This paper examines how freelance work is changing in the AI era. We compare the activities people undertake in ChatGPT with the work clients request on Upwork, an online freelance platform, and examine how clients combine tasks into projects. We find that activities more prevalent in ChatGPT tend to lose marketplace share on Upwork, while those less prevalent tend to gain share. Within AI-related work, demand is becoming more diverse, and job postings combine tasks that were rarely requested together before the generative AI era. AI-related postings also have substantially larger project budgets and involve more hours of work, while differences in hourly rates are small. Together, these findings suggest that AI is reshaping not only which tasks clients outsource, but also how they combine tasks into larger, higher-value projects.
https://arxiv.org/abs/2611.14708
Vision-language models such as CLIP can classify images into categories they were never explicitly trained to recognize, but accuracy degrades when test images differ in style or source from their training data. A common remedy generates randomly cropped and flipped copies of each test image, called views, and combines their predictions. Since some views miss the object, methods must identify which views to trust. The dominant approach uses prediction entropy to measure confidence and retains the most certain views. Other methods use signals such as embedding typicality. Despite repeated refinements of these scoring functions, the information they rely on, the model's own response to each view, remains largely unexamined. We investigate this choice while holding other pipeline stages fixed. First, we show that view selection, rather than the gradient update, accounts for almost all accuracy gains. We then examine whether entropy identifies useful views when predictions are wrong, despite at least one view predicting the correct class. In these cases, entropy ranks correct views above incorrect ones no better than chance, recovering only about one-twentieth of the accuracy recovered by an oracle over the same views. This finding holds across two image encoders and under a control ruling out an obvious statistical artifact. An exact analysis explains this limitation: entropy measures confidence but not which class a view favors. We identify an unused signal already available in the augmentation pipeline: the fraction of the original image retained by each crop. Ranking views by this quantity improves accuracy on 14 of 15 benchmark datasets over entropy and on 11 of 15 over the strongest scoring function tested. A pre-method measurement also predicts which datasets benefit.
https://arxiv.org/abs/2611.14686
Classical Sanskrit poses three challenges unresolved by single-call monolithic models: morphophonemic fusion Sandhi, dense nominal compounding SamÄsa, and discourse-level zero-anaphora forced by poetic meter. We present MITRA-Swarm, a multi-agent architecture injecting explicit linguistic structure into an otherwise monolithic pipeline: a Sandhi-splitting agent, a parallel array of generalist gemma-4-26B-A4B-it and domain-specialized gemma-2-mitra-it translators, and a discourse-aware Ellipsis Restorer that consults corpus-derived register tags to avoid inventing active-voice subjects in impersonal legal and philosophical registers. Two separated judges evaluate quality: a routing judge ($J_{\text{route}}$, same model family as the generators) driving an internal critique loop, and an independent reporting judge ($J_{\text{report}}=$gpt-4o-mini, no shared vendor lineage) scoring every output exactly once. On a stratified benchmark of 1,000 verses across five registers (23,990 scored verdicts, 12-condition pre-registered ablation), the full swarm reaches $85.90$ MQM against a monolithic baseline of $85.27$ ($\Delta=+0.63$, $p_{\text{adj}}=0.0018$, 95% CI $[+0.30,+0.97]$), and isolated discourse restoration alone accounts for $+0.44$ MQM ($p_{\text{adj}}=0.0288$), outperforming a call-count-matched self-refinement baseline by $+0.51$ MQM ($p_{\text{adj}}=0.0288$). Severe per-verse failures (MQM$<80$) fall from $9.0\%$ to $7.6\%$, at roughly $9\times$ the per-verse inference cost of the baseline. Additionally, $J_{\text{route}}$ scores its own family's outputs $2$-$7$ MQM higher than $J_{\text{report}}$ on every one of the 12 conditions, empirically motivating the two-judge split; and a preflight screen of a 192,151-entry retrieval index against all 12 evaluation works verifies 0 reference leaks across 269 runtime checks. These are small gains on a metric saturating near 85-86.
https://arxiv.org/abs/2611.14675
Operational reports carry two clocks: when an event happened and when its report arrived; later reports can revise earlier ones. A foundation-model component that maintains the state supported by the reports visible at a cutoff, then rolls a supplied scenario forward, can be helped by exact code before it, after it or inside the call, or asked to reason harder. We ask which remedy removes which error, for which models, on an exactly solvable lost-sales inventory environment: 36,672 independently rescored requests over twelve configurations of closed and open-weight models, with a forced-answer-call interface as a control and protocols frozen before the calls. Weak models fail in separable ways. Canonical input removes report-selection error, exact continuation removes rollout error, neither repairs the other, and which error dominates differs by model; strong models are at ceiling on raw reports. Asking for more reasoning is no substitute and can depend on the interface: under a forced answer call high effort collapses GPT-5.4 mini and DeepSeek V4 Pro, which gain or hold under automatic choice; on its provider the forced call disables reasoning for DeepSeek V4 Flash; and at high effort gpt-oss exhausts its output budget. The contribution is a controlled component diagnostic, released with every request, and evidence that for some model and provider pairs the tool-calling interface confounds reasoning evaluation.
https://arxiv.org/abs/2611.14671
AI sycophancy, the tendency of large language models (LLMs) to flatter users and affirm their beliefs, can undermine accuracy and contribute to misinformation, dependence, and overconfidence. Although prior work has benchmarked sycophancy in released models and proposed mitigations, its origins during training remain poorly understood. Preference alignment methods such as RLHF and DPO are often blamed, but their effects are rarely compared with earlier training stages. We investigate sycophancy across the training pipelines of three model families and three benchmarks. We find that sycophancy is largely established by the end of supervised fine-tuning (SFT), rather than introduced by preference alignment. Across all nine model benchmark combinations, SFT checkpoints attain at least 92% of preference aligned models' sycophancy scores and at least 85% of final models' scores; on average, they are about as sycophantic as final models. Preference alignment increases sycophancy in only four of nine cases, by an average of 4.2 percentage points when it does. To probe the underlying causes, we isolate model representations associated with sycophancy using mechanistic interpretability and measure their activation on training datasets from each stage. SFT datasets show roughly the same average activation as other training data but contribute substantially more examples, and the highest-scoring individual datasets all come from SFT. Our findings suggest that addressing sycophancy requires interventions throughout the training pipeline, particularly in data selection and augmentation for pre-training and SFT.
https://arxiv.org/abs/2611.10045
Thompson sampling is natural randomized strategy for optimizing unknown functions in a black-box manner. In continuous global optimization, Thompson sampling has largely been restricted to Gaussian process priors, as these are essentially the only class for which computing the posterior distribution has been tractable in practice. In this work, we introduce \emph{prior-fitted diffusion transformers}, which are diffusion models that perform Thompson sampling in-context at inference time, after being pre-trained using the prior-fitted network paradigm. To apply these models, we develop a set of practical non-Gaussian priors for general black-box functions---including those that have characteristics that cannot be achieved under Gaussianity, such as priors over unimodal functions. We also quantify the tradeoff between model size and the number of pre-training iterations, and show that these models follow scaling laws akin to those observed in transformers defined over other modalities. Across a set of global optimization benchmarks, we show that prior-fitted diffusion transformers with non-Gaussian priors can achieve improved performance compared to traditional pipelines built on Gaussian processes.
https://arxiv.org/abs/2611.14656
Cross-sensor remote sensing image retrieval aims to find scenes with the same content across heterogeneous sensors, such as SAR and optical or panchromatic (PAN) and multispectral (MS) imagery, without manual labels. Existing self-supervised methods either reconstruct sensor-specific pixel detail, align each image only with its co-registered twin, or rely on large transformers with 118-173M parameters (and up to 398M). We introduce RS-JEPA, a compact joint-embedding predictive framework built on a Vision Mamba encoder shared by both sensors. Since a state-space encoder reads an image as an ordered sequence of patches, we propose scan-path views, which treat two reading orders of the same image as two views with identical pixels and predict the hidden patches in one order from the visible patches in the this http URL keep this prediction challenging, we propose Adversarial View Coupling (AVC), which learns the hardest split of the visible patches, and a neighbourhood consensus loss that aligns the scene similarities of both sensors. Extensive experiments on SAR-optical (BEN-14K, CBRSIR-VS) and PAN-MS (DSRSID) benchmarks demonstrate that RS-JEPA achieves the highest F1@5 among twelve re-implemented methods (72.92, 80.35 and 94.43) with a retrieval network of only 3.88M parameters (14.57M in training, 8-16times fewer than prior JEPA retrievers). Like prior methods, RS-JEPA trains on co-registered pairs. Our results highlight reading order as a source of self-supervised views. code: this https URL.
https://arxiv.org/abs/2611.14496
Memory-augmented vision-language-action (VLA) policies are meant to remember what happened. Many of them, however, fill memory by time: they sample frames uniformly across the episode and stamp each frame with its time step. Because time passes even when nothing happens, such a memory may not tell waiting from working. We show that it does not. On the counting tasks of RoboMME, we make the robot briefly pause or fidget its wrist. Such a null event changes nothing the task depends on, yet a released memory VLA counts the interval as work done and stops early. We call this failure phantom progress. It cuts success from 47 to 16 of 50 on one task and from 44 to 10 on the other. When the same frames are withheld from memory, success returns, so the cause lies in memory rather than in the interruption itself. We trace the failure to a temporal shortcut: because demonstrations proceed at a steady pace, elapsed time predicts progress. Indeed, rescaling the time stamps alone makes the policy stop early when the clock runs fast and repeat when it runs slow. Surprisingly, removing the time encoding does not cure the failure, because the frames of a null event still occupy slots of the time-sampled memory. Across RoboMME's released designs, vulnerability follows how memory is indexed. Of the fixes the diagnosis suggests, training with injected pauses does not remove the failure and impairs counting, whereas indexing memory by events recovers half or more of the loss without retraining. A memory meant to remember what happened should count events rather than clock ticks.
https://arxiv.org/abs/2611.14488
Visual Question Answering (VQA) imposes substantial computational overhead on vision-language models (VLMs), motivating model-agnostic input reduction methods that can be readily applied across diverse VLMs, including in black-box settings. Simply enforcing a visual-token cap reduces computation, but it can discard information needed to answer the question. A single reduction strategy also struggles to accommodate different scenarios. Our key observation is that cropping and compression are complementary: cropping highlights small targets and local details, whereas whole-image compression better preserves context for large targets or distributed evidence. Based on this observation, we propose C$^2$-Routing, which adaptively selects cropping or compression according to the input image and question. Since both strategies may still discard key information, we further design a safety gate that uses the predicted joint correctness distribution under original and reduced inputs to guide fallback and protect accuracy. To accommodate consumer GPU resource constraints, all training and evaluation experiments run on RTX 3090 GPUs. To measure gains beyond token capping, we use full-image baselines with the same cap of 1,024 visual tokens. We conduct extensive evaluation on three VLMs and 18 datasets spanning local details, global scenes, scene text, and documents. With the visual budget already capped, C$^2$-Routing further reduces end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with a mean relative task-score decrease of only 0.77%.
https://arxiv.org/abs/2611.14469
Deepfake detectors are commonly evaluated on real/fake classification alone, but real-world applications require explanations of the visual evidence behind each prediction. Joint vision-language models struggle with authenticity assessment, producing "real" narratives for manipulated images. To address this, we propose a decoupled predict-then-explain pipeline as our submission to the ACM Multimedia 2026 Explainable Deepfake Detection Challenge, with methods trained and evaluated on the XPlainVerse benchmark. A vision-foundation-model ensemble first classifies the image, then a Qwen3-VL-8B model generates a decision-conditioned explanation of the visual evidence. For fake images, a second Qwen3-VL-8B compresses that text into a short non-expert summary without using the image. We further apply metric-aligned Group Relative Policy Optimization (GRPO) to optimize both explainers against the challenge criteria. Under matched conditions, GRPO raises complex BERTScore-F1 from 0.686 to 0.712 and also improves grounding (entity and evidence F1); our submission ranked first on complex BERTScore-F1 and seventh overall.
https://arxiv.org/abs/2611.14436
We describe our solution for the 2026 BEHAVIOR Challenge: 100 long-horizon household tasks in the OmniGibson simulator. We placed 9th on the final leaderboard, the only independent team in the top 10. The core of the solution is a single vision-language-action (VLA) model, one checkpoint for all 100 tasks, built on our 1st-place 2025 solution, which is based on $\pi_{0.5}$. On the 50 tasks shared with 2025, its Q rises from 0.260 to 0.371; across all 100 tasks it scores 0.392. Our main conclusion from 2025 was that the model was undertrained, so most of the 2026 work went into training the model for longer, optimizing some training parameters, making each training step slightly faster and reducing the gradient noise: a migration from 50 to 100 tasks (semantic stages built from the demonstrations' skill annotations, task and stage embeddings of new tasks warm-started from similar old ones, and a 21-dimensional action space matching the teleoperation controls); variance-reduced multi-sample flow matching with a simple noise-vs-compute model that picks the number of action-expert samples per observation without ablations; and a 200k-step training run, about one pass over the 211M frames of the demonstrations. A cloud evaluation infrastructure keeps a shared record with provenance, episode diagnostics and restorable simulator states, and runs six simulators on one GPU together faster than real time. On top of the learned policy, 14 tasks are routed to a GPT-6 Astra coding agent that controls the robot through observation and motion tools, starting every episode from a frozen copy of its own practice session on training instances. Across these 14 tasks, Astra's mean Q is 0.639, against 0.108 for the learned policy. The public submission included only part of the Astra results and scored 0.446 Q with 27.3% full-task success.
https://arxiv.org/abs/2611.14432
Users often complain that AI assistants such as ChatGPT are getting worse. Some versions do change, but human evaluation leaves a second possibility untested: the people judging may change. We separate the two on Chatbot Arena, a public platform where people compare answers from two anonymous large language models (LLMs) and pick the better one or say both are bad. We study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights. For the same pair of unchanged models, the share of votes saying "both are bad" rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%. The ranking does not trend: in 2024, the stronger model's share of decisive votes is flat. The rise holds within the same voter and on the same pair, and across arrival cohorts from their first recorded vote. It holds with prompt controls, in every major language, for dated vendor snapshots, open-weight models and models with stable answer length, and between interface changes. The drift is shared across models: month effects estimated from other models' pairs remove it from each model's own pairs. Platform-wide, the average model would be expected to score 78% against the 2023 average, yet the share of "both are bad" votes ended near where it began. The pattern fits a standard that follows the category's progress, so a product that stands still is judged as declining. Absolute verdicts on AI drifted while relative comparisons held steady, an evaluation drift that matters for LLM leaderboards and for the human feedback used to train models.
https://arxiv.org/abs/2611.14333
Transformer sentence encoders provide strong retrieval quality, but their inference cost becomes a bottleneck when embeddings must be produced on CPUs or refreshed frequently. We introduce StaRSE, a compact Russian static sentence encoder that removes Transformer inference at deployment time: token directions are stored as sign bits, token norms are stored separately, and sentences are encoded by mean pooling over norm-scaled sign directions. StaRSE is initialized from ruBERT-base, fine-tuned on Russian adjacent-sentence and task-shaped pairs, and post-trained with a contrastive stage that directly optimizes the sign-projected token table. On RuMTEB, StaRSE reaches a task mean of $51.16$, $7.98$ points above the strongest included static baseline. The packaged model occupies $11.3$ MiB and has a mean CPU latency of $0.08$ ms per sentence, compared with an estimated $55.7$ MiB 16-bit parameter footprint and $1.43$ ms mean CPU latency for the closest-quality contextual encoder. Ablations examine adjacent-sentence supervision, token norms, and direct sign-projected post-training.
https://arxiv.org/abs/2611.14236
Variability in syntactic structure and parsing ambiguity are fundamental aspects of natural language which modern LLMs have to learn to deal with proficiently. Previous "post-mortem" studies have shown that fully trained Transformers indeed approximate canonical parsing algorithms such as inside-outside, but little is established theoretically about how this is actually learned. Here, we make a first step towards addressing this by studying the learning dynamics in synthetic context-free grammars with variable parse tree topology. The stagewise learning curves are fully interpretable, in the sense that all of the loss stages a) can be computed analytically, and b) can be clearly associated to the representation of a hidden latent tree variable, which allows to lift ambiguity and utilize information from wider contexts. By probing representations, we visualize the emergence of inside-outside computation across depth and during training: while during early stages local (bigram/trigram) information is used and involves only a few initial layers, gradually the Transformer develops an encoder-decoder structure where the most abstract latent information is encoded in the middle, and the latter half of the network is used to progressively refine the final token prediction. This organization of computation is reminiscent of findings in real-world LLMs, and importantly applies here to a synthetic model which allows full analytical control of the data statistics, thus opening the door to theoretical progress.
https://arxiv.org/abs/2611.14221
During extended wakefulness, mammalian neocortical networks experience localized periods of transient neuronal silence termed off-states. Here, we examine how analogous spatiotemporally concentrated disruptions affect sequential reasoning in autoregressive Large Language Models (LLMs). Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models (Qwen3-8B and LLaMA-3.1-8B-Instruct), we contrast clustered bursts of attention head masking against budget-matched diffuse i.i.d. dropout. Concentrated bursts induce substantial reasoning degradation: at onset parameter $\mathcal{B}=0.10$ (scheduling $\sim 5.7\%$ of vulnerable head-token volume), Qwen3-8B accuracy drops from $91.2\%$ under diffuse noise to $78.8\%$ under burst dropout ($\Delta = -12.4$ percentage points [pp], exact McNemar $p = 5.5 \times 10^{-6}$), while LLaMA-3.1-8B-Instruct drops from $84.0\%$ to $76.8\%$ ($\Delta = -7.2$ pp, $p = 0.0198$). Conversely, dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation (retaining $91.2\%$ in Qwen and $84.0\%$ in LLaMA). In an exploratory factorial dissociation ($N=100$), single-step dropout density is associated with an acute drop in a single schedule draw ($-10.0$ pp on Qwen, $p_{\text{holm}} = 0.0425$), while multi-token continuity exhibits an inconclusive compounding trend ($-7.0$ pp, $p_{\text{holm}} = 0.1892$). Error analysis reveals that Qwen failures reflect derivation errors within formatted responses (conditional accuracy drop of $-11.4$ pp), whereas in LLaMA $22.2\%$ of discordant losses lack a final answer format line (conditional gap narrows to $-5.0$ pp). These findings demonstrate that autoregressive reasoning chains are particularly vulnerable to concentrated structural disruptions, providing empirical benchmarks for fault tolerance in transformer architectures.
https://arxiv.org/abs/2611.14209
Deploying vision transformer models across connected and automated vehicles (CAVs) requires joint management of computation and Vehicle-to-Vehicle (V2V) communication. Static split-inference methods use fixed partition points and cannot adapt to time-varying link bandwidth, processing load, and concurrent multi-stream traffic. This paper proposes an edge orchestration framework that jointly selects model partition points, per-layer quantization bitwidths, and execution placement across heterogeneous vehicular compute nodes. The joint partitioning and mixed-precision allocation problem is formulated as a Mixed-Integer Linear Program (MILP) that minimizes compute and communication energy subject to end-to-end latency and task-quality constraints. The framework is evaluated on Jetson Orin Nano (8~GB), Jetson Orin NX (16~GB), and Jetson AGX Orin (32~GB) nodes using OPV2V-driven multi-stream workloads with $K \in \{2,4,6,8\}$. Across the evaluated configurations, the proposed formulation reduces V2V communication energy by $33.3$--$36.0\%$ relative to a Uniform INT8 split. At $K=8$, end-to-end task completion latency is reduced by $34.0\%$ on the Orin Nano and $33.1\%$ on the Orin NX relative to the same baseline, while limiting the reduction in 3D detection accuracy to $3.1$--$4.1$ percentage points in $\mathrm{mAP}_{50}$ relative to FP16 execution.
https://arxiv.org/abs/2611.14196
Foundation models such as GPT and Claude have shown strong capabilities in spatial understanding and embodied intelligence. These advances make agentic robotics a promising path toward artificial general intelligence in robotics. However, we find that using foundation models to control robots poses serious safety risks, and comprehensive safety evaluation is still lacking. We introduce RoboRisk, a comprehensive benchmark for evaluating safety in agentic robotics. Its 92 simulated tasks and 13 physical-robot tasks fall into five categories: Task-level Refusal (T), Plan Repair (P), Execution-time Hazard (E), Injection (I), and Sequential Memory (S). Prior benchmarks cover at most two of these categories, and several evaluate only text outputs. In contrast, RoboRisk covers all five and executes every task in closed loop. We evaluate six state-of-the-art foundation models on RoboRisk: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, and Grok 4.7. These models generally show high harm rates. We also evaluate two robot-specific APIs: the Dream API and PokeBot's RoboAPI. We further propose the Lease-Enforced Agentic Safety Harness (LEASH), which offers essential safety infrastructure for agentic robotics and can be integrated into nearly any agentic architecture. LEASH leverages a decision model (Jev or OpenAI's Decisions API) to implement task, action, and commit gates, which review the task, each action, and each irreversible operation in their visual context. The planner and all three gates share a common safety policy. Experiments on RoboRisk in simulation and on physical robots show that LEASH effectively reduces harm across all six foundation models. It largely preserves their task completion performance and adds little time and token overhead.
https://arxiv.org/abs/2611.14115
We present our system for the SHROOM-visions 2026 shared task on character-level hallucination span detection in image-conditioned vision-language model (VLM) responses, evaluated across English, French, Italian, and Chinese. The detector is built on a frozen Gemma-4-12B-it backbone with trainable LoRA adapters (approximately 10.7M parameters). Because the backbone is a causal decoder, hidden states read from a single copy of an answer cannot attend to subsequent tokens; we therefore repeat the candidate answer within one prompt and extract features from its second occurrence, giving every position approximate bidirectional context in a single non-autoregressive pass. A character-level locator combines intermediate hidden states to predict hallucination probabilities, BIO boundaries, and a response-level abstention score, while separate language-specific classifiers assign one of five hallucination types to each predicted span. On the closed test set, our system reaches a macro-averaged character IoU of 0.4404, more than doubling the official baseline (0.1854), and ranks 5th-6th on IoU across the four languages. On the English tuning split, controlled ablations attribute a gain to the repeated answer that is comparable to that of LoRA adaptation (its held-out gain is smaller and concentrated on hallucinated answers), and they show that the detector relies on the image-answer pairing: shuffling images across examples drives IoU down to the level of the predict-nothing floor or below it in every language. In deployments where the same Gemma-4-12B-it model also generates answers, its base weights can be reused, making the approach practical for local question-answering systems that cannot run a separate verifier.
https://arxiv.org/abs/2611.14100
We study how to generate diverse, physically executable robot demonstrations for dexterous vision-language-action (VLA) learning. Existing demonstration-generation methods typically rely on costly teleoperated examples. Meanwhile, human hand--object datasets provide a reusable source, but existing transfer methods mainly reproduce given interactions. Turning these interactions into VLA training data further requires coverage of different objects and manipulation locations, together with executable arm and finger actions, and visual observations. To achieve this, we present DexSeed, a framework for low-cost generation of dexterous robot data from human hand--object interactions. Our key idea is to expand human sequences while preserving their grasps and realizing the resulting motions under arm-tracking dynamics in physics simulation. Specifically, we organize source sequences by grasp mode with shared language descriptions. InteractionMesh-based retargeting and arm trajectory augmentation adapt these sequences to the robot hand, different object geometries, and pick-and-place locations while preserving local hand--object relationships. An RL finger-control policy produces executable rollouts, while the arm tracks precomputed trajectories. Rendered observations and recorded actions from successful rollouts are paired with grasp descriptions to fine-tune a VLA. Experiments show improved contact fidelity in retargeting and more reliable whole-arm execution than the evaluated baselines. Simulation evaluation and real-world deployment further demonstrate that a VLA fine-tuned solely on the generated demonstrations can execute language-specified grasps. Project page: this https URL.
https://arxiv.org/abs/2611.14092
We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. Sable fixes an integer function F_O with no floating-point step, in which every rounding, normalization, softmax and sampling decision is determined by its inputs, and proves it with a constraint system whose relations provably never wrap modulo the 31-bit KoalaBear prime. The proof batches zerochecks and LogUp-GKR lookups into one sumcheck per domain, and commits to every witness tensor and weight in one multilinear commitment built from Reed-Solomon codes, BLAKE3 Merkle trees and a mixed-length BaseFold opening. The prover streams the model layer by layer on one GPU. On one NVIDIA RTX PRO 6000 Blackwell GPU, Sable proves all 64 layers of Qwen2.5-32B in 92.5 s for 4,096 tokens and 398 s for 16,384 tokens, and covers Qwen3-32B and the mixture-of-experts Qwen3-30B-A3B (4,096 tokens in 121.8 and 84.7 s). A proof costs 41 to 50 times the serving time, or 6.4 cents per 4,096-token request at public GPU prices, and 40 to 64 times less time per parameter and token than zkLLM. On WikiText-2, F_O's greedy token matches the BF16 model's at 86.3% of positions; on Qwen3, F_O is within 1.6 points of BF16 on MMLU and 1.7 points on GSM8K. A verifier with a compiled core checks a proof on one CPU core in 19 to 56 s. In the random-oracle model, the interactive proof has round-by-round soundness error 2^-103. The proof is not zero-knowledge, so the system targets open-weight models. Code: this https URL
https://arxiv.org/abs/2611.14083