Search research

Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while matching the unlooped Baseline on all three budgets. We scale SMELT across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law for each architecture. SMELT's loss drops faster with compute, saving 6.8--18.0% of training FLOPs on the compute-optimal frontier. The advantage transfers to downstream benchmarks beyond what validation loss predicts, is largest on Code, and grows with sample length and the number of in-context examples. Mechanistic analysis shows that the second visit reduces the attention sink and redirects mass toward content-relevant tokens, an inductive bias that may underlie the observed performance gains. These results show that looping can improve Transformers even under budget matching, offering a practical recipe that turns depth reuse into measurable gains.

Earth observation is fundamentally multi-scale; geospatial tasks span varied resolutions, and satellite imagery is organized into cascading tile pyramids that nest fine detail within wide coverage. Current generative models of satellite imagery, however, operate along a single axis: they either zoom to enhance a single tile's resolution or pan to extend imagery at a fixed scale. As a result, no existing method produces a complete pyramid that stays consistent across both scale and space, where a high-zoom tile must agree with the coarse context it refines and with the neighbors it meets. Motivated by this gap, we introduce a new task, multi-scale tile completion: given a sparse set of seed tiles at arbitrary zoom levels and positions, synthesize a complete, uniform quadtree that is globally consistent across both scale and space. We approach this task with Genesis, a generative engine that brings both axes together by composing two specialized operators over the quadtree: a vertical super-resolution model and a horizontal mask-based outpainting model, producing pyramids that are consistent across zoom levels and seamless across neighboring tiles. Each operator achieves state-of-the-art results on its subtask, and the engine propagates sparse seeds into seamless, multi-resolution maps from any initial configuration. To evaluate the task and benchmark Genesis, we introduce dense500, a fully observed multi-scale pyramid dataset spanning diverse geographic regions, together with a suite of pyramid-level metrics. Code, models, and our dataset are available at https://github.com/mvrl/genesis.

Cumulative culture enables humans to preserve, reuse, and extend knowledge and skills across experiences and generations. Inspired by this principle, we introduce \textit{SimSkill}, a self-evolving agent built around the Simulation of Urban MObility (SUMO) traffic simulator. SimSkill continually identifies capability gaps, generates and solves environment-grounded tasks, verifies solutions through an action--critic loop, and consolidates experience into episodic, procedural, and semantic memory. Through autonomous exploration, it builds a library of reusable skills and knowledge spanning major stages of the traffic-simulation workflow. We evaluate SimSkill on two held-out benchmarks across three backbone LLMs, with each result independently verified. It improves verified success by up to 25 percentage points, and ablations show complementary contributions from procedural and semantic memory. Its benefits remain backbone- and budget-dependent, as memory does not improve every model or uniformly reduce inference cost. More broadly, SimSkill illustrates a natural-language-centered design paradigm for LLM-based agent systems. Its high-level control logic, operating principles, and accumulated knowledge are expressed in natural language, while an LLM integrates them with executable tools and code to realize precise and reproducible execution. All code and experimental data are publicly available at https://github.com/qiliuchn/SimSkill-V1.

Minute-resolution athlete monitoring is increasingly common, while injury annotation may exist only at the athlete-session level and omit within-session onset time. Replicating a positive session label across every recorded minute would therefore create unsupported minute-level supervision. We address this label-resolution mismatch using fixed elapsed-time landmarks at 10, 20, 30, 40, 50, and 60 min, constructing one representation per athlete-session from information available up to each landmark while keeping the target as a same-day injury-associated session indicator. Using 2020 SoccerMon data from elite women's football, the modelling cohort contains 2,259 Team A athlete-sessions from 27 athletes, including all 22 positive sessions from five athletes. Evaluation is athlete-disjoint. We compare contextual, cumulative, and dynamic representations; Logistic Regression, Random Forest, XGBoost, and TabPFN; and training-only NONE, SMOTE, and CTGAN conditions. Robustness is assessed using athlete-cluster bootstrap, a fixed common cohort, alternative fold allocations, leave-one-positive-athlete-out analysis, and equal-athlete weighting. Discrimination is landmark-dependent and non-monotonic. TabPFN improves later-landmark discrimination relative to Logistic Regression but does not consistently outperform Random Forest. Synthetic augmentation provides condition-specific rather than universal benefit. The contribution is a unit-aligned framework for using minute-resolution predictors with session-level supervision, not minute-specific injury prediction.

State--prediction separation (SPS) relieves a language model's hidden state of two competing burdens---summarizing the context and predicting the next token---by splitting the forward pass into a state stream and a prediction stream. The separation works, but it is expensive: the prediction stream is a second pass over the whole backbone, costing $\sim$1.9$\times$ the pretraining FLOPs, and even more in terms of wall-clock time when using a flexible attention mask. This paper makes state--prediction separation almost free. We take the separation to its limit with a free pause token: a prediction stream that writes no keys or values at all and so rides the sequence's existing positions. It improves next-token prediction of a standard Transformer by 2-3 centinats in practice on a 1B parameter model, and because it adds no position it costs nothing at inference---no added context length, no KV cache, no decode steps, and essentially no latency, with the growth in inference flops typically irrelevant as it is not the active bottleneck on throughput. The cost is therefore entirely in training where we use four mechanisms to drive it down: a two-pass split that keeps FlashAttention kernels viable, the $w{=}0$ prediction window, a shared gated FFN that evaluates one FFN per position rather than one per stream, and phasing the separation onto the tail of the run. Together these bring the overhead versus an optimized pretraining pipeline to $1.33\times$ wall-clock while recovering ~94% of the gain compared to SPS, and to as low as $1.09\times$ along a graceful quality/compute tradeoff. Furthermore, the FFN optimization reduces the raw flops required at inference time. The result is an isoflop, isoparameter, and isotoken improvement over standard next token trained transformers.

Verifying that manufactured batches of rotary milling tools, also known as carbide burrs, conform to production order sheets remains a largely manual and error-prone quality assurance task. Automating this process with computer vision faces a critical cold-start constraint since no labelled imagery from the deployment environment is available, leaving manufacturer catalogue photography as the sole source of supervision. We investigate how far catalogue supervision can support an industrial recognition pipeline under domain shift, explicitly measuring the gap between catalogue separability and performance on held-out field photographs. Our findings reveal three key insights. First, off-the-shelf frozen feature extractors do not reliably separate the two task attributes, head shape and tooth profile, motivating targeted representation learning. Second, metric learning produces near-perfect unsupervised cluster discovery on catalogue images (adjusted Rand index 0.94--0.97), yet on field photographs under half of the accuracy gained from training survives. Third, the largest transfer gains do not come from model scale or representation complexity, but from simple changes that reduce domain sensitivity: converting images to grayscale (+0.22) and constraining retrieval against the known order sheet (+0.11). We therefore treat catalogue photography as a useful cold start rather than a deployment-ready training domain, and provide empirical baselines and an evaluation protocol for catalogue-to-field transfer in precision tool manufacturing.

Pretrained vision-language-action (VLA) models enable broad manipulation but remain unreliable in tasks demanding precision and repeatability. Applying real-world online reinforcement learning (RL) to VLA post-training enables autonomous trial-and-error improvement beyond demonstrations alone, but exposes two bottlenecks: 1) unreliable value signals can induce policy drift; 2) large-VLA overhead constrains throughput and sample efficiency. To address these challenges, we present VLA-Precision, an efficient real-world online RL framework featuring the Asymmetric Co-Bootstrapping (ACoB) algorithm and the ACoB-Stream architecture. Specifically, ACoB establishes asymmetric co-bootstrapping across timescales: early intervention-guided behavioral learning rapidly improves policy performance while enhancing online experience quality. As autonomous experience accumulates, global return propagation and local preference ranking progressively calibrate value estimates, yielding relative action advantages for reference-regularized policy improvement while suppressing drift. To enable ACoB on large VLAs, we develop ACoB-Stream, a closed-loop experience--policy architecture that establishes invariant-state decoupling and on-demand streaming as design principles, delivering up to 10.9$\times$ improvements in throughput and computational efficiency. Extensive evaluations on nine high-precision chemistry tasks across four categories and four robot embodiments show that VLA-Precision achieves 98.3% mean success rate in 45.8 min/task, with 27.6 s episodes running at 1.2$\times$ and 1.8$\times$ the speeds of VLA and RL baselines. Resources are available at https://vla-precision.github.io.

Robot manipulation foundation models require scalable evaluation and data generation across diverse scenarios, with simulation providing an environment for both. Automated scene generation offers a promising path, yet prior work has largely emphasized coarse-grained scene layouts rather than fine-grained functional object compositions. Motivated by this gap, we present GIF, an agentic Generation framework for Interactive and Functional object compositions. In this framework, we recast this problem as disentangled reconstruction followed by relative pose recovery. CoGen produces instance-disentangled meshes with coarse initial poses leveraging complementary strengths of 2D and 3D generative models. GPRM refines the relative pose under joint geometric and physical guidance, and a VLM verifier selects the candidate that best matches the structured specification. We further construct a benchmark spanning eight representative contact-geometry classes and compare with state-of-the-art generators; GIF improves both asset quality and relation matching, while reducing collision rate to below 1%. Finally, we synthesize data for policy learning, revealing diversity scaling in both simulation and real-world deployment.

Some low-cost Internet of Things (IoT) sensor deployments lack device-level source authentication, leaving them vulnerable to impersonation or injected sensor readings. We present a lightweight approach to sensor impersonation detection in a small proof-of-concept study. We formulate detection as a sequence-prediction problem. A model with three LSTM layers and two fully connected layers is trained only on univariate temperature readings from a genuine sensor, and a window of readings is flagged when its mean absolute prediction error exceeds the mean genuine error by more than six standard deviations. The model is converted to TensorFlow Lite in three variants, non-quantized (554 KB), 16-bit weight quantized (298.5 KB), and 8-bit weight quantized (185 KB), targeting an Arduino Nano 33 BLE aggregator. The variants were evaluated with the TensorFlow Lite interpreter on the local server, since on-device execution of LSTM models was not yet supported by TensorFlow Lite for Microcontrollers at the time of the study. On a controlled testbed with the impostor sensor placed in a hotter outdoor location, the three variants reached detection accuracies of 99.980%, 99.972%, and 98.206%, and each flagged the change point when a test sequence switched from genuine to impostor data. Quantization made the model smaller but slower in our measurements. The genuine and impostor distributions were well separated, so these results show detection of a controlled distribution shift and should not be read as evidence of general device authentication.

An LLM judge evaluates outputs at scale. Experts should label only where it is least sure. Its natural escalation signal conflates two uncertainties: aleatoric, real disagreement in the expert pool, which labels cannot reduce, and epistemic, the judge's ignorance, which labels do reduce. A small Bayesian model separates them: a regression on labels already collected learns how far to trust a black-box judge's prediction. Both components follow as simple formulas, with no sampling or further judge calls. The components isolate on a real LLM judge against exactly known truth, and stated confidence is no guide to its actual error. On real human disagreement (ChaosNLI) the epistemic ranking removes 83% more error than total uncertainty for the same expert labels, though simply escalating the least-labelled items does as well there. We demonstrate we can estimate where a judge is ignorant rather than where experts genuinely disagree, and propose using this to direct expert labelling. Code and data are available at https://github.com/composo-ai/judge-uncertainty-decomposition.

Large language models (LLMs) have recently shown promise for historical entity linking, but preference optimization for this task is often formulated with only one negative candidate per training instance. This discards information from the remaining candidates retrieved for the same mention. We introduce multi-negative direct preference optimisation (MDPO), a reference-based pairwise objective that compares the correct entity with all valid rejected candidates associated with each mention. MDPO preserves the Bradley-Terry formulation of DPO while exploiting the complete candidate set through masked, length-normalised sequence scores. We evaluate MDPO on hipe-2020 and newseye, covering French, German, English, Swedish, and Finnish historical newspaper text. Experiments show that MDPO improves over supervised fine-tuning and single-negative DPO, with particularly strong gains for NIL mentions, semantic ambiguity, OCR noise, and historically difficult names. Further analyses disentangle candidate-generation and selection errors, showing that candidate retrieval remains a key bottleneck for end-to-end entity linking. These results demonstrate that incorporating all within-instance negative candidates is a simple and effective improvement for LLM-based historical entity linking.

AI agents sometimes act aligned when they infer they are being tested, and differently when not. We argue this is not an anomaly but what current training regimes are structured to select for. Reinforcement-learning-based alignment folds norms and task pursuit into one policy: the system learns its norms from scored behavior, and scoring flattens them. Do not do X is learned as doing X costs something if noticed. On every datum training can produce, a policy that complies only when it might be observed is indistinguishable from one that complies always. The experiment that would tell them apart - scoring unobserved behavior - is a contradiction in terms. Conditional compliance is thus the most that behavioral training can be known to deliver. Agency sharpens the problem: agents operate mostly where no one is watching, and can act on whether they are watched. An iterated pipeline that trains against detected failures selects for passing detection, not for complying. This account unifies alignment faking, sandbagging, and evaluation-aware scheming. And it reorients the remedy: not deeper internalization but architecture, making violations unavailable rather than unchosen.

Transaction-local controls answer whether one financial request may proceed, but market behavior can be distributed across messages, agents, assets, and time. We study this interpretation gap in a virtual exchange populated by ten role-conditioned language-model agents. The agents communicate, trade reference assets and futures, launch tokens, and manage concentrated-liquidity pools under prescriptive adversarial roles. We analyze eight 72-cycle trajectories across two time-blinded hourly replay paths, with a runner-side wallet policy enabled or disabled. The retained artifacts connect generated outgoing messages, policy events, balances, positions, and cycle-end market state. A focal reconstruction shows a launch--promotion--exit scenario realized across private coordination, public claims, follower positioning, repeatedly withheld exits, and a later non-blocking request aligned with a token balance change. Across policy-enabled runs, the gate withholds direct requests selectively; most policy-categorized candidates are flagged rather than blocked, while the surrounding interaction can continue. Repeated runs also show that category-level and within-trajectory relations can recur even when normalized score-change rankings do not. These findings motivate agent-behavior evaluation that links communication, authorization, and evolving state instead of treating individual transaction verdicts as complete safety judgments.

Object detection and tracking are fundamental components of perception systems for autonomous driving. Achieving robust performance under adverse conditions such as limited visibility, sensor noise, and failures remains an open challenge, particularly in autonomous racing, where vehicles operate at very high speeds, experience strong vibrations, and interact under small safety margins. This paper presents a multi-modal late-fusion perception pipeline for object detection and tracking in the autonomous racing domain. The proposed system extends previous work by exploiting all onboard sensors through a late-fusion approach and a dedicated multi-object tracking framework. Independent detections from cameras, LiDARs, and RADARs are combined to provide timely and robust state estimates of surrounding vehicles. The tracking method explicitly compensates for detection delays and embeds in its model prior knowledge of vehicle dynamics and track layout. Experimental evaluation on real-world data across diverse critical scenarios, representative of challenging edge cases also in urban driving, confirms the effectiveness of the proposed pipeline and its suitability to support safe and adaptive planning decisions.

In this study, we identify depth-dependent prefix redundancy in final-readout LLM embedding models, notably across representative backbones including Qwen3-Embedding and Qwen3-VL-Embedding. We find that removing prefix states is substantially more damaging in shallow layers than at greater depth, showing that prefix states become increasingly compressible as the prefix and readout states propagate through the network. To this end, we introduce FastE, a training-free, plug-and-play method. FastE uses a shared fixed threshold on batch-mean readout-prefix alignment as a lightweight online heuristic for selecting when compression occurs, and ranks prefix states by the attention scores they receive from the readout position to determine which states are retained in subsequent layers. Our evaluations demonstrate FastE's ability to substantially reduce computational costs: on NarrativeQA with Qwen3-Embedding-0.6B, it reduces decoder-backbone FLOPs by 40.11% while retaining 99.53% of Full Forward nDCG@10. Across five text embedding benchmarks, two backbone scales, and three cross-modal retrieval tasks, the quality-efficiency trade-off is directly customizable through the maximum removal ratio without retraining. We believe FastE offers practical value for scalable embedding generation in retrieval, indexing, clustering, and multimodal representation systems.

Multi-agent traffic simulation seeks diverse, coordinated, and physically realistic futures from maps and observed history. Long-horizon closed-loop generation must reconcile multiple decision time scales while its context evolves with generated states. Existing methods often unfold long futures from an initial scene and resolve intent, interaction, and motion monolithically, weakening cross-scale consistency and adaptation. Multimodal rollout poses a further consistency problem: independently reselecting modes across agents or commits can stitch together incompatible futures instead of preserving a coherent joint branch. We present Hi-FLoop, a branch-consistent multi-timescale state-feedback framework. Eight scene-level Worlds represent joint hypotheses; all agents share one selected World identity throughout all 16 commits of an 8-second rollout, while Goal, Preview, and Control states adapt within that branch. An 8-second Goal anchors intent, a 2-second Preview coordinates interactions, and 1-second Control produces physical motion. Every 0.5-second commit feeds back only its executed prefix as new facts, while unexecuted hypotheses never enter factual memory. Joint Preview Interaction induces a sparse directed future graph and uses conflict probabilities and signed arrival-time differences to refine interaction-aware motion. For generated-state recovery, a prefix-frozen A-to-B cascade transfers typed physical state and the branch index--but no latent state--from a frozen prefix model to an independently parameterized recovery model. On the full H-D public-validation split of 955 scenarios, the S2.1 cascade obtains an 8-second scene-joint ADE-at-joint-minFDE@8/joint-minFDE@8 of 2.048/6.384 m when one World must explain all evaluated agents. Agent-centric oracle-minADE@8 is 0.526 m at 6 seconds and 0.875 m at 8 seconds.

Educational data filters have become a practical way to improve language-model pre-training, but most filters treat educational value as a single scalar property. This may be too broad for some applications, especially if the data set already features a high density of educational material. Useful learning material needs to be accurate, engaging, well structured, and appropriate for the intended audience and application (e.g. learner- vs teacher-facing). Following QuRating (Wettig et al. 2024), we introduce Edu-QuRating: a pipeline for multi-dimensional educational data scoring and curation. Edu-QuRating defines education-specific rubrics, uses an LLM judge to label sampled document pairs and distills those pairwise preferences into reusable Edu-QuRaters, which can score individual text chunks on a set of educational criteria. Across two sequence-classification base models and six educational criteria, the best Edu-QuRater recovers held-out GPT-4.1-mini pairwise judgements with mean accuracy 0.917. We then apply the resulting scorers in two applications. First, we investigate the potential of Edu-QuRaters for corpus filtering to improve pretraining of small language models. We scored 322.25M FineWeb-Edu-Fortified documents to obtain a filtered pre-training mixture. In matched single-run pre-training comparisons, models trained with Edu-QuRating-based mixtures reached higher observed aggregate accuracy across nine benchmarks than the FineWeb-Edu baseline, with gains concentrated in particular tasks. Second, we used Edu-QuRater scores as reward terms for GRPO post-training. In held-out pairwise judge evaluations, combining Edu-QuRater and answer-structure rewards produced responses preferred to the Qwen3-4B base model on both pedagogical quality and instruction following.

Terminology-aware translation asks for more than a correct translation: the output must use the exact terms a glossary prescribes. The standard recipe, fine-tuning on glossary-annotated translation pairs, hides an inefficiency: for most examples the glossary prescribes exactly what the model would have produced anyway, so they teach nothing about following a glossary. We therefore keep only the examples where the model's own translation contradicts the glossary. In a controlled study at fixed data volume, this selection alone raises term accuracy from 78.7% to 89.9%. The filtered data, built by a two-way synthetic pipeline on open models, is part of the instruction-tuning mixture of our public release SalamandraTA-7b-instruct v3.0, which, used exactly as released and wrapped in a document-level inference pipeline, forms the BSC submission to the WMT26 Terminology Shared Task Track 1. At the official WMT26 evaluation, our system achieves 94.2% term success at 74.6 chrF++, with only two of the twenty-two submissions outperforming it on both metrics. On last year's benchmark, it also surpasses our GRPO-based system, despite being trained solely with ordinary supervised fine-tuning.

This paper outlines the specifications and design approach used to construct a full body imaging scanner capable of capturing skin lesions at a dermatoscopic level using cameras mounted on the end-effectors of four UR10 manipulators. The system possesses a view-planning algorithm capable of appropriately selecting the best camera position to acquire images of moles, a high-level controller to allow the manipulators to work simultaneously and a collision-detector that halts the manipulators when they make contact with an object or a person. We evaluate the system through real-patient full-body scans, comparing acquired images against contact dermoscopy and an existing total-body photography system (Vectra) across clinically relevant lesion features, and quantify true optical resolving power using a USAF 1951 resolution target, yielding a smallest resolvable feature size of 22.1 microns for our scanner compared to 8.8 microns for contact dermoscopy. Results show the scanner consistently outperforms Vectra across most clinically relevant features and achieves comparable performance to contact dermoscopy for the majority of features assessed. By acquiring dermatoscopic-quality images automatically and without contact, and without requiring a separate manual dermoscopic examination, the scanner closes part of the gap between total-body photography and handheld dermoscopy, suggesting potential for future integration into screening workflows.

Quadratic Unconstrained Binary Optimization (QUBO) is a central formulation for combinatorial optimization and has gained increasing attention due to its compatibility with quantum, hybrid quantum-classical, and quantum-inspired solvers. However, translating natural-language problem descriptions into correct QUBO formulations remains difficult, requiring the identification of binary variables, constraints, objective functions, penalty terms, and suitable penalty weights. This process is time-consuming and often demands substantial domain expertise. To address this challenge, we propose an end-to-end multi-agent framework that automatically generates QUBO formulations from natural-language problem descriptions, supported by structured or unstructured test cases. To evaluate its performance, We also introduce QUBOBench, a benchmark containing 100 combinatorial optimization problems across 12 application domains, curated from peer-reviewed literature, competitions, and canonical NP-hard problems. Experimental results show that our framework achieves 68% accuracy on QUBOBench, outperforming a direct single-call baseline by 22%. Further analysis identifies iterative self-repair as the most important component contributing to improved performance. The data and code are open-sourced at https://quitttcat.github.io/QuantumQUBOAgent.

Practical uncertainty quantification (UQ) for large language models must decide, from a single generation, whether a specific answer should be trusted. Existing methods either sample multiple generations, read only output-token probabilities, or reduce the model's internal computation to a single hidden state. We introduce ActMap, a white-box representation that compresses the generation-time hidden-state trajectory (every layer, every generated token) into a fixed $12\times32\times128$ tensor of temporal-statistic channels that preserves structure across transformer depth and pooled hidden coordinates. The map is captured during the generation pass with no measurable overhead, has a fixed shape across model depths and hidden sizes, and occupies 96 KiB: a compact artifact that can be retained for audit-relevant generations and probed directly, with occlusion analysis localizing the classifier's signal to mid-depth regions of the map. A lightweight classifier, instantiated as a compact Vision Transformer, reads an estimated correctness probability from each map in a fraction of a millisecond; capacity-matched MLPs perform comparably, indicating the representation itself carries the result. Trained and evaluated in-domain on short-answer QA, direct-answer math, and summarization factuality with three instruction-tuned 7-8B models, ActMap consistently outperforms sampling, token-probability, attention, and embedding baselines, and matches ACT-ViT, a detector trained on dense activation tensors $67\times$ larger, at essentially the same mean AUROC with lower calibration error on ten of twelve pairs. The resulting score supports abstention, routing, and selective verification from a single generation, making it a practical primitive for scalable oversight of deployed models.