinction ratio, shot noise, correlated relative-intensity noise (RIN), thermal drift, propagation loss, and channel crosstalk---alongside an analytical cycle-level interface timing model to expose electronic-optical bottlenecks. Using simulated MNIST multilayer perceptron (MLP) benchmarks, we demonstrate how PISim identifies optimal precision operating points under multi-impairment stress (peaking at 6 bits for Mach--Zehnder and microring modulators), quantifies a 52.5 percentage-point accuracy gap between VCSEL-array and parallel-fanout laser architectures arising, under the modeled conditions, from power partitioning and correlated source noise, and models interface-limited throughput scaling across emitter-array widths from 32 to 128 channels. This comprehensive framework establishes an extensible, reproducible foundation for accuracy-constrained Pareto optimization prior to photonic tape-out.
https://arxiv.org/abs/2611.16522
While quantization compresses LLMs and accelerates inference, outlier activations impede efficient low-bit representation. In particular, massive outliers cause substantial performance degradation. Although rotation-based methods that rely on randomized Hadamard transforms for online inference aim to mitigate this, they can produce fragmented distributions with empty bins when a massive outlier dominates a token's activation. This fragmentation severely undermines uniform quantization performance. Moreover, massive outliers concentrate in the down-projection input, where the preceding nonlinearity requires the rotation to be applied at inference time. This constraint prevents training-based approaches such as SpinQuant from learning the online rotation at this position, as a learned dense matrix cannot use the Fast Walsh-Hadamard Transform and would require a full matrix multiplication during inference. To address this, we propose SignRot (Sign-adjusted Rotation), which identifies massive outlier channels and constructs a sign adjustment vector that unifies the signs of the corresponding Hadamard row, collapsing the bimodal distribution into a unimodal profile amenable to quantization. Since the aligned outlier shifts the activation mean away from zero, we pair SignRot with asymmetric activation quantization, whose zero-point absorbs this shift. As a lightweight, drop-in improvement compatible with any Hadamard-based quantization method, SignRot requires only a single calibration forward pass and adds only element-wise sign flips at inference, without an additional matrix multiplication. Applying SignRot to QuaRot achieves a WikiText2 perplexity of 5.88 and an average zero-shot reasoning accuracy of 72.73, surpassing SpinQuant on the LLaMA3-70B model. Code is available at this https URL.
https://arxiv.org/abs/2611.16502
A typed-decision model such as Laya reads a text and a list of options that the caller defines, for example ticket categories, and returns a probability for each option. Code around it compares the top probability with a threshold to answer, abstain or escalate to a person. Earlier work reports, for a Turkish model, that temperature scaling can raise held-out calibration error (ECE). This tutorial gives a five-step recipe (measure, diagnose, fit, validate, deploy) in 130 lines of numpy with standard methods, none new. On 1,065 synthetic, LLM-labelled items the released Laya checkpoint has a pooled ECE of 0.073 but a bucket-macro ECE of 0.143 (0.127-0.184), the mean over buckets (questions grouped by type and option count): the 12-option bucket is over-confident by +0.22, the three-to-five-option bucket under-confident by 0.05, and they cancel. One global temperature makes pooled ECE worse (0.173); a cross-fitted temperature per bucket brings macro-ECE to 0.073 (0.078 over 10 fold assignments) and reaches 0.089 with 10 labelled items per bucket, but holding out one of seven evaluation sets gives 0.099. After the refit, a gate aimed at 10% error answers 46.3% of items (10.5% wrong) against 6.9% on the published confidences, but one cut-off per bucket on the published confidences answers 45.4%, so most of the gain comes from separating buckets. Confidence-only maps (Platt, histogram, isotonic) need 20 or more labelled items per bucket to approach the temperature; with 320 per bucket the lowest macro-ECE belongs to Platt. In these files the retrained checkpoints store no per-bucket table and the stored 12-option value 0.1006 is clamped to 0.5 at runtime; the refit helps 31 of 33 checkpoints on random folds and 21 with a set held out. Fit and test share one generator and LLM labels, so these are findings about these checkpoints and this data.
https://arxiv.org/abs/2611.16499
A decision encoder such as Laya reads a question with named options (for example, which department should handle a ticket) and a text, and picks an option. It reads both in one forward pass and returns a probability for every option, without writing any text. We ask how many LLM-written training rows such an encoder needs and how wrong those labels may be. We ran 19 fine-tunes of Laya (ModernBERT-large) on nested subsets of a 2,818-row pool and scored them on 1,065 held-out items from the same generator, so every test item is in-distribution and no out-of-family test was run. Accuracy rises from 69.4% before fine-tuning to 74.3% with 5% of the rows, 84.8% with 50% and 86.9% with all rows. The curve flattens, and a saturating fit with bootstrap intervals cannot pin down its plateau. Seed spread at 20% of the rows (s.d. 1.6 points) exceeds some of the steps between sizes. Giving small subsets as many updates as the full pool gains 1.9 points at 5% and 1.9 points at 20% of the rows (one seed; p=0.14 and 0.13), and at 20% that is below the 2.3-point difference between two ordinary seeds. A single run with 10% symmetric label noise shows no detectable change (-0.2 points, 95% interval -1.7 to +1.4); 20% costs 3.4 points and 40% costs 8.7 points. Confusable-pair noise at 20% costs 7.4 points, and under noise the model moves from over-confident to under-confident. An independent labeller disagrees with the gold labels on 3.5% of held-out items, and about half of its choice disagreements (16 of 33, leaving each item out of the partner map) name the partner option that the confusable-pair noise uses. Laya is +0.9 points ahead of a TF-IDF classifier trained on the same rows at full size (the interval includes zero) and +21.1 points ahead at 5% of the rows. Everything is synthetic, LLM-labelled and from one backbone, one trainer and mostly one seed.
https://arxiv.org/abs/2611.16498