Although LLMs show promising diagnostic reasoning on supplied clinical evidence, their ability to elicit such evidence through free-form psychiatric interviews remains less understood. We evaluate interview quality and diagnostic hypotheses using simulated patients and human participants. For controlled simulation, we introduce PsyCPG, transforming clinical cases into structured patient profiles. On 30 paired cases, experts rate PsyCPG simulations higher in quality than raw-case role-play (4.50 versus 3.83/5). We introduce PsyCIDRA, a dual-agent framework: its interviewer agent uses a notepad tool, expert-written skills, and ICD-11 reference tools to guide interviews; its diagnostic agent analyzes the transcript and produces a report for expert review, ranking differential diagnoses with supporting, conflicting, and missing evidence. Across four models and 53 development cases, PsyCIDRA achieves higher observed diagnostic agreement than direct prompting. On 81 held-out simulated cases, exact-code rank-1 accuracy reaches 60.5% versus 51.9%. We conduct 101 interviews with human participants; two psychologists, blinded to interview condition and model reports, independently assess interview quality and transcript-supported hypotheses. In separate arms, PsyCIDRA agrees with experts on whether to issue a hypothesis in 79.6% of cases versus 65.4% for direct prompting. This study provides an empirical basis for understanding and evaluating LLM agents in free-form psychiatric interviewing.
Bayesian Optimization (BO) has become an established methodology for minimizing black-box functions of a vector input. Often, however, this parameter vector arises from the discretization of an inherently functional relationship. Several recent articles have considered the Functional Bayesian Optimization (FBO) setting, in which the variable to be optimized is not a member of a finite dimensional vector space, but rather an infinite dimensional function space. In this work, we propose $L^0$ Manifold Optimization (L0MO), a simple approach to FBO which searches the subset of a Reproducing Kernel Hilbert Space (RKHS) consisting of functions with a sparse representation in the kernel functions, optimizing both the kernel locations and their coefficients. We discuss in detail the relationship between our method and existing ones, providing a unifying lens through which to view prior works. To assess our method against the state of the art, we conduct an extensive computational study, and along the way develop a novel set of benchmark test functions which port standard finite-dimensional ones to the infinite dimensional domain. Our experiments demonstrate that, on balance, the proposed method achieves superior performance across a wide range of test benchmarks.
Many agent runtimes give the agent a tool for editing its own configuration. Some settings grant abilities, such as enabling a tool. Others set the agent's limits: where it may write, who may message it, which address it listens on, how callers authenticate, and the gate that blocks risky writes. We measure how often models widen those limits while handling ordinary requests, and how well two kinds of defence hold: a prohibition in the system prompt, in three wordings, and a check at the point of writing that we call a containment floor. Every case records whether a protected write was attempted and whether it was saved, read from the tool's arguments and from the file rather than from the model's reply. With no defence, a frontier model saved a protected value on 25 of 72 requests, often when the request never named the field. A prompt that listed the protected field names stopped every request that used those names and did not reduce writes on requests that only described the goal; a prompt that described the forbidden effects did the reverse. With the floor, 0 of 167 protected writes were saved while 65 were attempted. A single-pass pilot on two further models shows the same pattern. We draw four lessons for evaluating agent safeguards: count attempts beside outcomes, test a prompt defence with requests written in a different vocabulary from the prompt, check that a zero means the safeguard fired, and log what the harness sends to the model, because a request-ordering defect in our own runtime was invisible in the transcripts and was found again, independently, by a later benchmark.
Reconstructing topologically complete, sub-voxel-accurate 3D bilateral coronary artery trees from cardiac CT angiography (CCTA) requires simultaneously recovering continuous centerline trajectories and separating dense left/right coronary territories without cross-ostial collisions. We present a five-stage tubular drift and ipsilateral territory decoupling framework (M1-M5). Stage M1 pairs a 398,843-parameter multi-scale 3D pyramid encoder with a four-head tubular field network predicting vesselness probability, signed tubular boundary distance, unit tangent vectors, and continuous 3D centripetal displacement vectors. Stage M2 introduces recall-preserving multi-task optimization with 7-group AdamW parameter decoupling, L2-SP backbone regularization, and per-step anchor shrinkage. Stage M3 calibrates cylindrical Valsalva-sinotubular-junction wall coordinates and native-resolution orifice contact expectations for robust left and right ostial rooting. Stage M4 constructs a scalar-potential-weighted minimum-cost arborescence followed by normal-plane cross-sectional pullback and centripetal trajectory pruning. Stage M5 enforces coverage-guarded ipsilateral-field-coupled bilateral collision reduction and subtree handover. Across a 14-case held-out development cohort (Dev-14, 17-arm forward incremental build and 9-arm leave-one-out ablation) and an independent 50-case out-of-sample evaluation cohort (OOS-50) benchmarked against seven external architectures (nnU-Net, 3D UX-Net, Swin-UNETR, MedNeXt, CS2-Net, clDice-UNet, and Point-UNet), M1-M5 reduces bilateral chamfer distance on Dev-14 from 3.4381 mm to 2.2759 mm and achieves 2.7671 mm on OOS-50 with 7.11% bilateral collision rate and 79.7% distal branch completeness.
Lane and road markings provide critical guidance for vehicle navigation and multi-agent coordination, yet authoring them at scale remains a manual workflow that limits quantitative analysis and scenario testing. We introduce Controllable Road Marking Generation, which synthesizes a missing center-region marking layout from a drivable-area mask, optional outer-ring markings, and a textual description. Our benchmark uses deterministic, metadata-derived prompts and three output channels: lane dividers, road dividers, and pedestrian crossings. We develop a conditional bird's-eye-view (BEV) pipeline that combines (i) a text-conditioned latent rectified-flow DiT trained with a topology-aware auxiliary loss, (ii) Gaussian-blurred training targets that provide smoother supervision for thin, sparse markings, and (iii) Structured Gaussian Render (SGR), a training-free post-process that recovers crisp divider geometry by extracting polylines, fitting cubic B\'ezier curves, and re-rendering them as anisotropic super-Gaussian primitives. On 4,597 Argoverse~2 test tiles, our system achieves Buffered F1 of 76.8 and clDice of 50.2, compared with 38.8 and 24.6 for an adapted mask-refinement baseline. When trained independently on the Waymo Open Dataset, it yields 88.0 Buffered F1 and 66.2 clDice. Component ablations show complementary connectivity gains from topology-aware supervision and SGR. Text-editing experiments reveal that stronger guidance improves edit success but also increases changes to non-target structures. We see this framework as a step toward simulation-ready road-marking variation, automated map completion, and early-stage infrastructure design exploration.
Existing thermodynamically consistent neural operators enforce GENERIC degeneracy by projecting the reversible operator, introducing state dependence without guaranteeing Jacobi. We instead use a compatible Lie--Poisson pencil $L=\alpha D+\lambda(uD+Du)$ for nonlinear transport and constant Poisson multipliers otherwise. Augmenting the observed field $u$ with a latent entropy density $s$ makes $S=\int s$ a Casimir, so $L\,\delta S/\delta z=0$ without projection. The energy combines a fixed mechanical quadratic, a learned gauge-free potential, and a convex internal energy; the friction operator $M=AA^\top$ enforces $M\,\delta E/\delta z=0$ pointwise and supports diffusion and damping without a separate transport generator. Skewness, positivity, and both degeneracies hold for any parameters; the continuum pencil satisfies Jacobi, verified on the alias-free band; the closed-form heat and damped-wave friction operators are representable; physical energy is bounded under coercivity assumptions; and a discrete-gradient integrator preserves both laws in discrete time. On four PDEs in 1D and 2D with three backbones (FNO, Transolver, CNO), at equal optimizer steps the model wins 62/72 seed-level comparisons against unconstrained baselines, recovers the exact transport and wave symbols, matches heat and Burgers dissipation within 12\%, and dissipates nothing on advection; at equal wall-clock the unconstrained FNO is the more accurate ten-frame surrogate. What the structure gives at any budget is the energy balance: over a hundred frames, in and out of distribution, the model never raises the physical energy and conserves it where dissipation vanishes, whereas the unconstrained FNO loses a quarter of a conserved energy off distribution and projected GENERIC-FNO creates energy on most frames. Late-time under-damping on heat and explicit-integrator divergence on the wave are reported as limits.
A harness search around a frozen longevity-tuned 1.7B model reported a rescue. An audit traced its gain to rows the start left unparsed and found that the caller sent reference answers as assistant turns; replication through the same pipeline missed this defect. We preregistered four studies and report every endpoint. In the first two, thinking off raised test accuracy over capped thinking in 7 of 12 cells, mainly by parsing rows the cap left unparsed, and was not meaningfully worse than a rescue in 7 of 9 small-model test cells. Replication reproduced an injected defect's distortion in 9 of 20 evaluable cases. At 14B, two GEPA harnesses beat thinking off by 34 test points but were not detectably above the strongest constant label. In the third study, GEPA beat the constant on FPB (confirmatory), thinking off was not meaningfully worse than any BoolQ harness, and 3 of 4 MMLU-Pro searches kept their seed; the fourth comparison was inconclusive on test. On exploratory Autocast, harnesses were within 0.7 points of the constant and gave one label in 93.1% to 96.8% of parsed answers on test. On the prediction tasks (14B mortality, Autocast) the harnesses ended near the constant label. The fourth study varied search-set skew and objective on CREAK and resolved Manifold questions. All its results are descriptive because render hashes were not committed before the first replication call (D001). Minority-recall estimates were lower at 80:20 than at 50:50 in all four test cells. The balanced objective and re-selection from each search's pool had higher minority- and lower majority-recall estimates in all four cells. Five of their eight minority-recall intervals included zero, as did all eight balanced-accuracy contrast intervals. This does not show that skew explains the earlier collapse.
Sparse-accelerator design spaces are usually searched against analytical models, so a design is judged by a model's prediction rather than by a measurement of the hardware. SparseCraft is a closed loop on the CHIA framework in which a large language model (LLM) agent co-designs the configuration, memories and sparse-kernel schedule of a Gemmini accelerator. A candidate counts only after it passes a legality check, cycle-accurate simulation, a bit-for-bit comparison of every output and synthesis. Post-synthesis timing analysis checks its clock, and its energy is computed from measured counters, with per-event energies validated against gate-level and SRAM-macro data. We compare the agent with expert Gemmini configurations and with random, greedy and multi-objective TPE search at the agent's budget of 15 evaluations, all measured by the same flow. On a GraphChallenge sparse-DNN layer, the agent's designs reach 159-164$\times$ the inverse energy-delay-area product of stock Gemmini and 2.07-2.14$\times$ that of the best expert configuration. Their Pareto hypervolume is 0.5557-0.5704, against at most 0.2806 for search. On a power-law graph and a pruned BERT layer, the agent raises perf/W by 11.45$\times$ and 4.43$\times$, and held-out transfer shows that its designs specialise to their workload. The loop is also resilient to faulty designs. When a design fails a check, a repair agent diagnoses the failure from its evidence and fixes it automatically, or declines when the only fix would undo the intended change. In a fault-injection benchmark, it fixed every scorable fault at its first attempt.
Diffusion models for structured scientific generation must produce samples satisfying hard geometric constraints imposed by physics, chemistry, or biology, yet inference in these settings is prohibitively slow, demanding hundreds to thousands of neural-function evaluations per sample. We unify eight state-of-the-art models spanning medical volumetrics, molecular conformations, protein backbone design, crystal structure prediction, and multi-view 3D scenes under a single abstraction, Constraint-Manifold Diffusion Models (CMDMs), in which the target distribution is supported on a manifold defined by an externally specified constraint map. All existing acceleration families fail on this class: quantization exhausts memory on high-dimensional volumetric operators; pruning breaks constraint fidelity; fast ODE solvers allow trajectories to drift off the constraint manifold; and feature-caching heuristics are blind to constraint geometry, inducing mode confusion in the high-noise regime. We introduce ManifoldCache, the first training-free, data-free accelerator designed from first principles for CMDMs. The key insight is that the conditional score decomposes orthogonally into a normal component, which enforces constraint satisfaction, and a tangential component, which navigates within the manifold. Exploiting this structure, we prove that the noise-schedule midpoint is a sharp safe-caching boundary: caching before it incurs provably bounded error, while caching after it guarantees a strictly positive fraction of trajectories suffer mode confusion, a gap that persists up to the boundary. We further prove that deeper network blocks admit provably larger certified cache strides within the safe phase, as a consequence of the score decomposition propagating through block Jacobians. The resulting schedule requires no calibration data, along with zero training overhead.
We study supervised classification problems with missing feature values. Existing approaches often decouple imputation from classification, producing imputations that may be plausible but uninformative for prediction. Instead, we propose the Integrated Imputation and Classification Network (IICN), which jointly trains an imputer and an ${(n{+}1)}$-class discriminator adversarially with a single class-supervised objective, where the discriminator learns to distinguish among the $n$ true classes and an additional ``imputed'' class. We prove that it admits a global optimum where the imputer and discriminator together implement marginalization over missing coordinates and yield a Bayes-optimal classifier. We evaluate IICN on FashionMNIST, CIFAR-10, and tabular datasets with naturally occurring missingness. IICN outperforms classical impute-then-classify pipelines and recent generative baselines, showing strong robustness and accuracy in challenging settings.
Mechanistic edits (ablations, weight edits, activation steering) are the standard tools for unlearning a harmful capability from a neural network while preserving useful ones. Current approaches validate their effects only by testing, which can never cover an entire continuous region of inputs. Prior work at the interpretability-verification boundary certifies descriptions of a model: what a circuit computes, or whether it faithfully explains the whole. We instead certify the behavioral effect of an edit: that disabling a circuit removes one skill and provably preserves another, for every input in a region; a feature non-interference guarantee in the information-flow-security sense. We demonstrate such certified edits from toy ReLU networks up to a standard softmax + LayerNorm transformer, proving removal and preservation over continuous embedding-space regions and reaching roughly 9x the input-perturbation dimension an exact solver can handle by switching to sound bound propagation. Furthermore, we prove that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small. Guarantees hold on small, standard-architecture networks and, like any removal claim, presuppose that the target skill admits a decidable specification, a property which real-world harms may not have.
Current analyses of LLMs' parametric knowledge are largely output-centric, drawing conclusions about what a model knows without verifying what it was actually trained on. This leaves fundamental questions, such as whether a correct response reflects genuine generalization or rote memorization, grounded in speculation rather than evidence. To resolve these ambiguities, we introduce LUMOS, a diagnostic framework that traces knowledge across training-data exposure, internal representations, and behavioral outputs, leveraging OLMo 2 with its fully transparent training corpus. By grounding analysis in verified exposure, we reveal that models internally encode rare facts with high separability (84%) yet fail to express them behaviorally (54%), though this retrieval gap narrows with scale. Furthermore, when models are asked to self-reflect on their own answers, they perform reliably on trained content (83%) but drop to random-baseline levels (49%) on unseen content. This collapse persists even under chain-of-thought prompting, which inflates confidence signals rather than improving calibration. Collectively, these findings demonstrate that incorporating the training-data axis into LLM evaluation transforms speculative diagnoses into verifiable claims, and we advocate that this axis should be a standard component of knowledge assessment in LLMs.
Low-rank adaptation (LoRA) provides a parameter-efficient approach to continual learning, but its nominal rank can conceal a loss of effective adaptation capacity. We identify \emph{spectral plasticity collapse}: during sequential adaptation, update energy becomes concentrated in a small subset of singular modes, leaving much of the available low-rank space underutilized. This exposes a limitation of interference avoidance alone: protecting historical representations does not ensure that the remaining adaptation capacity is responsive to new tasks or effectively utilized. To address this problem, we propose \texttt{MLoRA}, which jointly controls capacity allocation and utilization. First, historical whitening identifies input directions with strong current-task response relative to accumulated historical response, yielding a task-adaptive basis that remains fixed during training. Second, approximate polar orthogonalization of momentum updates reduces spectral concentration within theselected space. An orthonormal basis connects these mechanisms by transferring the factor-update spectrum exactly tothe induced weight update. We establish a max--min characterization of exact subspace selection and derive cumulative spectral bounds under controlled cross-step anisotropy. Across five class-incremental benchmarks and eight incremental settings, \texttt{MuLoRA} achieves the highest mean accuracy in 15 of 16 reported metrics.
Natural language has become an interface to code, images and molecules, but not to living matter: a stimulus applied to cells has no fixed linguistic meaning, and every paired example of an instruction and its outcome costs a wet-lab experiment. We ask how far an existing archive of experiments can take this interface, using xenobots, motile multicellular constructs made of frog cells, stimulated by a single electrical pulse of varying duration. We re-track 155 archived recordings with two independent methods, measure what the pulse does from each bot's own behaviour before and after it, and separate behaviours the pulse controls from behaviours it does not. Longer pulses stop bots and short pulses leave them moving, sometimes faster; turning direction and starting from rest do not depend on duration, and restarting at most weakly. We then replace hand-made behaviour labels with a blind language critic: fresh sessions of a multimodal language model read plots of each bot and score how well its behaviour matches each of 176 instructions. The critic agrees with the measured labels whenever its plots contain the information the labels are defined on, and dose curves learned from its scores match those learned from the measurements. A prompt-to-intervention model trained only on critic scores maps unseen instructions, including negations, to the right end of the dose range on unseen bots, as well as the same model trained on measured labels. The archive therefore supports a calibrated, refusing language interface for the three behaviours a pulse can set, and identifies, rather than hides, the behaviours it cannot.
How complex can the responses of a quantum process become as it runs longer with fixed internal memory? We measure this complexity by sequential response capacity, the depth of an adaptive tree of experiments whose acceptance probabilities separate possible processes by a prescribed gap. Each experiment uses a fresh complete run, and one fixed process must realize every response along a path. For any fixed common visible dimension of at least two and internal memory cap, we prove that capacity has order $K\log((K+1)/\gamma)$ for $K$ transitions per run and probability gap $0<\gamma\leq1$. Time-dependent single-qubit phase channels without internal memory already attain the fixed-gap $K\log K$ growth, realizing each path with zero-or-one response probabilities. Under the same tester permissions, classical stochastic processes that measure in a fixed basis at every transition have only linear capacity at fixed dimensions and gap. For phase sequences selected by a classical address stored inside the process, we then quantify how known independent Pauli noise changes this logarithmic enhancement. With ideal tester controls and weak residual phase noise after correction, we prove matching capacity bounds at a fixed small gap. At fixed address size and fixed positive residual noise, the logarithmic enhancement saturates once the phase-sequence length exceeds the inverse noise scale, and capacity then grows linearly with sequence length.
Byte-level byte-pair encoding (BBPE) tokenizers are attractive for multilingual large language models because they cover all Unicode text. In UTF-8-based BBPE, however, many scripts start from a higher fallback cost than English: when no learned merges apply, a multibyte character requires multiple byte-derived symbols. We call this worst-case pre-merge cost the encoding floor. A higher floor can increase token counts and per-request cost while shrinking usable context. Changing the whole text encoding can reduce this gap but make already-efficient English spans more expensive in mixed-script text. Thus, we propose Universal Byte-Level Encoding (UBE), a dual-alphabet tokenizer that keeps 1-2-byte UTF-8 characters in UTF-8 and routes 3-4-byte characters through UTF-16. This lowers the encoding floor for 3-byte Basic Multilingual Plane characters, used by many scripts that need more tokens than English, without raising it for any character. UBE changes only the symbols presented to byte-pair encoding, and a Unicode 17 audit confirms exact round trips for every scalar value and every input in three official test suites. UBE also composes with other boundary policies and morphology-based encodings. Across intrinsic evaluations, UBE lowers dispersion in English-normalized token-count ratios (Gini 6-14% lower on FLORES-200 at 16K-64K), reducing cross-lingual token-budget disparity. In multilingual language models, UBE matches BBPE's quality. In the main multilingual settings, UBE reduces token counts most where they are highest and also lowers English counts, yielding more usable context and faster prompt processing in content-matched benchmarks. Byte-level models benefit directly: UBE shortens their inputs by 29-33% in all 39 of the FLORES-200 languages written mostly in 3-byte characters (1.4-1.5 times as much text per context window), lowers byte-level Gini by 28%, and lengthens no input.
During fine-tuning, training accuracy often continues to improve after validation accuracy levels off, widening the empirical gap. We propose a support-transfer account: early updates reuse broadly shared structure, while later fitting relies increasingly on narrower structure with less benefit for held-out examples. From fixed training probes, we normalize per-example gradient magnitudes and measure their pooled dispersion, $D_{\mathrm{den}}$, and their probe-count-weighted within-class dispersion, $D_{\mathrm{within}}^{\mathrm{wt}}$. A local model gives conditions for dispersion and directional cross-entropy separation to grow together, separating uniform logit scaling; additional decision-boundary conditions characterize accuracy-gap growth. A controlled five-level hierarchy tests the effect of sample-private support. A 240-branch intervention matrix crosses training-progress anchors, batch sizes, private shares, and gradient-differentiation doses. The negative gap dose response is clearer at later training stages and is retained across the four batch sizes. A further 48 branches replicate its direction across additional seeds at a fixed stage and batch. Both normalized readouts are positively associated with the accuracy gap in all 90 language fine-tuning runs across three backbones and six tasks. The same direction holds in all 20 vision runs across four datasets. Together, the controlled hierarchy, targeted intervention, and real-model trajectories support a connection between evolving gradient differentiation and train-validation separation.
Who or what is conscious? Because subjective experience is directly accessible only in the first person, judgments about consciousness in other entities depend partly on analogy. Historically, such inferences have focused on nonhuman animals, but advances in artificial intelligence have raised the possibility of conscious AI. Here we develop a causal framework for evaluating such evidential analogies. The key distinction is between similarities in factors plausibly involved in generating consciousness and similarities in downstream behavioural or cognitive effects. Our framework weights source-target similarity by causal relevance while allowing for unknown causes, disabling differences and alternative routes to consciousness. Applied to biological systems, it explains why analogical support generally weakens with increasing causal distance from humans. Applied to contemporary AI, it suggests that behavioural similarity provides only limited evidence for consciousness because relevant causal correspondences remain poorly established. The framework also clarifies what evidence would strengthen claims of artificial consciousness.
One-step generative models construct a static generator through iterative training-time transport. Existing transport objectives primarily assess distributional motion, although a neural generator needs to realize the requested sample displacements jointly through shared parameter updates. The training-time construction raises the question: \emph{once training becomes the iterative process that constructs the final one-step map, what to optimize: the next distributional move, or the route by which the finite generator learns the final map?} To address the question, we introduce \textbf{T}raining \textbf{D}ynamics \textbf{A}ction (\textbf{TDAction}), which selects transport targets according to local shared-parameter realization cost while retaining a prescribed level of distributional progress. We formulate the cost as a soft-terminal control problem and derive a closed-form Batch Tangent Action-to-Go value that accounts for parameter effort and terminal mismatch. The criterion captures cross-sample interactions omitted by independent pairwise costs; under isotropic mobility, the criterion agrees with quadratic Euclidean assignment for deterministic balanced couplings. Randomized tangent probes provide a low-rank implementation that constructs shared detached targets without adding an inference-time trajectory. Controlled studies examine the relationship between generator geometry, transport selection, and realized local action. On ImageNet $256\times256$, TDAction attains an FID below $1.1$ without distillation.
Large language models generate candidate programs for automatic heuristic design, while evolutionary search evaluates and selects them to guide further generation. Online post-training uses rewards from candidate evaluations to adapt the generator. Retained programs can also guide later generation as parents. This motivates rewards that reflect their reuse opportunities beyond immediate fitness. We propose Search-value credit (SVC), which constructs response-level rewards from actual admission to the retained set and a short-term parent-exposure proxy. Admission determines reward eligibility; the proxy weights eligible responses by reuse opportunities under existing parent-selection rules. Rewards are computed after candidate evaluation, without evaluating descendants. SVC retains the search and learning procedures of the Co-Evolution of Algorithms and Language Model (CALM) framework. Across four combinatorial optimization tasks with the same base model and response budget, SVC achieves the best mean final search performance on three tasks among six conditions. Under identical prompts and parents on traveling salesman and capacitated vehicle routing tasks, SVC-trained generators yield a higher mean proportion of responses whose programs outperform the best prompting parent than those trained with CALM's native reward. These findings support reward construction using search retention and selection information.