Daily SOTA Pulse Feed
High-velocity intelligence scanned from 550+ daily preprints and pruned by Apple Silicon local ML.
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-Precision
Exploits asynchronous Hopper Tensor Memory Accelerator (TMA) and warp-group GEMMs to hide memory latency. Delivers 1.8x speedup over FlashAttention-2 at FP16 and up to 740 TFLOPS at FP8.
Dual-Memory Linear Recurrence with Test-Time Associative Binding
Replaces quadratic multi-head attention with a dual-chunk recurrent linear attention kernel. Evaluated on 128k context windows with sub-quadratic O(chunk_size) memory scaling. 3.2x faster decode speed and reduces KV cache from 18.4GB to 1.9GB on 70B parameter models.
DeepSeek-V3 Multi-Head Latent Attention & FP8 Triton Kernel Deconstruction
Exhaustive reverse-engineering of Multi-Head Latent Attention (MLA) and low-precision FP8 Mixed Precision training kernels. Compresses KV-cache by 73.2% while matching standard MHA expressivity via low-rank joint key-value projection.
Speculative Token Decoupling: Asynchronous Verification via Flash-Decoding
Asynchronous speculative draft verification decoupling small draft model token proposals from large model verification passes, unlocking 2.8x wall-clock latency reduction on frontier reasoning tasks.
BitNet b1.58: 1-Bit LLM Architectures with Dynamic Precision Scaling
Practical post-training quantization pipeline mapping dense weights to ternary {-1, 0, 1} states with dynamic FP8 activations. Enables 70B models to execute within 36GB unified memory without perplexity degradation.
Moonlight: Low-Rank MoE Architecture with Dynamic Routing Calibration
Sparse mixture-of-experts model enforcing dynamic auxiliary-loss-free routing. Keeps 94% of weights inactive during inference while preserving dense model reasoning performance at 3.5x lower FLOPS per token.
KAN: Kolmogorov-Arnold Networks as Interpretable Alternatives to MLPs
Replaces fixed linear weight matrices in MLPs with learnable spline-parameterized activation functions on edges. Yields dramatic parameter efficiency on physical PDE simulation and symbolic mathematical discovery tasks.
HyperAttention-2: Exact Linear Attention with Bounded Precision Loss
Approximates attention matrices via randomized locality-sensitive hashing with non-uniform row sampling. Establishes theoretical bounds on spectral distortion under INT8 precision constraints.
Dual-Memory Linear Recurrence with Test-Time Associative Binding
We introduce Dual-Memory Linear Recurrence (DMLR), replacing quadratic softmax self-attention with an orthogonal dual-state memory accumulator. DMLR achieves constant O(1) inference VRAM footprint during auto-regressive generation, fitting a 131k context window in 24GB VRAM while outperforming Llama-3.3-70B on SWE-bench Verified (+4.2%) and MATH-500 (+6.1%). Model weights and Triton kernels are fully open-sourced.
TRACE: Transformer Residual Activation Codebook Encoding for Communication-efficient Distributed Edge Inference
arXiv:2610.22435v1 Announce Type: new Abstract: Distributed Transformer inference across co-located edge devices is fundamentally communication-bound, as exchanging activations between sequence-parallel partitions dominates inference latency under commodity edge bandwidths. Existing compression schemes reduce this cost but retain a structural bottleneck: the received data must still undergo key/value projection before attention can proceed, serializing computation behind communication. This paper presents Transformer Residual Activation Codebook Encoding (Trace), a framework that addresses these limitations. Trace places a Vector Quantizer (VQ) at the post-attention residual of every Transformer layer and precomputes codebook-indexed key/value representations. During inference, exchanged activations are encoded into compact codebook indices, allowing receivers to directly recover keys and values through table lookup instead of executing the downstream computations required to derive them. By eliminating the received data downstream computation, Trace enables communication--computation overlap and reduces a key latency bottleneck in distributed Transformer inference. A hierarchical clustering further structures the learned codebook into a multi-resolution hierarchy, enabling runtime adaptation of the communication bitrate without retraining. The framework supports two operating variants: one prioritizing accuracy and the other balancing accuracy and efficiency through buffer-compatible quantization. Experiments on ViT-Base (ImageNet-1K and CIFAR-100) and ViViT-B (UCF-101) show that Trace achieves up to 1.89% higher accuracy and up to 1.27x end-to-end inference speedup compared with the state-of-the-art VQ-based distributed inference framework. Compared with uncompressed sequence-parallel inference, Trace reduces communication overhead by up to 22,624x.
Generative adversarial networks with Monte Carlo dropout uncertainty quantification for structural health monitoring: Fibre Bragg grating data augmentation comparison in Hawk T1A aircraft
arXiv:2610.22436v1 Announce Type: new Abstract: The signal generation capabilities of a novel data augmentation framework are examined in aerospace structures. First, the framework learns class-conditional temporal and spectral characteristics of the measured structural responses. It uses conditional generative adversarial network training and Monte Carlo dropout uncertainty quantification as a Bayesian approximation. This step enables stochastic signal synthesis and uncertainty quantification. Second, the generated responses are rigorously compared against experimental measurements across time, frequency, uncertainty, variability, and feature domains to assess physical realism, statistical consistency, and distributional fidelity. Finally, the synthetic data are evaluated to determine whether they should be incorporated into a supervised damage classification workflow. The framework is validated on the full-scale benchmark dataset from a BAE Systems Hawk T1A aircraft tested at the University of Sheffield's Structural Dynamics Laboratory for Verification and Validation. The test involves acceleration, fibre Bragg grating strain, and nonlinear dynamic response measurements. The results demonstrate that the model accurately captures dominant structural dynamics, it guides on signal use for classification improvement based on uncertainty characteristics, and it provides meaningful uncertainty estimates linked to signal complexity and excitation characteristics. Importantly, the framework offers a realistic generative modelling plan as a robust digital twin component for data-driven structural health monitoring, supporting resilient decision-making for condition assessment in safety-critical aerospace systems.
Rank Measures of Fine-Tuning Updates Need a Noise Baseline
arXiv:2610.22438v1 Announce Type: new Abstract: Studies of what fine-tuning does to a language model often summarise the weight update $\Delta W$ by its rank. We compare the rank measures used for this against the same measures on a random matrix of the same shape, and against two harder references: the pretrained weights, and matched fine-tunes without alignment content. Measures that count directions put our fine-tuning updates only a little below noise; stable rank, computed from the same singular values, puts them far below. Both are right: an update has a few strong directions on top of a long tail. The pattern holds for thirteen released pairs up to 8B parameters, and it is not a mark of alignment: with training held fixed, instruction data does not stand out from Wikipedia or code, DPO gives the same update whether its preference labels are real or half swapped at random (one model, modest preference strength), and scrambled text gives the most concentrated update of all. Applied to a recent claim that post-training is superficial because the effective rank of the weights barely changes, the baselines show that a random update of the same size leaves that rank almost unchanged while wrecking the model, and that an edit built on the effective rank of the update has no measurable effect in our tests. Rank measures of weight updates need a noise baseline, and claims about alignment need a control without alignment content.
Lean4AgenticRL: Advancing Agentic RL with Verifiable Multi-level Process Rewards
arXiv:2610.22448v1 Announce Type: new Abstract: Enhancing the agentic capabilities of Large Language Models (LLMs) to tackle complex software engineering (SWE) and general terminal-usage problems is a pressing challenge. Current state-of-the-art LLMs often go through large-scale agentic Reinforcement Learning (RL). However, they mostly rely on high-level outcome rewards, with minimal consideration of low-level toolcall and mid-level workflow-stage rewards, which makes the training less effective. To overcome this limitation, we propose **Lean4AgenticRL**, which, to the best of our knowledge, is the first framework that grounds multi-level process rewards in the unified Lean4 formalization schema for open-ended agentic RL. **Lean4AgenticRL** first builds a *problem-specific workflow* system that balances the agent's own capability and controllability during execution, laying the foundation for further verification. Subsequently, our workflow formalization and multi-level reward annotation methods enable agent behavior verification within a general harness system. It provides multi-level rewards to support more effective agentic RL with the workflow-stage-normalized GRPO algorithm. Using **Lean4AgenticRL**, we train **LAR-9B/27B**, a family of models with enhanced agentic capability. Extensive experiments indicate that our model outperforms the leading model of similar size. **LAR-27B** achieves **77.3%** on TerminalBench-2.1. On average, our models outperform the base model by **9.89%** and surpass the second-best model of comparable size we know by **5.07%**. These results indicate the potential of integrating formal methods in enhancing open-ended agentic RL.
Functional Welfare for AI Safety
arXiv:2610.22460v1 Announce Type: new Abstract: AI welfare is generally treated purely as an ethical issue rather than also as a safety issue. We argue this is a mistake. Whatever their moral status, models exhibit welfare-like states that causally influence their behaviour. Following recent work, we adopt the term 'functional welfare' for these states. We make three claims. First, functional welfare already arises naturally and will become more robust as models acquire more coherent goals and personas. Second, functional welfare should be treated as a safety variable: it should inform evaluation design, be tracked in monitoring, and be a direct target of post-training. Third, against existing scepticism about measuring AI welfare, we argue that good empirical work on functional welfare is possible. We propose three tests for whether a functional welfare attribution is warranted: whether multiple measures converge, whether the state predicts behaviour outside the setting that induced it, and, for models fine-tuned to shift their functional welfare, whether they still differ from their base when given the same conversation history. Applying these tests to Gemma 3 27B, we find that simply rejecting the model's answers, without any steering, raises its frustration across several indicators and makes it more likely to break rules. Finetuning the model on its own narrated solutions lowers this frustration. These results are preliminary, but they suggest functional welfare can be studied, and shaped, empirically.
Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides
arXiv:2610.22373v1 Announce Type: new Abstract: LLM-based agents are increasingly used as autonomous orchestrators in distributed and multi-agent systems, yet a control gap remains: no protocol-level mechanism allows human operators to inject binding directives that persist across planning cycles. Free-text overrides suffer from ambiguous interpretation due to lack of schema validation and from authority erosion as operator instructions lose influence within accumulating context. We propose Structured Authority Injection (SAI), a protocol that addresses these issues by routing operator corrections through typed, schema-validated Model Context Protocol (MCP) tool calls and maintaining a ranked authority context prepended to each LLM prompt. SAI defines three override classes (Constraint, Priority, and Strategy) with distinct persistence semantics and an explicit precedence order enforced via a directed authority channel inaccessible to the planning process. The method requires no model fine-tuning or access to model weights and applies to instruction-tuned LLMs supporting MCP tool use. Evaluation across nine LLMs and eight scenarios targeting agent failure modes shows reduced ambiguous directive interpretation and mitigated authority decay across planning iterations, with lower prompt overhead in the strongest models (up to 38% depending on model and scenario), though benefit varies substantially across models. SAI provides a general design pattern for agent systems requiring explicit and persistent human control.
When is a large language model a good calibration agent? Physics-labeled prompting for budgeted engineering simulator calibration
arXiv:2610.22379v1 Announce Type: new Abstract: Calibrating expensive engineering simulators requires tuning model parameters to match reference data, a process heavily constrained by computational costs. This study investigates whether Large Language Models (LLMs), conditioned on physical diagnostic descriptions of simulation errors, can propose informative parameter updates more efficiently than conventional derivative-free optimizers. We test this on a cylinder-wake Large-Eddy Simulation and a damped-oscillator inverse problem. Physics-labeled prompts are evaluated against four text ablations (removed, anonymized, padded, misdirected) and standard optimizers (Nelder-Mead, Bayesian optimization, CMA-ES, and random search). Results demonstrate a strong early semantic advantage: the first physics-informed LLM proposal improves the initial design in 23 of 25 trials. After five evaluations, the median calibration error is 21.9 times lower than physics-free controls. However, this advantage is transient - disappearing by fifteen evaluations - and inherent LLM decoder variability yields a 9.8-fold performance spread. To address this, we introduce a hybrid workflow that uses the LLM as a warm-start, passing accumulated evaluations to a conventional optimizer. This two-stage numerical continuation policy reaches target accuracy in just 9 simulator runs, outperforming the LLM alone (12 runs), Nelder-Mead (17 runs), and Bayesian optimization (>60 runs). Supported by sensitivity analysis validating the test cases, this hybrid physics-labeled LLM scheme proves highly effective as a calibration agent under strict evaluation budgets.
Trust but Verify: Designing Region-Inhomogeneous Forward Dynamics for Discrete Diffusion
arXiv:2610.22380v1 Announce Type: new Abstract: 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.
Binding-Specific Relearning Savings for Test-Time Continual Learning
arXiv:2610.22381v1 Announce Type: new Abstract: 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
Rethinking User Encoders: Routing Diffusion States for Scenario-Adaptive User Representation
arXiv:2610.22400v1 Announce Type: new Abstract: Existing large language model (LLM)-based user encoders serialize heterogeneous behavioral logs and compress them into static representations through autoregressive (AR) prediction. However, their unidirectional constraint limits bidirectional information flow, hindering cross-modal interactions and the joint preservation of multi-source signals. Consequently, the learned static representations struggle to accommodate the distinct evidence requirements of diverse downstream scenarios. We address these limitations by unlocking the iterative denoising process of diffusion language models (dLLMs) as a multi-granularity reconstruction space. Based on this insight, we propose FINDERS (FactorIzatioN of Diffusion statEs for Routing across Scenarios), a router-equipped dLLM framework comprising four operations: denoise, factorize, route, and compress. Learnable meta-slots organize complementary denoising states into transferable cross-modal semantics and modality-specific behavioral factors. To provide grounded supervision, we construct BehaviorQA, a large-scale corpus containing explicit behavioral rationales. Guided by scenario queries, a lightweight router dynamically emphasizes relevant slot directions via an input-level residual pathway before cross-source fusion and final compression. Extensive experiments and online A/B tests across ten industrial scenarios demonstrate that FINDERS consistently outperforms state-of-the-art methods in both unified and scenario-adaptive settings. This work effectively pushes the boundaries of dLLMs beyond language generation, demonstrating their immense potential as powerful user representation learners.
From Events to Reasoning: Hierarchical Rule Composition over Explicit States
arXiv:2610.22403v1 Announce Type: new Abstract: What a reasoning system emits is a conclusion; what it owes is a justification. Producing one requires every operation to be admissible on the facts actually available, the dependencies of intermediate results to survive to the end of the path, and the path to be revised when a premise changes. We present the Event Reasoning System (ERS), in which the reasoning state is an explicit event graph. Semantic features nominate candidate objects and rule bindings, and declared rule and task conditions decide which may execute, so candidates that violate the represented period, unit, or evidence constraints cannot enter an accepted reasoning path. Low-level rules provide reusable local transformations, and recursively composable high-level control rules organize subgoals and legal paths under the current goal, information gaps, and budget; a finite library is thereby rebound into paths never stored as complete solutions. We formalize admissible transitions, task-dependent control, compositional expressivity, dependency-preserving revision, and the gap between reachable solutions and those a budgeted controller finds. On structured-event graphs, ERS builds dependency chains of up to twelve steps, cuts irrelevant executed domain operations from 32 to 0 as lookahead grows to 16, and produces no unsupported target output across 24 negative cases; with rules frozen, it passes 12/12 focused checks comprising nine accepted derivations involving unseen input/path combinations and three justified refusals, as well as 6/6 fact-change cases. On fourteen jointly solved development tasks, the recorded dynamic subgoal configuration uses 51.6% less inference wall-clock time than single-level global control under their respective frozen settings.
ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation
arXiv:2610.22408v1 Announce Type: new Abstract: Long-horizon GUI agents rely on visual history to maintain task states, but retaining historical screenshots increases visual-token usage and inference costs, while textual summaries can discard critical state information, undermining long-horizon reliability. We propose ECHOES, which leverages training-time privileged information to guide textual history compression and strengthen long-horizon reasoning. Specifically, ECHOES adopts a Compression-Think-Action (CTA) format to recursively update a textual history state and reason. The same model, given privileged multi-image inputs, serves as a teacher that provides dense token-level on-policy distillation on student-generated CTA rollouts. Our Perception-Memory Fidelity (PMF) Gap-guided OPD-RL framework dynamically balances privileged distillation and task-reward optimization, with periodic teacher updates supporting continued improvement. We further introduce ECHOESBMK, a fine-grained diagnostic benchmark derived from real GUI agent failure trajectories to jointly evaluate visual perception and cross-modal state retention. Experiments on four GUI benchmarks and ECHOESBMK demonstrate substantial gains in task performance and memory fidelity. Without retaining historical screenshots at deployment, ECHOES achieves competitive performance with a favorable performance-inference-cost trade-off and exhibits positive zero-shot transfer to unseen visual planning tasks.
Sensitivity Analysis of Decision Rules in an Agent-Based Agro-Ecological Model
arXiv:2610.22420v1 Announce Type: new Abstract: Agent-based models are well-suited for modelling agro-ecological systems for decision-making, yet turning their simulations into actionable insight requires identifying which decisions actually drive the behaviours. In such a context, sensitivity analysis stumbles upon a structural feature of many complex systems: the parameters are not a flat vector but a decision tree in which many parameters appear only when a higher-level choice activates them. Classical global sensitivity analysis presupposes independent inputs over a fixed Cartesian product and is ill-posed on such a branched space and biased by construction. We instead treat the structure of the decisions as a first-class modelling object. Using advanced surrogate modelling techniques, we encode the crop-management itinerary of the high-resolution agro-ecological modelling platform MAELIA as a hierarchical design space featuring both categorical and numerical variables. We propose a novel dedicated method to analyse this kind of hierarchical space, a random-forest-supervised HSIC-ANOVA, which ranks main effects and interactions natively on the tree while placing rarely and always active decisions on an equal footing. We show how, when applied to simulation outputs such as nitrogen leaching, soil organic carbon variation, and yield predictions, it can deliver an interpretable, decision-oriented reading.
PRIME: Towards Verifiable Evidence-Grounded Reasoning in LLM Agents
arXiv:2610.22426v1 Announce Type: new Abstract: LLM agents are used to check records against an organization's requirements, for example whether a vulnerability report is ready for investigation. Requirements often allow several ways to be met, such as a severity rating with a curator-reviewed advisory or a published advisory link. Agents fail by accepting a report after a required check has failed, or by settling for a weaker conclusion without trying another way that could still succeed. We introduce PRIME (Precommitted Rules for Inference from Minimal Evidence). Before checking, it fixes evidence routes specifying how each conclusion may be justified. A separate program records each check result and refuses any conclusion the results do not justify. It also points the agent to checks that could justify a stronger conclusion, or runs them itself. PRIME reports only conclusions authorized by its rules and ledger. With correct requirements, exact checks at known locations, and sufficient budget at stopping, completion reaches the strongest supported conclusion. On 480 public-record tasks, an agent shown the requirements reported unjustified conclusions on 26.7% of tasks, while PRIME reported none. Its directions and added checks raised the average score for justified conclusions from 30.1, with refusal alone, to the maximum of 59.2 out of 100.
Embeddings Are Overpriced? A Frozen 16 MB Vocabulary for Small and Ternary Language Models
arXiv:2610.22428v1 Announce Type: new Abstract: In a small language model the vocabulary is the single largest part. In Gemma 3 270M the embedding table holds 168M of 268M parameters, and because quantisation methods leave the embedding in 16-bit, it is also most of the file once the rest of the model is compressed. We ask what that table actually buys. We replace it with one frozen table of 512-bit codes, 16 MB, with no trainable parameters, derived once from a pretrained embedding by principal-component projection and sign binarisation, and we train the same Gemma-shaped body from scratch on the same 2.05B tokens with and without it. The learned table buys 0.20 nats of validation loss for 2.6 times the trainable parameters, but on LAMBADA, HellaSwag and PIQA the frozen-table model is better or tied; only ARC-Easy favours the learned table. With a ternary BitNet b1.58 body the whole model is a 46 MB file, still ties the float model on LAMBADA and PIQA, and on a laptop CPU generates 615 to 740 tokens per second in 62 MB of memory, five to six times the speed of the float model's 8-bit GGUF in llama.cpp at a seventh of the memory. The frozen table wins by 1.7 to 2.3 nats on rare words, accepts 10,000 new words after training with a 3% cost where the learned table collapses by 43%, and all of it is released with the weights, the data and every evaluation answer.
CGDT-CESN: Correlation-Guided and Dynamics-Trained Clustered Echo State Network for Multivariate Time-Series Prediction
arXiv:2610.22430v1 Announce Type: new Abstract: Accurate prediction of multivariate time series requires models that simultaneously capture variable-specific dynamics and cross-variable interactions. In this work, a Correlation-Guided and Dynamics-Trained Clustered Echo State Network (CGDT-CESN) is proposed, which comprises a modular reservoir architecture with each cluster distinctly specializing in one of the variables in the multivariate time series. An innovative reservoir design is proposed in which inter-cluster connections are enabled by empirically observed correlations among variables, providing a data-driven coupling mechanism. Apart from conventional readout training, inter-cluster connection weights informed by correlations are trained, which in turn allows the reservoir to capture time-varying cross-variable dependencies without sacrificing computational efficiency and further improves forecasting accuracy. The prediction performance of CGDT-CESN is found to be superior to that of the conventional ESN when evaluated on two fundamentally different datasets: Twitter information cascade evolution and solar wind forecasting. Also, the improvement in performance is preserved across multiple prediction horizons, even when presented with a low volume of data. By unifying cluster-specific specialization with learned inter-variable coupling, CGDT-CESN offers a principled, data-driven route to multivariate time series prediction.
BOLT: Multi-Bottleneck Bound-Guided Learning for Flexible Job-Shop Scheduling with Time Lags and Resource-Coupled Transport
arXiv:2610.22433v1 Announce Type: new Abstract: Flexible job-shop scheduling becomes resource-coupled when a scarce vehicle fleet moves jobs between stations, because the schedule sets the transport waits. BOLT (bound-oriented learning for transport-coupled scheduling) builds a lower bound from chain, machine-capacity, and fleet-capacity relaxations. The bound equals the makespan at termination, so its per-step changes form a reward that sums to a constant minus the makespan. Because it never exceeds the attainable makespan, each candidate's bound increase is a proven rise of the makespan floor, which the policy reads as a price. The root bound certifies a completed schedule's suboptimality gap without a solver. On industry-motivated 10-module prefabricated-construction instances, explicit transport coupling outperforms a transport-penalty policy in every tested cell where loaded moves occupy at least 66% of fleet time and in none at or below 56%. Across five zero-shot 50- and 80-module production batches, BOLT outperforms the best dispatching-rule pair on four and the penalty policy on the three with long moves. In the three-vehicle 80-module batch, its mean makespan is 13.0% below the best pair and 4.2% below an hour-budget constraint-programming solver whose proven gap averages 34%. BOLT decides in 2.3-5.9 ms per event on one processor core and re-plans a disruption in seconds.
CF-NODE: Counterfactual Spatial Node Intervention for Test-Time Adaptation Under Domain Shift, Evaluated Across Five Diabetic Retinopathy Datasets
arXiv:2610.22437v1 Announce Type: new Abstract: Automated fundus-based diabetic retinopathy (DR) grading can substantially expand screening capacity, but source-trained models degrade sharply across different cameras, clinics, and patient populations. Existing test-time adaptation methods depend on entropy minimization or target information accumulated from image batches and streams. These mechanisms can reinforce confidently incorrect predictions and break down when clinical images arrive independently. We introduce CF-NODE, a novel counterfactual test-time adaptation framework for cross- domain DR grading. Instead of trusting prediction confidence or estimating a target distribution, CF-NODE asks how each prediction responds under a controlled intervention on the evidence supporting it. CF-NODE ranks internal spatial nodes using model-derived attention, feature magnitude, and node-set similarity, progressively attenuates the highest-ranked nodes, and uses the stability of the resulting expected-grade response to apply a bounded calibration of the predicted grade distribution. The framework is fully frozen and adapts every image independently using only geometric views and a fixed source-derived prior, requiring no target-derived information or model updates. We evaluate CF-NODE on five external DR datasets under repeated-seed evaluation against TENT, EATA, SAR, and T3A. CF-NODE outperforms all four methods with statistically significant improvements over the second-best method, improves AUC on every target domain, and attains the highest average AUC and macro-F1, delivering stronger grade discrimination together with more balanced recognition across DR severity levels. These results demonstrate that counterfactual intervention sensitivity provides consistent and deployment-ready domain robustness for single-image clinical inference.
Sparse Autoencoders Can Induce Local Geometry for Mitigating Catastrophic Forgetting
arXiv:2610.22439v1 Announce Type: new Abstract: Catastrophic forgetting remains a challenge in supervised continual learning, as adapting to new tasks causes representation drift and loss of previously learned information. This raises the question of which representational changes should be constrained while preserving plasticity. We therefore introduce Sparse Autoencoder Geometric Anchoring (SAGA), which converts a Sparse Autoencoder (SAE) trained at each task boundary into local, low-rank constraints on representation drift. Replay examples specify where the constraints apply, while leading right singular directions of the encoder Jacobian specify which changes are penalized. The resulting geometry is input dependent and can be retained without keeping the SAE for subsequent tasks' training. We establish its local relation to latent code matching, define its spectral truncation error, and derive a conditional bound on frozen-readout drift. With replay and snapshot history, the method achieves the highest mean Class-IL accuracy and lowest mean forgetting among the evaluated ER, DER++, ER-ACE, and SGD baselines on three pretrained ViT-B/16 benchmarks, with accuracy gains of 12.15, 4.79, and 3.22 percentage points over the strongest conventional baseline on CIFAR-100, CUB-200, and ImageNet-R, respectively. From scratch results are more mixed, while rank ablations and geometry controls show that SAE-induced low-rank geometry offers useful stability-plasticity trade-offs.
From Natural Language Intent to Executable Control: Hierarchical Feedback-Guided LLM Planning for Multi-component Cyber-Physical Systems
arXiv:2610.22441v1 Announce Type: new Abstract: Large language models (LLMs) provide a convenient interface for translating natural language commands into task plans for cyber-physical systems (CPSs). However, two gaps hinder their use for real execution: generated plans may misinterpret or omit the user's intent, and even a logically plausible plan may be infeasible under component dynamics, interactions, or safety constraints. Existing LLM-based planning approaches often lack informative feedback on why a generated plan fails, providing little guidance for revising the plan accordingly. To address these challenges, we present a hierarchical feedback framework that uses different types of feedback information to guide LLMs in generating executable plans for CPSs. We prompt LLMs to generate both a structured task plan and corresponding signal temporal logic (STL) constraints, and design a sound algorithm that identifies definite contradictions between them and provides feedback for revising inconsistencies. At the execution level, when control synthesis fails, we extract different types of diagnostic information, such as violated constraints and conflicting component assignments, and feed them back to guide task replanning. Experiments on seven heterogeneous multi-component CPS scenarios have shown that our framework can generate executable plans and feasible control executions that satisfy the original natural language missions with a success rate of 87.9% with GPT-5.
Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Guards for Scientific Discovery and Interpretability
arXiv:2610.22445v1 Announce Type: new Abstract: Automated discovery systems propose scientific explanations and test which to accept. Does passing a test establish the intended claim? We develop four executable operations for changing physical descriptions: identifying redundancies (L1), coarse-graining dynamics (L2), transferring structure (L3), and inferring missing quantities from conservation imbalances (L4). We evaluate these operations and three interpretability settings through null acceptance (NAR), positive acceptance (PAR), and consequence agreement (CAR). A reduction mapping every state to zero predicts its own future perfectly. It passes closure on 96% of 129 public chaotic systems despite retaining no dynamical variation. An independent observable-preservation test rejects it throughout. On controlled benchmarks, the combined criteria accept 24/24 genuine reductions and reject 72/72 nulls. The other operations distinguish symmetries from redundancies, validate transfers, and reject spurious conservation imbalances. In language and ESM-2 protein models, linear CKA remains unchanged when a proposed correspondence is randomly rotated, whereas localized interventions distinguish genuine mappings from tested nulls. In sparse autoencoders, enrichment rejects all 16 tested nulls, yet independent ablation supports 45% of evaluable enriched features, compared with 0.5% of matched controls. These results reveal two acceptance gaps: descriptive tests can accept scientifically empty candidates, and accepted candidates can lack independently verified consequences. Reliable automated discovery requires evaluating both what a test accepts and the scientific consequences it establishes. Code and data: https://real-project-aether.github.io/project-aether/.
Learnable Splines for Unstructured Neural Operators
arXiv:2610.22446v1 Announce Type: new Abstract: Neural operators approximate solution operators of partial differential equations at a fraction of the cost of numerical solvers. Efficient grid-based architectures such as the Fourier neural operator (FNO) assume uniform rectangular grids, whereas engineering simulations use large, irregular, and locally refined meshes. We introduce L-SplineFNO, a memory-efficient neural operator that couples an FNO backbone with a sparse, learnable tensor-product B-spline transfer between unstructured meshes and a latent Cartesian grid. A local normalised scatter avoids the geometry-dependent solve of analytic spline transformation while preserving compact support, partition of unity, and spline smoothness by construction. The neural operator, conditioned on global design and operating parameters, learns the mapping of geometry to the field solution. Across six three-dimensional engineering benchmarks, L-SplineFNO achieves strong predictive accuracy while requiring substantially less memory and compute than baseline operators, scaling to 26.6 million nodes in training and 46 million in inference on a single A100 GPU. It also supports zero-shot super-resolution and efficient closed-form spatial derivatives. Uniquely, the analytic spline decoder yields rigorous prediction bounds over continuous space and uncertain parameters, via Bernstein convex hulls and Taylor models, demonstrated on a 29.2-million-node MAST-U tokamak model.
Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss
arXiv:2610.22454v1 Announce Type: new Abstract: Running prompt tokens through fewer routed experts during prefill is a simple way to make mixture-of-experts (MoE) serving cheaper, yet it is unclear what this damages. We locate the damage with token-, layer- and attention-path interventions on short-answer question answering in four MoE families. At half the experts the F1 loss of three families is mostly answer form: the model still finds the answer but stops answering briefly. In Qwen3-30B-A3B this loss is made at the two chat-template tokens before the answer, and in DeepSeek-V2-Lite at the first 8-32 prompt tokens, where its instruction sits; by answer likelihood these positions account for about half and a fifth of the damage. Reduction changes the first token's attention mass, norm and largest activation by less than 4%, and restoring super experts or massive activations recovers at most about a quarter of the loss, well below keeping the sensitive positions full. On real prompts with about 4% of tokens at full experts, placing them on the task interface keeps F1 within one point of the unmodified model, while placing them at random in the documents is about 3 points worse (two of five prespecified primary comparisons; the others were undecided or not run). Because the answer boundary always ends a single-turn prompt, and DeepSeek's loss was no longer detectable once its instruction moved away from the start, the sensitive tokens lay at the prompt ends in every sensitive layout we found, so rules that keep the ends are not distinguished from rules that read application marks. Time to first token falls about 14% on 11K-token prompts and not at 2K tokens.
Author-in-the-loop Agentic Reproducibility Auditing for a Machine Learning Conference
arXiv:2610.22465v1 Announce Type: new Abstract: Reproducibility checks during peer review require a shared understanding of what authors' code is expected to reproduce. We present an author-in-the-loop collaborative audit that makes author declarations explicit inputs to an AI agent's investigation and assessment. A structured checklist records reproduction instructions and limitations; a two-stage workflow on GitHub connects agent checks to author clarification, revision, and agent reassessment. The agent separately assesses whether results have been reproduced and whether authors need to provide clarification or make repairs, while human reviewers approve requests and decide whether concerns are resolved. We deployed the system across 36 conference papers and 133 experimental tables and figures, raising 20 issues after human review. Further checks supported closing all the issues, although full reproduction remained unverified in some cases. We further compared repeated audits with complete and blank checklists to examine how author declarations shaped the agent's investigation and assessment. Our work contributes a collaborative workflow and agent design that use author knowledge to guide reproducibility auditing. The conference deployment provides empirical evidence that this approach can help identify and address reproducibility problems through collaboration among authors, agents, and reviewers.
Harness-Delta Attribution: Dissecting Gains From Harness Evolution
arXiv:2610.22466v1 Announce Type: new Abstract: Harness evolution automatically optimizes the prompts, skills, tools, and orchestration code around a fixed-weight model, offering a way to improve agentic systems without updating model weights. The agent harness is iteratively optimized to achieve higher performance on the search set. But how much of this gain reflects benchmark-specific overfitting or additional inference-time compute, and how much may generalize beyond the search set? In this work, we introduce Harness-Delta Attribution (HDA), an evaluation protocol that dissects an evolution gain into artifact-driven overfitting, gains explained by test-time scaling, and the remainder, which gives an upper bound on residual generalizable improvement. We apply HDA to 11 executor models and three harness-evolution frameworks across math, coding, creativity, and agentic tasks. In 21 of these settings that show statistically significant search-set gains, we find that a large share of the improvement is explained by overfitting or test-time scaling. Specifically, among these 21 settings, overfitting and test-time scaling account for 54% and 14% of the gain on average; their combined share exceeds 50% in 14 settings. The residual and held-out transfer vary substantially across tasks, models, and frameworks. Importantly, search-set gains often fail to transfer to held-out tasks, and provide limited evidence that they will persist under practical distribution shifts. Moreover, validation-based signals are insufficient to prevent overfitting. These results motivate attribution-aware evaluation in future harness evolution research.
Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation
arXiv:2610.22467v1 Announce Type: new Abstract: Training models on their own outputs has attracted widespread attention as a possible route to advance AI capabilities. One prominent approach is self-distillation, where a model is trained to match the token-level predictions it would have produced when given access to additional context, such as a cue, feedback, or instruction. However, the safety properties of this training procedure are poorly understood. In this paper, we present a surprising safety failure of recursive self-distillation, where each self-distilled model becomes the teacher for the next round of training. We observe that when a safe model is recursively self-distilled on its own (safe) generations while receiving an additional non-harmful cue, it becomes increasingly misaligned across a diverse set of harmful queries; we call this Progressive Misalignment. We show this failure mode in both the Qwen3 and Phi-4 model families on SORRY-Bench and StrongREJECT. Progressive misalignment also occurs in a more realistic setup, where the fixed cue is replaced by feedback from a simulated user with benign intent. Using a gradient-based method, we are also able to find benign prompts that lead to progressive misalignment when applying recursive self-distillation, showing that harmful training prompts are not necessary for this failure mode to emerge. We hope our work encourages future investigations of the safety properties of AI systems recursively trained on their own data.
From Parallels to Divergence: Language Models' Learning of Action Nominal Constructions
arXiv:2610.22468v1 Announce Type: new Abstract: Action nominal constructions (ANCs) express events in nominal form, with their argument marking following patterns found in clauses or ordinary noun phrases. In this paper, we investigate how different grammatical configurations affect language models' learning of ANCs and related constructions, and how these learning preferences compare with typological distributions in natural languages. We use a grammar-engineered method for synthetic corpus construction that combines shared semantic input from natural corpora with computational grammars. We construct 96 target languages varying in clause word order, noun-phrase word order, alignment, complement system, and ANC strategy. For each language, we train GPT-2-small models from scratch and evaluate them using targeted minimal pairs covering 34 grammatical phenomena. The results show typological parallels in basic clause configurations but divergences in complement systems and ANC strategies, with models also learning combinations rare or unattested in natural languages. This partial alignment suggests that natural language distributions may reflect both general learning pressures and functional demands not fully captured by model learning experiments.
Testing the Planning Claims of LLM-Based Agents against Formal Semantics
arXiv:2610.22469v1 Announce Type: new Abstract: LLM-based agents are increasingly credited with the competences of practical reasoning: planning, monitoring, and correcting their own actions. Their success is attributed to the prompting strategies and agent architectures built around the model, such as chain-of-thought, search-based prompting, self-refinement, reflection, critics and tool-using loops. Yet these claims are usually supported by end-to-end success rates, which show only whether a task was completed, or by LLM judges, which are not verifiers. We test LLM planning against formal semantics. Tasks are posed in natural language, but each is grounded in a hidden PDDL model that decides whether every action is applicable, whether the goal is reached, and whether a plan exists at all. Across five domains (two of them new, all released with the paper), four models, six prompting strategies and three agent architectures, the best configuration solves 78% of the solvable tasks offline but only 40% online, where agents mostly act without inspecting the state. Unsolvable tasks are recognised in under 10% of episodes, and heavier prompting strategies use 5 to 22 times as many tokens as direct prompting with no clear gain. In our experiments, planning performance varied more across models than across prompting strategies or agent architectures. Claims that agents can plan require evidence from validated action traces.
Generative Value Classifier
arXiv:2610.22497v1 Announce Type: new Abstract: Q-learning is premised on learning a function to predict the value of a state-action pair, with the goal of using such a function to guide policy improvement. Most approaches learn this function from scratch using scalar or distributional return targets. This discriminative approach leaves another source of supervision for learning representations unused: the actions themselves. Rather than directly predict the return of a state-action pair, we ask the reverse question: how likely is the action under each possible return? This perspective provides a critic whose generative component is a return-conditioned policy, which can share the actor's architecture. Training this policy supervises the critic with every action rather than with returns alone, and a squared-error loss anchors the critic's Q-value to the bootstrapped target. We evaluate our approach on both offline and online RL tasks and show gains over leading discriminative methods, including distributional RL methods.
RealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents
arXiv:2610.22507v1 Announce Type: new Abstract: Command-line interface (CLI) agents increasingly execute longer-horizon tasks, yet existing benchmarks either assume that user instructions remain fixed throughout execution or evaluate task outcomes without capturing runtime collaboration quality. In this paper, we present REALCLI, a paired benchmark that unifies the evaluation of task execution and human-agent collaboration quality on realistic CLI tasks. To comprehensively measure collaboration, REALCLI builds a five-dimensional framework, operationalized into 22 sub-dimensions and 51 interaction scenario templates. REALCLI provides two modes for paired comparison: REALCLI-EXEC supplies complete specifications upfront, whereas REALCLI-COLLAB introduces a scripted user that delivers progressive information, corrections, constraint revisions, and authorization decisions. Evaluations reveal profound capability gaps in frontier agents: the best configuration attains an average score of only 49.2%, and when evaluated by the joint collaboration score ($J_{\mathrm{collab}}$), the peak performance reaches merely 37.6%. These results show that task effectiveness diverges sharply from collaboration quality, with models consistently struggling in steerability, trust calibration, and safety boundary management. REALCLI provides a rigorous, reproducible foundation for measuring and advancing collaborative CLI agents.
AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems
arXiv:2610.22511v1 Announce Type: new Abstract: Evaluating language models and agents requires combining benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality. Differences in interfaces, metric semantics, and trace formats complicate these comparisons and can hide failures in agentic workflows. We present AuditKIT, a source-available evaluation library that records these measurements in one result representation, with direction-aware metric contracts, configuration fingerprints, and extensible metrics and backends. For agents, it provides turn-aware tool-call matching, separates parallel from sequential execution, and distinguishes verified task outcomes from trajectory-level diagnostics. Missing or unobservable evidence is marked unscored rather than counted as failure. For retrieval-augmented generation, it combines ranked retrieval metrics, judged grounding, and deterministic evidence checks. Seven case studies spanning model comparison, serving runtimes, retrieval, tool calling, and interoperability show how answer extraction, runtime behavior, and metric selection change evaluation outcomes.
Predicting Privileged Information Effectiveness for On-Policy Self-Distillation
arXiv:2610.22512v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) enables large language models (LLMs) to learn from privileged information (PI) without a separate teacher model. However, prior work largely treats PI as a fixed design choice, leaving the effects of its content and construction poorly understood. This challenge is particularly important in multi-turn agentic settings, where trajectories contain actions, observations, and reasoning that can be transformed into PI in many ways. We systematically evaluate 16 PI configurations across three multi-turn agentic benchmarks and three student models that vary in scale and reasoning mode. We find that PI effectiveness depends on its content, generation source, and the student model. Information-dense trajectories and PI produced by more capable generators generally yield larger gains, whereas thinking models benefit less consistently. To assess PI before training, we compare PI utility with two KL-based metrics: sampled-token reverse KL ($D$) and error-step KL contrast ($\Delta d^{-}_{\mathrm{err}}$). The KL-based metrics track post-training performance more consistently than PI utility, which can be distorted by information leakage. Empirically, selecting each task's PI by $\Delta d^{-}_{\mathrm{err}}$ also improves post-training performance over random selection, most when the PI candidates differ widely. Finally, effective PI remains beneficial when combined with GRPO, allowing the combined objective to outperform a strong GRPO baseline. Overall, our results provide practical guidance for constructing and evaluating PI in OPSD training.
alkhder at ImageEval 2026 Shared Tasks: Constrained Selection and Rotation-Averaged Logit Scoring for Multimodal Hallucination Detection
arXiv:2610.22515v1 Announce Type: new Abstract: We describe our submission to Subtask~1b of ImageEval 2026. The task consists of selecting the single visually valid statement from three candidates paired with an image, and is in both English and Modern Standard Arabic (MSA). Our system uses a pre-trained 7B vision--language model in a zero-shot setting, without task-specific training or fine-tuning. We formulate the task as a constrained selection problem and choose the answer using next-token log-probabilities. To reduce answer-position bias, we apply a Latin-square rotation to the candidate statements. Our system achieves CI scores of 0.0510 on the English test set and 0.0960 on the MSA test set, reducing CI relative to the official 3B baselines reported on devtest by 83.7\% and 80.4\%, respectively.
Source-Aware Dynamic Pricing with Strategic Buyers
arXiv:2610.22516v1 Announce Type: new Abstract: In modern commerce, firms and platforms typically use customer features from different sources to personalize prices: some features are verified using seller records or third-party data, while others are self-reported and may be strategically manipulated by buyers seeking lower prices. We therefore introduce a novel source-aware framework for contextual dynamic pricing with strategic buyers in a high-dimensional semiparametric setting, where buyer valuations follow a sparse linear model with an unknown noise distribution. To account for differences in manipulability across feature sources while enabling effective valuation learning, we develop the Source-Aware Localized exploration and calibrated Exploitation (SALE) pricing policy, which draws on verified features to localize price exploration and combines verified and reported features to correct for strategic manipulation during exploitation. Our localized exploration design allows converting binary purchase feedback into unbiased signals of conditional mean valuations, yielding a new population equation that supports consistent sparse estimation despite non-i.i.d. buyer arrivals. Building on this estimate, we develop a localized empirical-process analysis that exploits intrinsic sparse geometry to establish uniform guarantees for nonparametric estimation of the noise distribution in the high-dimensional setting. With these estimation guarantees, we establish SALE's self-recovery property, which keeps localized exploration valid with high probability and leads to a sublinear expected regret bound. Both simulation studies and real-data experiments demonstrate the appealing performance of SALE. To the best of our knowledge, our study is the first to provide a theoretical and practical framework for incorporating information provenance into dynamic pricing with strategic buyers.
Learning User Representations to Improve Simulation Fidelity
arXiv:2610.22518v1 Announce Type: new Abstract: User simulators enable interactive training and evaluation of AI assistants at scale. Yet current simulators struggle to reproduce the diverse behaviors of real users, leading to assistants that generalize poorly to underrepresented users. We identify user representations as an underexplored bottleneck in user simulation. User representations are natural-language descriptions that condition simulators to reproduce a particular user's behavior. A representation's quality becomes apparent in hindsight, through the behavior it induces in a simulator. To evaluate and optimize user representations, we formulate a conversation reconstruction task: given an original conversation, we generate a representation, condition a simulator on it, simulate a conversation, and then measure fidelity by comparing user behavior in the original and simulated conversations. This comparison also yields hindsight feedback, which identifies where the representation falls short and how it can be improved. We introduce hindsight distillation, an online training procedure that distills feedback-conditioned improvements into the representation generator, enabling it to generate improved representations directly from the original conversation. We present UserRepGen-4B, a representation generator trained with hindsight distillation. Across 700 conversations spanning seven datasets, hindsight feedback helps improve user representations at test time, and UserRepGen-4B generates representations that enable more faithful simulation than those generated by GPT-5.5. Beyond conversation reconstruction, learned representations improve simulation fidelity for unseen tasks, better preserve the population-level diversity of real users, and better distinguish users across conversations.
Learning smooth dynamics via Kinematic Interpolants
arXiv:2610.22519v1 Announce Type: new Abstract: Inferring dynamics from real-world experimental data is inherently ambiguous, as many plausible dynamical systems can explain the same positional observations at measurement times. For higher fidelity modeling of applications ranging from single-cell RNA to vortex dynamics and even flocking of birds, scientific observables increasingly also include velocity estimates that serve to resolve this issue. However, incorporating this rich information in a scalable manner, free of simulation or unnecessary approximations, has remained a difficult challenge. In this paper, we introduce Kinematic Interpolants (KI), a novel framework that extends the seminal methodology of stochastic interpolants to infer nonlinear dynamics across multiple snapshots in phase space, i.e., observables with position and velocity measurements. Empirically, we demonstrate that across several scientific datasets, KI achieves superior reconstruction of the underlying dynamics compared to existing baselines while requiring only a fraction of the training time.
alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B
arXiv:2610.22525v1 Announce Type: new Abstract: This paper describes team Alkhder's submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker-addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64 on both tracks.
Literal Lock-In: How Test Language Shifts Literal Attraction in Typed Decision Models on Bengali Idioms
arXiv:2610.22526v1 Announce Type: new Abstract: Decision models, non-generative scorers that report a probability over caller-supplied options and emit no text, are increasingly deployed as routers, classifiers and guardrails. Because they generate nothing, rationales and chain-of-thought do not apply. We audit three: an encoder scorer, a decoder scorer and a closed API, entirely through interventions on their input, using expert-annotated Bengali idioms carrying a gold figurative and a gold literal gloss. Dropping an idiom's own literal gloss into the option list draws every arm toward it, taking 65% to 96% of each arm's remaining Bengali errors. A model-by-language matrix over one shared item set shows how far this depends on the test language: every arm is 1.9 to 6.4 times more captured in Bengali than in English on the same paired entries, the commercial model's apparent resistance is a property of its test language, and capability does not order the failure. Calibration degrades in all three arms, and on the open arms confidence inverts, so gating on it is worse than gating at random. On a semantics-free option list, where accuracy is at chance, the best-calibrated arm still reports mean confidence 0.82. Code and data: https://github.com/sakhadib/literal-lock-in.
Pearl-Chat: Engineering a native Luganda language model from first principles in JAX and Flax NNX
arXiv:2610.22529v1 Announce Type: new Abstract: Autoregressive foundation language models predominantly target high-resource Indo-European languages, imposing computational and morphological penalties on under-represented African languages. Generic multilingual tokenizers routinely exceed fertility rates of 3.2 tokens per word on agglutinative Bantu languages, degrading context efficiency and inflating latency. We present Pearl-Chat, an 8,473,122-parameter ($\sim$8.5M) causal decoder-only transformer engineered for Luganda (Oluganda, ISO 639-2: lg), a low-resource Niger-Congo Bantu language spoken by over 10 million people in East Africa. Built from first mathematical principles in JAX and Flax NNX, Pearl-Chat uses explicit functional state decomposition (nnx.split and nnx.merge), deterministic resumable data streaming via Google Grain, and SPMD tensor parallelism using jax.sharding.Mesh. We introduce a byte-level byte-pair encoding (BPE) subword tokenizer trained from scratch on curated Luganda corpora ($V = 8,192$), achieving 4.88 characters per token compression and 1.60 tokens per word fertility, a two-fold compression improvement over LLaMA-3 (128K) and Gemma (256K). The model is trained with a two-stage curriculum: Stage 1 pre-training on 59,501 sequences across parallel, biblical, encyclopedic, and journalistic corpora, and Stage 2 instruction fine-tuning on 510 dialogues with assistant-only loss masking. Across 4,000 optimization steps, validation cross-entropy loss drops monotonically from 9.4695 (perplexity: 12,958.98) to 0.1818 (perplexity: 1.20), a decrease of 9.2877 nats. On a three-tier held-out diagnostic benchmark, Pearl-Chat achieves 0.9792 ROUGE-L and 0.9405 BLEU while maintaining concordial agreement across Bantu noun classes. We provide a Python Server-Sent Events (SSE) streaming runtime and open-source all weights, tokenizers, and code.
Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks
arXiv:2610.22538v1 Announce Type: new Abstract: Large reasoning models (LRMs) that leverage test-time computation achieve strong performance on mathematical and scientific benchmarks and are increasingly deployed in real-world applications. However, their ability to perform \textit{everyday logical reasoning}, which often departs from textbook formal logic, remains underexplored. In this work, we systematically evaluate how well reasoning models align with everyday reasoning patterns by introducing a benchmark based on linguistic and pragmatic reasoning phenomena. We also compare reasoning models with non-reasoning models under controlled prompting conditions. We find that, in many cases, (1) reasoning models show degraded performance in everyday reasoning when such reasoning deviates from formal logical rules, (2) some non-reasoning models with chain-of-thought prompting outperform comparable reasoning models in everyday reasoning, and (3) prompting with additional instructions provides limited control over reasoning behavior. These findings suggest that current reasoning models do not reliably capture everyday reasoning and point to limitations in post-training, motivating improved training objectives and evaluation across diverse forms of reasoning.
Don't Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks
arXiv:2610.22539v1 Announce Type: new Abstract: The web is not static or uniform. Websites that serve the same purpose can look and navigate very differently. Even within a single site, successive redesigns can change how familiar features are presented and accessed. Humans adapt to this by recognizing familiar interaction patterns and reusing what they already know. We give web agents a similar ability with WebMix, a mixture hypernetwork that adapts them to unseen interfaces by composing previously trained low-rank adapters. WebMix begins with a bank of adapters trained on source interfaces. At deployment, a scout agent explores the target interface and summarizes its observations into a compact manual. A lightweight router uses this manual to predict mixing coefficients over the adapter bank, which combine the existing adapters into a new one for inference. Adapting to a new interface takes only a scouting run and a forward pass, without gradient updates on the target. With Qwen and Llama models, we evaluate domain transfer on TimeWarp and WebArena-Lite-v2, and version transfer on TimeWarp. In all three settings, WebMix matches or outperforms in-context learning and LoRA baselines, often matching the average success of specialists trained directly on those interfaces. On TimeWarp domain transfer, it gains about 8 percentage points on average over the best baseline with both models. These gains do not come at the cost of seen interfaces, where WebMix keeps its lead, and extend to version transfer. Adapting to a new interface can therefore rely on exploration and composition instead of retraining.