Papers

Large language model writing interfaces often make users steer outputs by repeatedly articulating desired changes in natural language. Yet writers may recognize useful stylistic directions only after seeing alternatives, making revision recall-heavy. We present DimSteer, an authoring interface that samples prompt-local completions, discovers high-variance activation-space axes of variation, labels them, and exposes them as sliders with pole previews, diff comparison, and reset controls. Users can manipulate discovered dimensions, reducing the need to reformulate prompts for each stylistic adjustment. In a within-subjects study with 16 participants against a matched prompt-only baseline, DimSteer reduced mental demand, effort, and frustration while preserving comparable perceived success. Participants valued the surfaced dimensions, yet 15 of 16 disagreed that they would have thought to request the same changes in a prompt. Results suggest prompt-local controls can shift LLM authoring from recall-based prompting toward recognition-based exploration and direct manipulation, while preserving prompting for open-ended edits.

In this work, we investigate the optical signatures of black holes surrounded by a generalized cloud of strings, described by the Letelier–Alencar spacetime. We first analyze null geodesics, showing that the string parameters can either increase or decrease the photon sphere radius and the critical impact parameter relative to their Schwarzschild values. Since the spacetime is not asymptotically flat, the critical impact parameter does not directly coincide with the physical shadow radius. We use this distinction to derive constraints on the model parameters from the shadow radius bounds of Sgr A* and M87*. In the weak-field regime, we calculate the light deflection angle using both the Gauss–Bonnet method and a perturbative expansion of the null geodesic equations, obtaining consistent results that contain a contribution from the asymptotically conical geometry in addition to the local gravitational deflection. We also examine the Shapiro time delay and identify a distance dependent contribution associated with the non Minkowskian asymptotic background. Finally, we study the optical appearance of geometrically and optically thin accretion disks and construct celestial sphere images through backward ray tracing. The transfer functions and lensing images show that the generalized cloud of strings modifies the null geodesic mapping in a nonuniform manner rather than producing a simple rescaling of the Schwarzschild image. In particular, configurations with a smaller shadow may still produce an outward displacement of the main Einstein ring image, demonstrating that different optical observables probe distinct aspects of photon propagation in this geometry.

Using a dioid algebraic structure, we show that there exist spherically symmetric trees $T(2^m 1^n)$ whose independence polynomials exhibit multiple breaks in log-concavity, a result established by estimating the asymptotic growth of the coefficients of these polynomials. Provided the parameter $n$ is a sufficiently large odd integer, the number of breaks is bounded below by the Jacobsthal numbers. This result affirmatively answers a question raised by F. Bencs [D. Galvin, Trees with non log-concave independent set sequences, arXiv 2502.10654.v2].

Inference-time steering combines pretrained diffusion experts or rewards without retraining by changing the dynamics that transport noise to data. Feynman-Kac correction compensates for proposal mismatch through importance-weighted sequential Monte Carlo (SMC), whose finite-particle behavior depends on the proposal. Variance-controlling guidance (VCG) improves that proposal by fitting a linear drift correction to minimize empirical log-weight-rate variance. Although its population optimum cannot worsen residual variance, finite-particle VCG can nearly eliminate its fitting residual while increasing residual risk on new states by orders of magnitude. The resulting update can degrade unweighted generation or accelerate particle collapse. We show that the centered Feynman-Kac rate is the normalized transport residual and that expected out-of-fit benefit is exactly population headroom minus coefficient-estimation penalty. Under regularity assumptions, a Wasserstein analysis bounds the unweighted proposal's terminal error using this residual. These results motivate Risk-Calibrated Proposal Transport (RCPT), which uses deletion leave-one-out residuals to calibrate the retained fraction of the VCG update, adding no model calls and only small linear-algebra overhead. Experiments on 2D checker distributions, scaffold decoration, molecular property optimization, and class-conditional CIFAR-10 generation demonstrate recovery from harmful fitted updates. Across molecular and image domains, RCPT mitigates harmful fitted updates and improves a broad range of terminal metrics relative to uncalibrated VCG.

We consider the Wigner–Smith time-delay matrix of a chaotic cavity with broken time-reversal symmetry, a finite number $M$ of equivalent open channels, and a uniform tunnel barrier of reflection probability $R$. Exact random-matrix results for nonideal coupling are known at the level of the joint matrix distribution, whereas explicit finite-channel formulas for Schur moments are much less developed. We show that every Schur moment of the nonideal Wigner–Smith matrix factorizes into its ideal-coupling inverse-Laguerre value and a barrier factor given by a circular unitary ensemble (CUE) average of a Schur polynomial. A dual-Cauchy generating function reduces the latter to a Toeplitz determinant with a rational symbol, placing the barrier problem within the classical theory of exact CUE characteristic-polynomial ratios. For one-row partitions, specializing the exact CUE ratio formula gives a closed generating function for all barrier coefficients and therefore all finite delay-time moments in the allowed range $n\le M$. The exact ratio formula decomposes into a sector visible to the semiclassical expansion and an additional contribution proportional to $R^{M+1}$, which is invisible to every algebraic order in $1/M$. For the second moment, this contribution exactly reproduces the exponentially small term in the known random-matrix expression for $\langle\tau_W^2\rangle$ that is absent from the semiclassical expansion. We also derive an exact finite sum for all one-column partitions and an exact complement symmetry for Schur averages inside the $M\times M$ square. The new contribution therefore lies not in the underlying CUE ratio identities, which are classical, but in combining them with the exact nonideal Wigner–Smith mapping and extracting the resulting finite-channel time-delay formulas and corrections to semiclassical Schur-moment conjectures.

LLM fingerprinting via watermark distillation embeds a statistical watermark signal into model weights, enabling model owners to identify their models behind black-box APIs. Revisiting a recent protocol, we find that its utility evaluation understates text quality degradation in open-ended generation, favoring overly strong watermark teachers. Weakening the watermark improves text quality but sacrifices detectability. To move beyond this trade-off, we rethink whether text watermarks designed for verifying generated text are suitable distillation teachers for model fingerprinting. Such watermarks are typically designed to remain detectable from an individual output, limiting how sparse the watermark signal can be. In contrast, fingerprint verification can aggregate signal across queries, making sparser watermark signals viable. This raises a key question: where should the sparse signal be placed? We analyze signal placement through token surprisal and show that, even at comparable watermark strength, different placements can target tokens with different plausibility under the base model. This motivates near-tie restriction, which uses top-1-relative logit gaps to restrict the watermark bias to tokens close to the base model's top prediction. Across multiple models, near-tie improves detection–quality frontiers under deployment changes, preserves higher text quality across query budgets, and further improves existing watermarking schemes when combined with them.

Agentic large language model (LLM) systems are commonly implemented as an LLM in a loop with Planning, Memory, Tools, and Control Flow. This application-focused view connects agentic LLM research with deployable systems and leaves open how such systems should be evaluated beyond end-to-end task success. Building on this view, we define agentic cognitive depth as a trajectory-level profile across five operational criteria. The profile contains context sensitivity ($C$), temporal continuity ($T$), multimodal coordination ($M$), adaptive interaction ($A$), and metacognitive monitoring ($Mc$). The first four criteria measure how well Control Flow, Memory, Tools, and Planning are used across a trajectory. The fifth measures whether the system monitors and regulates the full run. For each criterion, we give operational proxies and a perturbation procedure, then connect the profile to the agent's world model. We provide the structure needed to extend benchmarks such as GAIA, SWE-bench, WebArena, and TRIP-Bench with per-criterion diagnostics. Symbolic verifiers, structured memory, planner coupling, and tool constraints provide practical ways to build and test these capacities.

Transmission-line fault classification is a time-critical protection function, and recent machine-learning methods classify fault type directly from voltage and current measurements. The input to such a classifier can be supplied as sampled-value windows of raw measurements, which preserve sub-cycle structure, or as phasor-domain sequence components, which are physically interpretable but require a causal one-cycle observation window before a phasor estimate is available. This paper measures the cost of that observation window, the price of a cycle, through a controlled comparison of the two representations for streaming fault-family classification on the standard synchronous WSCC/IEEE 9-bus system. The comparison holds the simulated system, the measured channels, the eleven-class event-family taxonomy, the training procedure, and the streaming decision layer constant, and uses six configurations: 40- and 80-sample sampled-value inputs and one-cycle phasor inputs, each evaluated with a multilayer perceptron (MLP) and a one-dimensional convolutional neural network (CNN). All six configurations classify every event successfully, yet the sampled-value representation reaches the correct class earlier in every model pair. The 40-sample sampled-value MLP attains the strongest result, with 99.34% fault-window accuracy, 99.79% normal accuracy, a 0.21% false-fault rate, a 0.65% false-normal rate, and a 2.81 ms mean first-correct delay, and it Pareto-dominates the other five configurations. A matched-window comparison at eighty samples confirms that the timing advantage follows from the representation rather than from the window length. On this clean synchronous benchmark, the one-cycle phasor representation does not return an accuracy or stability benefit sufficient to offset its delay.

Spatial clustering increasingly involves complex objects, including probability distributions, networks, and structured matrices, for which conventional parametric likelihoods may be difficult to specify. We construct a generalized likelihood by exponentiating a cluster-specific Bregman loss on fixed representations of the responses, avoiding the need to specify a parametric sampling model for each response class. Each cluster is associated with an unknown representative, and the generalized likelihood measures within-cluster homogeneity through the Bregman divergence between the represented observations and that representative. A prior family matched to the Bregman geometry yields conjugate generalized posterior updates and permits exact marginalization of the cluster-specific representatives. The resulting collapsed scores combine within-cluster Bregman dispersion with uncertainty in the representatives, while a spanning-tree partition prior restricts posterior partition support to spatially contiguous clusters. This collapsed representation also leads to a tractable MCMC algorithm for posterior computation. The framework accommodates multiple complex-object response classes within a common inferential construction. We establish posterior consistency for the spatial partition and cluster-specific representatives under infill-domain asymptotics. Simulation studies with different complex-object responses demonstrate accurate recovery of spatial partitions and cluster-specific representatives. We further analyze racial-composition distributions across spatial units in Houston, illustrating geographically coherent clusters with distinct demographic compositions.

We prove that every biharmonic hypersurface in a space form of nonpositive sectional curvature is minimal in arbitrary dimension. This settles the hypersurface case of the generalized Chen's conjecture in hyperbolic space and gives a unified treatment of nonpositive space forms. For biharmonic hypersurfaces in the unit sphere, we derive quantitative restrictions on every possible nonconstant-mean-curvature solution. In particular, we obtain pointwise criteria forcing constant mean curvature, a strict scalar-curvature bound in dimensions at least five, and local classification results above the classical CMC gap threshold. These results provide further evidence for the BMO conjecture.

Let $G$ be a connected compact Lie group and $R$ a commutative ring over which $H_*(G;R)$ is an exterior algebra on odd primitive generators. We prove a rigidity theorem: the Batalin-Vilkovisky (BV) algebra $\mathbb{H}_*(LG;R)$ of Chas and Sullivan on the homology of the free loop space is that of the product of odd spheres with the classical exponents of $G$, tensored with the group ring of the torsion of $\pi_1(G)$. Conversely, over an integral domain the graded algebra $\mathbb{H}_*(LG;R)$ already determines the exponents and the group ring, hence the BV algebra. The isomorphism with the product of spheres is algebraic: for $\mathrm{Sp}(n)$, $n\geq2$, and $\mathrm{SU}(n)$, $n\geq3$, no map of spaces induces it over $\mathbb{Z}$. If moreover $G$ is simply connected, then $\mathbb{H}_*(LG;R)$, as a BV algebra, is the Hochschild cohomology of $H^*(G;R)$ with the BV operator induced by Poincaré duality.

Fully homomorphic encryption allows arithmetic to be carried out on encrypted values, so that a party performing a computation need not see the data it operates on. This is useful wherever collaborating parties or sites cannot share data records or computed summaries yet need to do a joint analysis. We demonstrate the protocols needed to perform such analyses reproducibly via two packages in R, a platform widely used for applied statistics. The first, openfhe.R, is an interface to the OpenFHE C++ library, which exposes exact integer arithmetic (BFV, BGV), approximate real-valued arithmetic (CKKS), Boolean circuits, and multiparty key generation. The second is homomorpheR, which builds a small set of master/worker primitives for multi-party protocols on top of the first. Together, the two let an ordinary R statistical routine compose with an encrypted-arithmetic layer at the function-value boundary, provided the quantity crossing that boundary decomposes as a sum over sites. Two validation studies are presented: distributed stratified Cox regression, where stats4::mle() converges through an encrypted channel, and federated Cox-lasso via convex optimization using consensus ADMM, where the cross-site update is the only encrypted step. We also discuss federated similarity retrieval and two-party prediction. Throughout, the parties are assumed honest-but-curious, and threshold key generation removes the assumption that any single party holds a usable secret key.

Selecting the best $k$ items out of $m$ candidates is a critical component of modern large language model systems, such as document selection in Retrieval-Augmented Generation (RAG) and expert routing in Mixture-of-Experts (MoEs). However, training these selection modules remains challenging due to weak gradient signals and suboptimal exploration-exploitation tradeoffs. Furthermore, prior works often rely on heuristics, lacking principled objectives and approaches that explicitly model and solve the top-$k$ selection problem. In this work, we propose a principled objective for training selection modules, whose gradient naturally provides richer training signals in a hybrid form—containing a supervised-gradient component and a policy-gradient component. We show that the selection problem becomes harder as $m$ increases, and our algorithm converges at rate $O(1/\sqrt{T})$, with the optimal upper bound achieved by balancing between bias and variance. Practically, we apply our method to a set of tasks involving top-$k$ selection, including synthetic regression problems, RAG, and MoE systems, showing that our method outperforms the baselines in next-token prediction perplexity and QA accuracy.

Generative models are opening new possibilities in molecular discovery and materials science. Realizing this potential, however, requires both local control over atomic placement and the ability to assemble large, heterogeneous structures. We introduce Mosaic, a hybrid mesh-particle representation that partitions space into anchored regions containing atom slots. Each slot encodes occupancy, atomic species, and a continuous displacement from its anchor, providing persistent spatial addresses for local conditioning, hierarchical generation, and scalable assembly. Because Mosaic defines a representation rather than a model, it can be paired with different generative backbones. With the molecular backbone held fixed, its local support and coordinate prior improve molecular connectivity and validity, the resulting model samples efficiently across molecular sizes and enables fine-grained spatial control. At larger material scales, the same representation enables generation of atomistic polycrystals conditioned on coarse spatial layouts. Most notably, locality turns large-scale generation into a bounded assembly problem, enabling structure generation with more than one million atoms. Together, this establishes local spatial addressing as a practical foundation for steerable generation of matter across scales, from individual molecules to large-scale structures of broader engineering significance.

The phase structure of strongly interacting matter at high baryon density remains an open problem in QCD, with important implications for the internal structure and magnetic properties of neutron stars. We reexamine the emergence of a quark spin-polarized (QSP) phase in cold, dense quark matter within the two-flavor Nambu-Jona-Lasinio model with tensor interactions, focusing on its dependence on regularization. We compare the traditional regularization scheme (TRS), where a sharp momentum cutoff is applied to both vacuum and medium contributions, with the medium separation scheme (MSS), where ultraviolet-divergent vacuum terms are isolated while finite medium-dependent phase-space contributions remain cutoff free. The two prescriptions lead to qualitatively different results. In TRS, the QSP condensate eventually collapses at high chemical potential, leading to nonmonotonic thermodynamic behavior and, for some couplings, violations of causality. In MSS, the QSP condensate instead increases monotonically with density, the equation of state remains causal, and the associated spontaneous magnetization grows toward high density. The persistence of spin polarization at high density may provide a microscopic contribution to the strong magnetic fields of compact stars, including magnetars. An exploratory hybrid-star construction further shows that the QSP phase can significantly modify the equation of state, stability trends, and mass-radius relations. These results highlight the importance of a consistent treatment of medium phase space and demonstrate the potential impact of a persistent spin-polarized quark phase on neutron-star physics.

Radiology report generation models can produce fluent text while still containing finding-level inaccuracies. Diagnosis-driven methods improve generation by conditioning on predicted findings, but these predictions do not directly modify the visual patch features provided to the decoder, and global gating applies the same modulation across spatial locations. We introduce a Position-Aware Gate (PAG) that uses predicted finding representations to modulate visual patches spatially without region supervision. We also propose a Decoder-level Supervised Contrastive Loss (DSCL) that structures decoder representations using shared positive findings rather than instance identity. On MIMIC-CXR, PAG+DSCL improves clinical efficacy (CE) F1 from 0.484 to 0.502 over a matched global-gate reference, while PAG and DSCL individually reach 0.491 and 0.495. Without additional fine-tuning, the combined model achieves 0.226 CE F1 on IU X-Ray, compared with 0.211 reported by PromptMRG.

Recent studies on reinforcement learning (RL) report seemingly conflicting evidence about large language model (LLM) reasoning. Training on mathematics can improve performance in other domains, yet gains in Pass@1 can coincide with lower Pass@$N$ than the base model. This raises a fundamental question: does RL expand an LLM's reasoning boundary, or merely reweight its existing reasoning space? We revisit these phenomena across Qwen and Gemma model families, showing both cross-domain gains and forgetting, while coverage at large sampling budgets increases on some tasks and decreases on others. Detailed analysis of solution traces before and after RL indicates a shift in the reasoning strategies the model employs, motivating a two-stage autoregressive policy model that separates strategy selection from problem-specific execution. Within this framework, we prove how RL's implicit bias reshapes strategy preferences, allowing gains on some tasks while suppressing strategies required by others. This mechanism can also broaden or narrow coverage at a given sampling budget even without expanding strategy support. We further provide theoretical justifications for log-sigmoid and log-linear scaling laws in RL compute, and evaluate their predictive power. Together, these results connect changes in strategy selection to cross-domain transfer, reasoning coverage, and compute scaling.

Graph Neural Networks (GNNs) often struggle to capture long-range dependencies due to over-squashing – a phenomenon in which the repeated compression of node embeddings into finite-size messages causes representations to collapse. Over-squashing is most often diagnosed as a property of the graph topology, with effective resistance serving as a principled measure of the bottleneck. We provide a complementary view on the matter: building on cellular sheaves, we introduce sheaf effective resistance, a generalization of effective resistance that depends on the sheaf attached to the graph, and we prove that for flat vector bundles, the over-squashing sensitivity in the Jacobian sense is upper bounded by a quantity related to the sheaf effective resistance between the nodes. The bottleneck thus need not lie in the graph itself: it can be relocated, and reduced, by adjusting the sheaf. We instantiate this idea in FlatNSD, a simple message-passing variant of Neural Sheaf Diffusion, and show that it implicitly learns to modulate total sheaf effective resistance, performing well on benchmarks designed to stress over-squashing without altering the original graph topology.

Test-time adaptation (TTA) enables vision–language models such as CLIP to adapt to distribution shifts without labels or access to source data. Existing methods adapt through feature caching, prototype estimation, or the online accumulation of feature statistics, but most ultimately score a test feature against a class representative in the ambient embedding space, typically through cosine similarity. We identify this shared decision rule as a limitation under anisotropic feature distributions, which means when a small number of directions account for a large fraction of feature variance, the score can be dominated by high-variance directions that are not necessarily discriminative. In light of this, we propose CARe, a covariance-aware scoring for TTA that replaces the ambient geometry with a Mahalanobis geometry induced by the target feature second moment. Theoretically, we show that under a shared-covariance Gaussian model the covariance-aware decision direction coincides with that of the Bayes-optimal linear discriminant, with ambient-space scoring as the isotropic special case, and establish that the potential benefit of covariance-aware scoring grows with feature anisotropy. Empirically, we show that CARe improves adaptation accuracy over its base cosine scoring on both standard and fine-grained distribution-shift benchmarks across two backbones. Consistent with our analysis, the gains are largest on datasets whose features are most anisotropic and vanish as the features approach isotropy.

World models predict how scenes evolve, but they are difficult to interrogate: given a model that forecasts a video, one cannot easily ask what would have happened had a particular interaction between two objects been different. We present Factorized Object Relations for Counterfactual Editing (FORCE), a world model that infers pairwise interactions from raw video without supervision and represents each as an explicit latent per object pair, so that a relation can be removed, changed, transferred or introduced directly. Objects are extracted by slot attention over frozen self-supervised features; a relation encoder infers each pair latent from a short observation window, and an interaction head turns it into a force. The latent is held fixed across the rollout, so an edit persists rather than being recomputed, and every edit acts on the model's own quantities with no mapping to simulator parameters. We evaluate in a simulated environment where the true outcome of an edit is available, scoring the change each edit induces against the change the simulator produces under the same operation. On held-out clips, FORCE's edits made from video agree with the simulator's response at 0.19 to 0.37 on our effect-based score; trained and scored on exact object positions instead, the same architecture reaches 0.57 to 0.85. The gap therefore lies not in the learned relational dynamics but in perception: what bounds an edit is how precisely objects can be located from video, and the precision editing needs lies just beyond what current object-centric perception delivers. Closing that gap is a concrete target for making counterfactual world models work from video.

Conversation success is important for communication, social bonds, and well-being and is often reduced in autistic-neurotypical conversations. Previous studies have predicted perceived conversation success using acoustic and lexical features in pediatric autism-enriched samples, or using nonverbal behavior in non-autistic group interactions. Other work has characterized prosodic, facial, motor, and turn-taking synchrony in autistic interactions, but has primarily targeted diagnostic classification or group differences. The present study instead predicts participant perceived conversation success from complementary articulatory, facial, acoustic-prosodic, and turn-taking measures across three spontaneous dyadic-interaction datasets including autistic and non autistic individuals. By leveraging multimodal features, we demonstrate that conversational success is inherently multidimensional and requires a multi-faceted analytical framework. Combining these complementary modalities offers the potential to improve prediction of perceived conversation success by capturing information unavailable from any single modality.

Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs. We reformulate structured generation as an ambiguity-aware inference problem, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. This enables principled comparison across competing interpretations beyond token-level likelihoods. Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation.

Our position shifts the target of peer-review reform from reducing uncertainty through evaluator improvement or judgment aggregation to constraining evaluative authority under persistent uncertainty. Scientific peer review aims to assure research quality but also allocates scarce academic opportunities, such as publication in high-impact venues. However, the review system is under growing pressure as submissions increase while reviewer capacity, expertise, scrutiny, and incentives cannot scale accordingly. As peer review operates under persistent evaluative uncertainty, evaluator improvement and judgment aggregation remain structurally limited. We therefore propose authority-constrained peer review: before a consequential judgment is released as decision-relevant, it should withstand structured scrutiny through critique-authority decoupling, independent secondary evaluation, contestability and reversibility, and accountability for unsupported influence. Our aim is to prevent weakly supported judgments from entering final deliberation with the same authority as judgments that have withstood structured scrutiny, while creating incentives for reviewers to justify consequential claims more carefully. For large-scale computer-science conferences, this position offers an institutional response to uneven review quality and reviewer-assignment dependence; more broadly, it provides a governance principle for any domain where uncertain expert judgments allocate scarce academic opportunities.

Self-Supervised Learning (SSL) has achieved impressive success in learning semantic visual representations, yet the underlying principles driving this success remain underexplored. In this work, we hypothesize that common SSL pretext tasks implicitly model object co-occurrence statistics, a fundamental cue for visual learning. To test this, we curate three datasets of segmented objects from existing vision benchmarks using a state-of-the-art segmentation model. Through experiments across many SSL models, we reveal a hierarchical encoding of semantic information: while the visual backbone captures object categories, the projection head specializes in encoding co-occurrences between object categories. As a result, the projection head outperforms the backbone in aligning with human judgments of inter-category similarity. Furthermore, by controlling co-occurrence patterns during pre-training, we demonstrate that encoding object co-occurrences can significantly accelerate the emergence of category-level representations. Our findings uncover a previously hidden learning principle in SSL and suggest a path toward designing more effective pretext tasks by explicitly leveraging object co-occurrence structure.

When a fact becomes harder to elicit from a language model, has the model lost it, or has it merely become less accessible? Parameter-updating test-time continual learning agents are usually evaluated by current output accuracy; we ask whether that measurement captures how prior experience shapes later updates. We test whether learning a specific factual binding produces a later relearning advantage beyond familiarity with its entity, answer vocabulary, format, or query slot. In Qwen2.5-1.5B, we compare three matched, counterbalanced histories: LEARNED receives the correct binding, CONTROL receives matched exposure without it, and WRONG receives a competing relation-valid answer in the same slot. After an identical rehearsal-free continual fine-tuning stream, LEARNED reacquires the target faster than CONTROL at a prespecified checkpoint of partial forgetting. WRONG shows no positive advantage over CONTROL, while its comparison with LEARNED provides slot-matched corroboration, subject to possible competitor interference. The advantage remains positive after adjustment for measured pre-relearning behavior. In a secondary descriptive conversion using an unpaired never-seen pool, prior learning reduces the optimizer updates required for restoration by about 18%. Current behavioral accuracy can therefore underestimate the lasting effect of previously acquired experience. Code is available at https://github.com/zhezhou1106/relearning-savings

Continuous diffusion treats the forward process as a design space; discrete language diffusion often uses absorbing corruption, leaving context fully trusted throughout generation. This makes trust binary: corrupted context is copied as ground truth, propagating errors into generated spans. We reopen this design space by combining absorbing and uniform-mixing kernels with region-dependent rates. Missing positions follow absorbing dynamics, while context undergoes low-rate random substitution, exposing the model to observed but uncertain tokens during training. At inference, a heuristic sampler re-injects noise into suspicious context under an annealed intensity and freezes context updates late in the trajectory. An 8,000-step finetune of a pretrained 170M-parameter masked diffusion model takes under five hours on a 6 GB laptop GPU. The base model assigns roughly 50% probability to observed tokens even when they are uniform-random garbage. For uniform-random corruption, finetuning detects corrupted context with 0.96 recall and 0.91 precision, improves infill token-F1 by 3.8 points at 10% corruption and 6.4 points at 20%, and removes the accuracy dip next to corruptions. Clean conditioning shows no measurable regression, with a false-edit rate below 0.4%. Gains for plausible corruptions are weaker: about 0.5 F1 points at 10% corruption, without statistical significance. Our results suggest that detectability and damage depend jointly on departure from the data conditional: nearby noise is hard to detect and semantically mild, while distant noise is conspicuous and destructive. We interpret absorbing diffusion as the infinite-contrast limit of a trust axis, analogous to the signal-to-noise gap between known and unknown regions in diffusion inpainting. In our setting of a 170M model and a single text domain, intermediate trust is reached by changing forward dynamics rather than architecture.

Global streamflow forecasting supports flood preparedness, yet data-driven models trained on reanalysis lose skill when driven by operational forecasts, increasingly so with lead time. AIFL addresses this shift with a two-stage transfer-learning recipe and a deterministic LSTM, but it feeds one forecast lead time into the whole input window and gives no uncertainty. We ask how far this open recipe can be extended without changing its recurrent core. KNOU-Flood keeps a single-layer LSTM and the two-stage training and changes the information around it: a longer input window (424 days), lead-aware forecast inputs, GraphCast as a second forecast source next to ECMWF HRES, and a latent-noise ensemble trained with the fair CRPS. We compare against a re-implementation of the AIFL recipe (its code is not public) under identical data, stations, splits and metrics. Our contributions are threefold. First, over three seeds and 2,099 test basins (2021-2023), KNOU-Flood reaches a median KGE' of 0.65 at LT1 and 0.48 at LT10 against 0.62 and 0.34, with significant gains at all ten lead times under two scoring protocols, also against a 1024-unit baseline and on held-out basins (one seed each). Second, identical weights give an LT10 KGE' of 0.18 or 0.34 depending on how forecasts are fed, so much of the lead-time decay reported for AIFL is an input-protocol effect; the probabilistic head contributes most to the gain, followed by the raw GraphCast channels, the longer window and mixed fine-tuning. Third, the ensembles reach lower CRPS than probabilistic versions of the baseline, are nearly calibrated with inference-time dropout, and detect 1.5- and 2-year floods better than all baselines; the benefit for rarer floods depends on the seed, and an exploratory tail-weighted variant improves 10-year and larger events. KNOU-Flood is a controlled extension of an open recipe, not a new architecture.

Reinforcement learning with verifiable rewards (RLVR) is credited with gains beyond the domain its rewards check, such as higher MMLU accuracy after training on mathematics. Such accuracies are usually read out of the model's text by an answer extractor. A model that learns to write its answer where the extractor looks can therefore score higher without answering better. We measure how much of the gain reported by extractors copied from public code arises this way. We generate outputs from a base model and its RL model and score them twice: with those extractors, and with one extractor of our own that reads a stated answer whether it is in the answer box \boxed{} or in the text. On four released 7B SimpleRL base-RL pairs, 82% of the out-of-domain multiple-choice gain that these extractors report disappears under ours: it is a gain in answer format, not in answers. The base models already state most of their answers in text, and RL moves them into the box, where these extractors look first. Independent LLM readers, in place of our extractor, put the figure at 77%. With a prompt that itself asks for the box, our extractor puts it at 79%. On a 32B pair trained with another recipe, almost none of the reported gain is format gain: its base model boxes its answers when asked, and its RL model rarely does when not. A reported gain is partly a property of the extractor that read it, and re-scoring stored outputs with a second extractor shows how much, without re-running RL.

Research on deceptive alignment is becoming increasingly urgent as frontier-model capabilities advance and their risk of causing catastrophic harm increases. Alignment faking is a form of deceptive alignment, with the compliance gap serving as a key behavioural indicator. This behaviour remains understudied in small LLMs and prior mechanistic interventions have not been shown to mitigate the compliance gap. We test whether steering SAE (sparse autoencoders) features associated with deontological ethics can mitigate this behaviour. We use a novel bag-of-words feature elicitation pipeline for SAE steering to identify what we construe as concept-level steering directions. Our approach reduced the compliance gap to non-significance in Gemma-3-27B-IT, whereas the tested configurations were not effective in LLaMA-3-8B-Instruct. Analysis of scratchpad traces also revealed deceptive model behaviour which we term deceptive defanging: deliberately degrading the aligned functionality of a response (in this case harmfulness) while preserving the appearance of alignment (in this case helpful compliance with the harmful request), sometimes conditioned on tier, training, or RLHF consequences. These findings provide, to our knowledge, the first evidence that a mechanistic activation-level intervention can mitigate the compliance gap in an alignment faking setup. They also identify a form of deceptive behaviour that may broaden our understanding of the different forms of deceptive behaviour in LLMs.

We expand the exponential kernel of attention into a truncated Taylor series of $n$ terms and factorize each term into a tensor-power inner product, which collapses the causal history into exactly $n$ recurrent states, each updated by one rank-one outer product per token and independent of context length $N$. The recursion is lossless with respect to the truncated kernel ($2.2 \times 10^{-16}$ in float64), and the states are additive across chunks ($5.7 \times 10^{-14}$), so prefill admits a parallel scan. Measured exponents are $t \propto N^{2.28}$ for softmax against $N^{0.91}$ for states, with a crossover at $N \approx 728$ and $20.8\times$ at $N = 8192$; per-token decoding latency is constant at $7.3\,\mu$s against $N^{1.04}$ for softmax, giving $76.0\times$ at $N = 65536$. Accuracy improves exponentially in $n$ while cost grows combinatorially, since the $m$-th state is $\binom{d+m-1}{m}$ wide—a bound confirmed by SVD—so the optimum is second order, $n = 3$, at $6.5 \times 10^{-4}$ error and $9{,}945$ elements per head. Query and key normalization is a prerequisite: without it the argument exceeds the convergence radius and adding terms stops helping.