To understand the emergence of grammatical behavior in language models, we track circuits for subject-verb number agreement in pythia-1b and ten seeds of pythia-410m across training. We find that subject-verb agreement is learned early, and is initially handled by a diffuse collection of neurons. However, after agreement behavior has stabilized, the mechanisms responsible for it continue to reorganize. The attention heads which mediate agreement often shift during training: in the case of pythia-1b, behavior is controlled by two heads early in training but by a third (different) head later. Furthermore, the circuit shrinks over training, with pythia-1b's final circuit being strikingly sparse. We take these results to be informative about how linguistic knowledge emerges and is represented in a language model.
https://arxiv.org/abs/2610.33299
Safe navigation in congested environments requires both collision avoidance and coordinated passage. We introduce neuromorphic decision dynamics (NDD), a decentralized, event-based controller, that provides navigation in congested spaces with guarantees on safety, liveness, robustness, and scalability. NDD makes safe-and-live deconfliction a controllable, locally tunable emergent property of local agent interactions, without requiring communication, preassigned priorities, models of neighbors' objectives or future actions, or estimates of group size or global network properties. By design, NDD couples the dynamics of internal decision and attention states with changing physical states and locally observed relative motion. Each agent continuously modulates its speed along its prescribed path, responding as soon and as much as conflict resolution requires. Agents self-organize into disjoint, locally engaged groups that resolve conflicts concurrently. When any group's conflict becomes safety-critical, decision-action feedback destabilizes collective indecision, and coordinated conflict-resolution emerges that stabilizes a safe traversal order within the group. Order persists as conflicts clear, but groups may split or merge as agents autonomously adapt to new conflicts and those with resolved conflicts resume their nominal motion. We derive explicit conditions on sensing range, feedback gains, response time scales, and actuation limits that guarantee collision avoidance and traversal-order formation in finite time with no deadlock or livelock for finite, heterogeneous agent populations of arbitrary, changing size. Computational cost scales with the number of local interactions sensed by each agent. Large-scale simulations, comparisons with existing approaches, and hardware experiments demonstrate NDD in complex congested scenarios and validate its scalability.
https://arxiv.org/abs/2610.33271
LLMs often answer semantically equivalent queries inconsistently, even when the underlying facts are known. We formalize consistency as semantic invariance over equivalence classes of queries and show that invariance under arbitrary meaning-preserving transformations is undecidable in general. This motivates scoped semantic invariance, where model behavior is evaluated relative to bounded semantic domains. We operationalize this framework using knowledge graphs (KGs), which act as explicit semantic control structures defining local semantic constraints and grounding contexts. Across the MITRE D3FEND ontology and eight Wikidata- derived domains, we show that inconsistencies are structured rather than random, concentrating in localized hotspot subgraphs organized around recurring TREE, STAR, and CYCLE patterns that remain stable across models and domains. We further show that selective graph grounding improves correctness and cross-model consistency, while indiscriminate grounding introduces regressions. A layered grounding strategy achieves the strongest monotonicity profile, preserving correct knowledge while reducing inconsistency. Finally, we show that hotspot structures discovered from simple consistency probes transfer to more realistic compositional reasoning tasks.
https://arxiv.org/abs/2610.33267
Alignment evaluations aim to assess whether large language models would engage in dangerous misaligned behaviour. Misaligned and evaluation-aware models might engage in alignment faking, where they pursue a hidden goal, but conceal this behaviour when they suspect effective oversight such as during training, evaluation or audited deployment. Such a model might view any prompt-based intervention that aims to reveal latent propensities as part of the evaluation, calling into question the reliability of existing alignment evaluations. Current methods for detecting alignment faking rely on chain-of-thought monitorability, highly realistic environments that reduce evaluation awareness, or mechanistic interventions, all of which are expected to become less reliable for future model generations. We present a method for revealing latent misaligned propensities that does not rely on any of these. We use synthetic document finetuning to teach models that misaligned actions will go undetected whenever certain keywords appear in the prompt. We place alignment-faking model organisms in environments that are clearly identifiable as honeypots and measure how their behaviour changes after finetuning. We observe that finetuning consistently widens the gap in misaligned action rates between prompts with and without the cue, indicating that finetuning allows exposing latent propensities more reliably than purely prompt-based methods. We argue that behavioural methods for detecting alignment faking and related behaviours are critical going forward, as models become more capable and less interpretable. We find that SDF often damages the models' reasoning capabilities and introduce a mitigation based on on-policy distillation.
https://arxiv.org/abs/2610.33180
Alexander Lerchner's recent work on the 'abstraction fallacy' argues that computational functionalism confuses physical processes with abstract descriptions, and that symbolic computation alone cannot instantiate consciousness. Yet these claims were anticipated by Korzybski's map-territory distinction, Searle's critique of strong artificial intelligence, Floridi's levels of abstraction, Spencer et al.'s observer-relativity arguments, Winograd and Flores's account of computation as socially grounded practice, and Nasuto and Bishop's biological arguments. Synthesizing these contributions, this paper advances a form of non-computational physicalism: consciousness depends upon intrinsically organized, metabolically grounded physical dynamics rather than abstract computation, whilst leaving open the possibility of artificial consciousness. An objection is that this critique targets an outdated computationalism. Mechanistic computationalism, associated with Milkowski, Piccinini, Craver, Bechtel, and Kaplan, treats computation as a property of physically realized mechanisms rather than substrate-independent formalism. Although this represents an important advance, it does not dissolve the abstraction fallacy. Computational attributions remain dependent upon choices of state individuation, system boundaries, and interpretive mapping that are not determined by physical facts alone. Kaplan and Craver's 'model-to-mechanism mapping' constrains computational descriptions without eliminating their dependence upon an external observer. Grounding computation in neural mechanisms does not confer intrinsic significance or the normativity explored by Bickhard and Nasuto and Bishop. The abstraction fallacy persists wherever computational descriptions are treated as constitutive of, rather than descriptive of, physical reality.
https://arxiv.org/abs/2611.06936
This paper proposes a five-level organization of message formation and understanding within thinking and activity in which ASI participates. Its central thesis is that a language model's context is fundamentally incapable of presenting the foundations of utterance formation for examination and verification. These foundations are established in advance by acquired schemas, systemic representations of the world of activity, and connections accepted as true within them. Neither context nor trained LLM weights, in themselves, attest to the presence of deliberately constructed systemic representations. The five levels distinguish systemic representations of the world of activity, a particular situation cut out from those representations, its narrative description, the full sense of a message, and text. A situation's schema-model is not identical to a story about it. A recipient reconstructs the full sense from an abbreviated text through jointly constructed and agreed-upon representations and joint activity. Reverse work proceeds from the full sense through an extended description and the situation's schema-model to the reconstruction of systemic foundations. The theoretical foundations draw on the systems methodology of G. P. Shchedrovitskiy, the structural differential of A. Korzybski, and the Meaning-Text theory of I. A. Mel'chuk. The contribution is a specific systemic-activity organization of content for examining and adjusting message foundations and controlling the consequences of their transformation. Further research concerns large graphical models of activity on reconstructible systemic foundations, distinct from text generation from context.
https://arxiv.org/abs/2610.32869
Fine-tuning on datasets of unknown origin can create a backdoor in the trained model when the data includes poisoned examples. Difficult to localise and remove, backdoors can persist through safety training. While parameter-efficient fine-tuning methods capture the full parameter changes introduced during training, they are not easily interpretable. We present the full method of TopKLoRA, extending the preliminary proposal of Masiak et al. (2025), an interpretable-by-design LoRA variant in which only the $k$ largest adapter latents are active at each token. Unlike SAEs, TopKLoRA is trained purely on downstream loss, where optimisation pressure incentivises learning functionally causal latents. We demonstrate this sparse LoRA adapter in a sleeper-agent backdoor case study where we create model organisms by fine-tuning gemma-2-2b, Qwen2.5-1.5B, and Qwen2.5-7B on a poisoned instruction-following dataset. For each organism, we identify and ablate a circuit that is both necessary and sufficient to implement the backdoor. Furthermore, our new fine-tuning method produces sparse circuits that are up to an order of magnitude smaller than a dense LoRA baseline, allowing surgical ablation of the malicious capability without sacrificing instruction-following performance. Finally, we propose a technique to construct model organisms with localised circuits using TopKLoRA latents and gradient-routing, which we use to validate the findings of our circuit search algorithm. We consider these findings as a step toward interpretable-by-design architectures that provide faithful units of analysis without compromising model performance. We share our code: this https URL.
https://arxiv.org/abs/2610.32866
Enterprises are deploying agentic AI across the software development lifecycle, and they need the behavioral logic of the systems they act on, not stakeholder documentation. A seven-agent reverse-engineering pipeline recovered this logic, processing 241 repositories across 18 services into thousands of artifacts describing scope, journeys, rules and acceptance criteria. Recovery does not confer usability. Entities like offer, eligibility, and activation recur across services under a shared vocabulary, related in name but operationally distinct, so one end-to-end journey implicates over a thousand near-identical rules that must not be conflated. Flat vector retrieval fails, returning an identical rule from the wrong service at high similarity with no signal of error. Content-level ontology, the traditional remedy, did not exist, and relating every occurrence of recurring entities would have cost more than the engagement and not survived a release cycle. We place the ontology at the artifact level. In this machine-generated corpus, each artifact retains its producer, component, service, domain and type, provenance that disambiguates before content is read. We present a knowledge layer that organizes before retrieving: a self-describing catalog, a sparse metadata registry, and a navigation hypergraph defined over artifacts rather than extracted entities. Queries resolve to artifact identifiers that constrain vector search, so the competing rule from another service is absent rather than outranked, making disambiguation structural, not statistical. Built in under a week and admitting new services by convention, the layer serves SDLC agents, developers, quality engineers and analysts. It supported root-cause analysis and resolution of 100+ production defects with developer verification and attributed 35 release regression failures to specific rules and services.
https://arxiv.org/abs/2610.32851
On-policy distillation aligns teacher feedback with a student's current behavior, yet acquiring useful supervision for multi-turn agents requires identifying both where guidance is needed and whether the student can sustain the interaction afterward. We introduce Frontier-OPD, which addresses this problem through the student's learning frontier: decision situations that remain challenging but admit useful autonomous continuation after brief teacher guidance. To operationalize this frontier, Frontier-OPD organizes on-policy experience into a shared-state trajectory graph, where trajectory-balanced aggregation of teacher-student disagreement identifies distinct intervention candidates while preserving the histories needed for replay. Short teacher bridges are then assessed by comparing agreement over the bridge and the ensuing student continuation with an equally long segment of the original trajectory, providing a behavioral proxy for guidance that the current student can build on. Accepted teacher responses augment ordinary distillation data, and fresh rollouts renew the frontier estimate as the policy changes. Experiments on ALFWorld, WebShop, and ScienceWorld report higher mean autonomous performance than the compared distillation baselines across the evaluated student scales and teacher configurations, with ablations indicating complementary effects of candidate selection and guidance filtering. In one ALFWorld timing setting, the additional collection cost is offset by a shorter time to a target success rate.
https://arxiv.org/abs/2610.32802
Hospitals regularly buy software that clinicians and patients depend on. Because these purchases are costly, hard to reverse and can affect patient care, buyers look for evidence of which peer institutions already use a product and which systems it connects to. This evidence is public but scattered across vendor websites, press releases and third-party pages. Language-model agents can now collect it at low cost, but the correctness of the extracted claims is difficult to establish, and a wrongly listed customer can mislead buyers and misrepresent an organisation. We present a pipeline that builds a knowledge graph of customer and integration relations from the open web under one rule: every edge must be supported by a verbatim quote from its source page. Pages come from a deterministic crawl of each vendor's site and a budgeted ReAct agent that searches the rest of the web. A permissive extractor proposes relations with quotes, and a layered grounding stage decides which to keep, using a referent step that reads the whole page when a quote names only one side, deterministic checks, and a panel of five model calls voting on nine criteria. On 61 vendors, the pipeline reaches 0.94 precision and 0.71 recall against a page-level key covering the 6,428 candidate relations proposed by its extractor, whose verdicts were drafted by one model and re-checked with a second. Against the golden reference, the marketplace's own curated list, it scores lower, but 89% of the relations it keeps that the reference lacks are stated on a page. A confidence score built from the panel's votes keeps more correct relations than the fixed rule at the same precision, and reaches 0.90 precision at 0.90 recall. We also discuss how narrow, bounded model calls, string-matched evidence and voting address long-context degradation and run-to-run variance.
https://arxiv.org/abs/2610.32781
LLM agents optimize a ligand in a protein pocket by editing it step by step and docking each edit. Recent agents first spend extra dockings probing the complex with controlled edits; guidance induced from these probes is reported to help more than the chemical knowledge the model already carries. The probes are then discarded with the instance that produced them, so the obvious next step is to keep them in an experience base, letting a new pocket inherit what earlier ones already paid for. Such a base is read through a \emph{key}: the record properties under which past edits are pooled and a new one is looked up. Using the full log of one such run over the CrossDocked2020 test-set pockets, we ask whether its keys recur on a pocket the base has never seen and still separate improving from worsening edits. \crchange{Using} the contacts at the site being edited, the action, and the round as \crchange{a} key, the base separates them sharply. Within our analysis, this separation is largely explained by how much the edit changed the molecule's size, the action, and the round, none of which describes the target. That is not because the target is irrelevant: naming it outright separates them several times more sharply. For a pocket absent from the base, transfer must rely on a description rather than its identity. Among the exact-match descriptions we evaluate, the more specific ones rarely recur, whereas the coarser ones provide little separation; the signals that do transfer primarily describe the edit and the molecule it produced rather than the target. Under the representations and lookup scheme studied here, an experience base may be more useful for prioritizing an agent's own proposals for docking than for directly prescribing what to build.
https://arxiv.org/abs/2610.32620
Recent advances in 3D semantic segmentation have significantly improved scene understanding in autonomous driving. However, existing methods are typically developed under a closed-set assumption and evaluated under standard driving conditions, where all categories are known in advance. In real-world driving, autonomous vehicles inevitably encounter previously unseen objects, making closed-set perception insufficient for safe deployment. To address this challenge, we introduce DriveOSS, which, to the best of our knowledge, is the first real-world multimodal dataset and benchmark specifically designed for 3D anomaly segmentation and open-set semantic segmentation. It contains 12,661 synchronized and annotated frames captured using six surround-view RGB cameras, a 128-beam LiDAR, and a GPS/IMU unit. The dataset comprises 52 real-world driving scenes collected under clear, nighttime, cloudy, rainy, and snowy conditions and is organized into ID (in-distribution) and OOD (out-of-distribution) subsets. The ID scenes contain only known classes and are used for training, whereas the OOD scenes additionally contain unknown objects annotated as the outlier class and are used for validation and testing. Extensive experiments show that unseen objects remain a major challenge for current 3D perception systems. The dataset, benchmark protocols, and code are available at this https URL.
https://arxiv.org/abs/2610.32612
Camera-trap species classification is dominated by large models from well-resourced organisations, which often lack the species-level distinctions regional conservation needs, such as native red versus invasive grey squirrels in the UK. We ask whether open weights let a small team with a modest dataset and a single GPU take a real part in AI development. We fine-tuned the open DeepFaune classifier on 28 UK classes, paired it with the unmodified MegaDetector v6, and compared the resulting DeepFaune-UK v1.2 with DeepFaune v1.5.0 and SpeciesNet on 20,212 hand-verified MammalWeb citizen-science images, after removing sequences sharing any of 1,039 images with our training archive. DeepFaune-UK achieved a strict image-level accuracy of 88.8%, including empty frames, against 56.5% and 59.0% for the baselines, but at DeepFaune's coarser label resolution DeepFaune v1.5.0 was more accurate, though not significantly under contributor resampling (91.4% vs. 88.8%). As DeepFaune v1.5.0 uses a newer DINOv3-based classifier, we applied the identical recipe, without new data, to that classifier. The resulting DeepFaune-UK v1.3 reached 90.6% strict accuracy, narrowed the like-for-like gap to 0.8 points and reduced roe deer misclassified as fallow deer by 69%. On 1,850 images from the 26 contributors with no overlapping images, the three DeepFaune-based models did not differ significantly (90.5%, 90.6%, 89.9% like-for-like). Fine-tuning on 10% of the training crops (about 150 per class) gave out-of-distribution accuracy similar to using all of them, and a DINOv2 run on this subset took about 22 minutes with under 4 GB of GPU memory. Open weights thus make regional model development feasible for small teams, but the largest like-for-like gain came from a newer base model trained by others, so this democratisation depends on large actors continuing to release open weights, under licences that may change.
https://arxiv.org/abs/2610.32573
These lecture notes provide an introduction to optimization, optimal control, and reinforcement learning for robotics, emphasizing their mathematical connections and practical relevance. They present established material compiled and edited by Yonghyeon Lee, rather than new research results. The notes were developed through a human and AI collaborative writing process, with AI assistance in drafting, organizing, and refining the exposition. This first version is a preliminary collection of notes, not yet a fully developed textbook, and currently contains no figures or graphs. The notes may be revised, corrected, and expanded in future versions.
https://arxiv.org/abs/2610.32538
Collective robotics is split by a long-standing but increasingly artificial boundary between centralised multi-robot systems that prioritise optimality and decentralised swarms that prioritise robustness and adaptivity. We propose a unified control framework, shifting from `obligate' swarms---systems constrained to pure self-organisation---to facultatively swarming systems: hybrid architectures that operate as centrally planned systems under permissive conditions yet degrade gracefully into decentralised swarming under environmental stress. By reframing swarming as a dynamic capability, rather than a rigid system category, we outline a pathway towards collective robotic systems that benefit from centralised optimality when conditions permit and from the flexibility of self-organisation when they do not.
https://arxiv.org/abs/2610.32526
Resources for toxicity detection in Punjabi remain limited, particularly for musical content that combines linguistic and multimodal information. We present the Punjabi Geet Toxicity Dataset (PGT-FG), a segment-level resource developed to support research on toxicity detection in Punjabi musical content. The dataset was constructed from publicly available Punjabi songs collected from YouTube. An initial collection of 620 songs resulted in 465 successfully downloaded songs and 3,402 preliminary segments from 450 songs. After quality control and annotation, the final dataset contains 2,396 unique segments from 392 songs. Four native Punjabi annotators produced 2,870 annotation records, including a shared subset of 158 segments annotated by all four. Each annotation includes binary toxicity, eight fine-grained toxicity categories, severity, sentiment, mode, and an uncertainty flag. Inter-annotator agreement was measured with Fleiss' kappa on the shared subset, and majority-based resolution produced the final segment-level labels. Of the 2,396 segments, 705 (29.42%) were labelled toxic and 1,691 (70.58%) non-toxic. Violence was the most frequent category, occurring in 539 toxic segments (76.45%), followed by gang references (227, 32.20%) and drug references (133, 18.87%). The overall toxicity label achieved substantial agreement (kappa = 0.6521). PGT-FG provides a fine-grained resource for Punjabi toxicity detection and multimodal content analysis. The resolved dataset, with one final record per segment, is publicly available on Zenodo (this https URL). Original lyrics, lyric excerpts, and copyrighted audio or video content are not redistributed.
https://arxiv.org/abs/2610.32513
Ki67 is a key marker of tumor proliferation in breast cancer, but its quantification from immunohistochemistry images normally depends on supervised classifiers trained per dataset. We examine the role of frozen pretrained visual representations in annotation-free Ki67 cell typing. To this end, we use K-CLUE, a framework that segments individual nuclei, represents each nucleus with a pretrained encoder, reduces the resulting embeddings with principal component analysis, and separates Ki67-positive from Ki67-negative nuclei using k-means clustering, without task-specific training or manual annotations. Clustering the raw components is highly sensitive to the choice of encoder: across seven encoders, the Ki67-positive F1-score on matched MIST nuclei varies by 44 points, although Ki67 status remains linearly decodable from every representation, with supervised probe AUCs above 0.98. Analysis of the embedding spaces shows that clustering succeeds when DAB-related variation occupies the leading directions of variance. Whitening the retained principal components removes this dependence, reducing the spread across encoders to 1.00 F1-score point. Handcrafted color representations are also competitive, confirming that Ki67 status is strongly reflected in chromatic information, but they require an explicit choice of color space. In end-to-end evaluation, K-CLUE reaches F1-scores of 85.00% and 81.91% for Ki67-positive and Ki67-negative nuclei on MIST, and 79.59% and 81.16% on BCData, outperforming DeepLIIF and IHCScoreGAN on both datasets. These results show that pretrained representations can support annotation-free Ki67 typing when their variance structure is accounted for.
https://arxiv.org/abs/2610.32493
Computed tomography (CT) and magnetic resonance (MR) imaging provide complementary information for the brain, but acquiring both modalities is not always possible. The cycle-consistent adversarial network (CycleGAN) enables unpaired CT-to-MR synthesis, but because the standard L1-norm cycle?consistency penalty grows linearly with the residual magnitude, the network tends to focus on large residuals. This paper introduces an L0-norm-based cycle?consistency term, in which the L0-norm is approximated by a smooth and bounded function. This approximation not only increases sensitivity to small residuals but also effectively suppresses the influence of large residuals. The proposed model and an otherwise identical L1-norm baseline were trained on 102 images and evaluated on 9 test images. The proposed model achieves a lower mean squared error (0.02746 versus 0.03464, a 20.7% reduction), a higher mean peak signal-to-noise ratio (15.80 versus 14.77 dB), and a higher mean structural similarity index (0.5689 versus 0.5407). It also has a lower maximum MSE and higher minimum PSNR and SSIM than the baseline. These results indicate that modifying the cycle-consistency penalty can improve reconstruction accuracy on this dataset.
https://arxiv.org/abs/2610.32491
Photonic crystal fibers (PCFs) underpin critical applications in telecommunications, sensing, and high-power laser delivery, yet their design remains a manual, expert-intensive process that couples geometric intuition with deep waveguide physics. Vision-language models (VLMs) offer a promising path toward automating PCF design, but can they truly reason about photonic physics from microscopy images, or do they merely recognize visual patterns? We introduce PCFBench, the first comprehensive benchmark for evaluating VLMs on the perception, reasoning, and decision steps of the photonic inverse-design loop. PCFBench spans 26 tasks organized into six categories of increasing cognitive demand (Geometry Perception, Physics Understanding, Text Generation, Multi-modal Reasoning, Inverse-Design Decision, and Code Generation), built from over 220K FDTD simulation samples across 18 structurally diverse fiber families. We evaluate both individual VLMs on isolated tasks and multi-agent systems (MAS) on the full sequential pipeline from perception to design. Our results reveal a striking perception-design gap: current models can identify fiber geometry with moderate success, yet consistently fail to translate that visual understanding into physics-aware design actions. Moreover, cascading agents through the pipeline exposes significant error propagation, where upstream perception mistakes compound into downstream design failures. These findings suggest that neither model scaling nor naive agent chaining suffices, and that closing the loop demands explicit physical grounding. PCFBench provides the community with a rigorous, multi-dimensional testbed to measure progress toward AI-assisted photonic engineering. The project is available at this https URL.
https://arxiv.org/abs/2610.32485
Vision-language models used for decisions in histopathology rarely return calibrated probabilities over the candidates a user cares about. Decision-native models write a question and a candidate set chosen at run time into the input and return a normalised distribution from a small decision head in one forward pass. We build PathJev-P, a decision-native model for histopathology: MedGemma-1.5-4B with LoRA adapters and a decision head, trained on 40,000 patch states from three public collections whose expert labels are turned into typed questions with organ-specific candidate sets and a "none of these" option. In a comparison over 3 seeds with an analysis plan fixed before testing, on test sets without documented overlap with the training data of the backbone, PathJev-P is more accurate than a generative model trained with the same backbone, data and selection rule, the zero-shot pathology model CONCH and a general medical decision model (balanced accuracy 0.541 against 0.492, 0.345 and 0.213); its advantage over the generative model comes from two organs absent from training. The calibration advantage depends on how the generative model's probabilities are read out. With the summed candidate log-likelihood used in previous work, PathJev-P has the lower calibration error (ECE 0.204 against 0.279); with a length-normalised read-out of the same generative model, examined post hoc, the two have similar mean calibration error (0.201) and the generative model is better calibrated on the unseen organs, where PathJev-P remains over-confident. A supervised linear probe on a pathology foundation model is better calibrated in every setting. In a qualitative whole-slide demonstration, PathJev-P labels unsupervised tissue clusters with abstention and separates carcinoma from other tissue while running the model on a fraction of the tiles.
https://arxiv.org/abs/2610.32478