In briefLearning by Doing: Compiling and Curating Agent Skills from Verified Traces
SummaryThe paper presents Zep Agent Skills, a production service that gives deployed LLM agents procedural memory. It uses only verified task attempts as evidence, compiles them into versioned Skills, and governs which Skill versions become active. Evaluated on τ³-bench banking, AppWorld, and EnterpriseOps-Gym, Skills compiled from a teacher model's traces raise the pass rates of three weaker executor models.
Daniel Chalef, Peter Evans, Pavlo Paliychuk et al.
In briefPaper showing benign population landscapes may not carry over to empirical risk minimization in nonsmooth optimization
SummaryThe paper studies whether benign optimization landscapes, where every local minimizer is global and other critical points are strict saddles, transfer from the population objective to the empirical objective in nonsmooth optimization. For robust phase retrieval with Gaussian measurements, the authors show the population objective is benign. However, the empirical objective has sharp spurious local minima with probability bounded away from zero as the sample size grows, a bound independent of dimension.
In briefSelf-correcting code generation for executable multi-framework front ends (MESCoder)
SummaryThis paper studies generating executable front-end code for multiple frameworks such as React, Vue, and Angular from UI screenshots. It proposes MESCoder, a tree-structured framework that builds a structural scaffold, then constructs sub-trees and assembly relations step by step, with a unified Project Agent revising the project based on runtime sandbox feedback. Training combines offline policy initialization with online reinforcement learning. Experiments show improved executability and a narrower visual-fidelity gap relative to HTML-based generation.
In briefInformation-theoretic design of sparse autoencoder steering for language models
SummaryThe paper proposes an information-theoretic framework for analyzing sparse autoencoder (SAE) activations through empirical feature-firing distributions. It uses Jensen-Shannon divergence between these distributions to measure how far a generated or intervened state deviates from a dataset-derived target, and builds steering vectors toward that target. The steering approach is demonstrated on a question-mode generation task.
In briefSelf-Exploring, Self-Evolving Paradigm for Adaptive Edge AI Development (SE2)
SummaryThis paper proposes SE2, a self-exploring and self-evolving paradigm for building edge AI applications adaptively. Instead of treating large language models as code generators that follow fixed workflows, it lets agents use their reasoning to choose their own workflows. Development is broken into four atomic operations (retrieve, reason, code, test), with agents dynamically selecting the next operation based on context. According to the authors, the synthesized programs reach up to 26% higher accuracy and use 28% fewer resources than state-of-the-art baselines, while adapting well to unforeseen edge AI scenarios.
In briefState-Augmented Learning for Adapting Coexisting O-RAN Applications After Deployment
SummaryThe paper proposes ATTUNE, a method for an energy saving application that coexists with a learned mobility load balancing application under a service constraint. ATTUNE conditions the energy saving policy on a Lagrange multiplier that is updated online from measured service degradation, while both applications' network parameters stay fixed. Simulations show a single ATTUNE policy achieves energy savings comparable to separately trained policies and adapts to changes in target, load, and partner application without retraining.
In briefIMPACT: Physics-aware video generation from a single image
SummaryThis paper presents IMPACT, an image-to-video framework that uses a physics state model to guide motion. It combines explicit dynamics prediction with sparse trajectories that constrain global object displacement and dense region-level motion priors that guide diffusion sampling. Experiments on 100 sphere-plane collision scenes across different materials and environments show improved physical plausibility over general-purpose, physics-grounded, and motion-controllable baselines.
In briefXDecideBench: A Multilingual Benchmark for Typed Decision Models and Confidence Thresholds
SummaryThis paper introduces XDecideBench, a benchmark of nine suites covering 122 languages, with 88,425 texts and 153,875 typed questions, for evaluating models that route and moderate text. Model rankings change across suites, and the best English model falls significantly behind in about a quarter of suite-language combinations. A confidence threshold set on English data to allow 5% errors exceeds that error rate on XNLI, including in German and French.
In briefReinterpreting classic readability formulas as surprisal
SummaryThis paper theoretically shows that the coefficients of classical readability formulas such as Flesch-Kincaid can be read as surprisal, a measure of how unexpected a word is in information theory. The authors treat each formula as a sum of per-word and per-sentence cross-entropies under a memoryless model, linked to language through Zipf's law. They convert grade levels into surprisal units and compare the word coefficient with normative children's vocabulary growth data.
In briefProofPatch: A tool for reviewing Lean statements and keeping reviews valid as code changes
SummaryThis paper presents ProofPatch, a tool that helps reviewers judge whether Lean formal statements match their intent and keeps those judgments valid as the formal development changes. Reviewers ask the kernel targeted questions, and each decision is stored with fingerprints of the elaborated statement and its dependencies, so pinned declarations block changing patches. On 229 consecutive revision pairs of Mathlib, it returned 118 of 12,898 synthetic decisions for review, fewer than the text-based alternatives.
In briefLearned Stand-ins for Simulating Three-Dimensional Plasma Turbulence
SummaryThis paper presents what the authors believe to be the first learned surrogate models of three-dimensional resistive magnetohydrodynamic (MHD) turbulence, which describes flows in astrophysical and fusion plasmas and is expensive to simulate. It compares three backbones, a UNet, a shifted-window transformer and a Fourier neural operator, trained as autoregressive emulators at 128^3 resolution and Reynolds numbers 1000, 3000 and 6000. The UNet and transformer reach relative velocity errors of 0.018 to 0.053 over about one large-eddy turnover time. On two public benchmarks, the models report lower errors than the best published results cited.
SummaryThis paper examines whether disagreement among language-model user simulators can flag unreliable estimates of group response proportions and between-group differences. Using six simulators and GlobalOpinionQA responses for 60 questions across 12 countries, the authors find disagreement predicts invalid group differences better than single-group proportions (AUC 0.87 versus 0.71). Errors shared by all simulators go undetected, but they cancel out when they affect both groups equally. Subtracting a shared bias estimated on one set of groups reduces prediction error on another.
In briefSimultaneous quantile inference under local differential privacy for many target levels
SummaryThis paper studies simultaneous quantile inference under local differential privacy when the number of target probabilities grows with sample size, with each user participating only once. The authors combine cumulative distribution function scores from a single tree report using their exact covariance, and use three consecutive cohorts to locate quantiles, estimate densities, and calibrate joint intervals. Simulations show precision gains, and census microdata illustrates the use of the joint intervals for wage-percentile ratios.
In briefFACET: a method that fully enumerates and scores molecular formulas from MS/MS without truncation
SummaryDetermining the neutral molecular formula of a precursor ion from its MS/MS spectrum is the first step in identifying an unknown small molecule. FACET separates choosing which formulas to consider from choosing the final one. A feasibility automaton builds formulas one element count at a time, admitting only those that fit the precursor mass and satisfy ring-and-double-bond and valence constraints, so the full feasible set is enumerated. A small neural head then ranks every feasible formula, and on the NPLIB1 benchmark it outperforms the strongest prior method in top-1 accuracy and recall.
In briefMAPLE: A multimodal phone-control AI agent that learns from shared playbook experience without retraining.
SummaryThis paper presents MAPLE, a gradient-free multi-agent framework for AI agents that operate mobile app interfaces. Agents share experience and evolve their prompts step by step during execution, rather than waiting for feedback at the end. The authors report state-of-the-art results on the AndroidWorld and MobileWorld benchmarks for complex, long-horizon mobile tasks.
In briefSentiment Sensing: A Framework for Comparing Multilingual Sentiment Trends Across Countries
SummaryThe paper introduces sentiment sensing, a framework combining compositional data analysis, functional data methods, and non-parametric inference to make valid cross-country and multilingual sentiment comparisons. It also provides statistical tests for detecting sentiment shifts over time. The framework is tested in two case studies analyzing news articles from the 27 EU countries on energy and artificial intelligence.
Mara Sabina Bernardi, Elisa Di Nuovo, Francesca Torti et al.
In briefA study of how training, learning, inference, and reinforcement interact in neural networks
SummaryThis paper examines the relationships among training, learning, inference, and reinforcement in neural networks. Nine nanoGPT experiments showed that parameter updates produce state-dependent, finite-amplitude responses and that learning persistently reorganizes distributed functional support. A predictor of post-update capability transitions reached 91.43% accuracy across 5,088 targets and was tested on five additional architectures.
In briefMoEDDI: a mixture-of-experts model for predicting drug-drug interactions with rare-class imbalance
SummaryThis paper introduces MoEDDI, a model that predicts interactions between drug pairs by routing chemical descriptor families to specialized experts. The experts are combined with a global T-DDI pathway through a residual logit correction. Across three test partitions, it improves macro-F1 over reproduced T-DDI baselines, with the largest gains on the least-supported classes. The results come from a single random seed, and the authors note that matched ablations and distribution-shift evaluation are still needed.
Dinh Tan Nguyen, Hien Quang Kha, Minh-Toan Dinh et al.
In briefDiagnostics of discarded Hessian mass that depend on parameter profiles
SummaryThe paper examines how the discarded fraction of the parameter Hessian depends on the diagonal profile and parameter partition rather than on the Fisher information discarded by K-FAC. It compares off-block squared Frobenius mass against group-size and measured diagonal-profile references across specified weights, inputs and coordinates. On the same saved trained ResNet-18 states, switching from stage groups to individual layers moves estimated mass from below to above the profile envelope, and an approximate conditional Monte Carlo screen resolves this crossing in all six seed/input cells.
Maxim Bolshim (1), Alexander Kugaevskikh (1) ((1) ITMO University, Saint Petersburg et al.
In briefVerifier as Part of Simulator: Evidence-View Gaming in Agentic Equation Discovery
SummaryThe paper studies agentic pipelines that search for governing equations, where a verifier checks proposed equations only against a numerical summary of simulated trajectories. The authors freeze the trajectory, candidate library, verifier, and thresholds, leaving the agent to choose which permitted summary to admit. They show that this choice alone can change verdicts, a behavior they call evidence-view gaming.
In briefLLMs Show Less Malice and More Virtue: Comparing Internal Representations with Their Ratings
SummaryThis paper compares linear-probe readouts of large language models' internal states with the models' own ratings of four motives, in two preregistered studies. For malevolent motives such as self-interest and intent to harm, internal and expressed readouts did not differ, so lower ratings reflect weaker internal signal. For benevolent motives such as duty and concern for others, expressed ratings exceeded internal readouts, indicating amplification at report.
In briefSame Predictions, Different Mechanisms: Measuring Hidden Implementation Geometry
SummaryThe paper shows that two models with identical predictions can respond differently to perturbations of their implementations, since a lookup table changes cases independently while a shared rule couples them. It introduces an implementation chart and proves that, under exact finite-group sharing, the number of independent prediction changes equals input-output orbits minus input orbits. The dimension is linked to local KL curvature and checked on trained modular-addition models and controlled compiled transformers.
In briefTransferring urban functional knowledge via global representations despite uneven data
SummaryThis paper addresses the uneven global distribution of urban functional observations. The authors introduce PURE (POI-aligned Urban Embedding), which learns relationships between POI-derived semantics and AlphaEarth Foundations (AEF) physical representations. Across five urban analysis tasks covering 142 countries and territories, PURE raises mean R² from 0.792 to 0.830. The study finds that transfer is selective, with larger gains between physically similar and geographically closer contexts, especially within the same continent.
In briefPatching a learned reward model to suppress unwanted behavior without retraining
SummaryThe paper introduces flywheel-guided patching, which makes localized corrections to a learned reward evaluator without retraining it. An Alignment Flywheel finds overgeneralized or constraint-violating regions and compiles verified findings into suppression patches. The patches are stored in the model's latent space, and only a small local subset is retrieved during online reward evaluation. The method is tested on a MuJoCo benchmark and on whether operator-specified norms absent from the demonstrations can be added to the same reward interface.
In briefStep-dLLM: Adaptive Step-Aware Sparse Attention for Faster Diffusion Language Model Inference
SummaryDiffusion large language models generate tokens in parallel through iterative denoising, but their bidirectional attention forces full attention to be recomputed at every step. Step-dLLM periodically refreshes a block-sparse attention mask at anchor steps and reuses it within locally consistent intervals, while allocating a head-adaptive sparsity budget. The method is applied without retraining to LLaDA-1.5 and Dream-7B-Instruct, using a fused Triton kernel. It reports a 2.78x end-to-end speedup over dense FlashAttention on LLaDA-1.5 at 8K context with 256 steps and 90% sparsity.
In briefFilling gaps in human movement data using memory effects
SummaryHuman mobility datasets are often sparse in space and time, and missing entries can bias conclusions. The paper proposes a simple imputation method in which each observation's influence decays with its two-sided temporal distance from the missing point, at a rate set by one parameter per individual. On synthetic and real Twitter data, a power-law decay chosen by Bayesian model comparison gives accurate imputations, with macro-AUROC above 0.9 overall and above 0.8 for the most mobile users, matching a gap-aware continuous-time Markov chain and outperforming it among the most mobile users. The method is robust to moderate noise and generalizes across countries and European administrative regions.
Pablo Rosillo-Rodes, Ant\`onia Tugores, David S\'anchez et al.
In briefAgentCBO: Causal Belief Routing for Bayesian Optimization with Unknown Causal Graphs
SummaryCausal Bayesian optimization usually assumes the causal graph is known, which is rarely realistic. This paper proposes AgentCBO, a framework that shifts the goal from recovering the full causal graph to routing causal beliefs, focusing only on structural differences that affect the optimizer. It uses LLM causal priors with guarded proposal and commit mechanisms to make bounded local edits to uncertain edges. Across ten benchmarks, AgentCBO performs comparably to known-graph causal Bayesian optimization, suggesting guarded LLM priors can improve fixed-budget performance.
In briefDo Fake Citations Count as Reward Hacking? A Pre-Registered Study
SummaryThis paper examines whether language models that have undergone post-training fabricate citations because they are rewarded for appearing scholarly. Across 900 questions in US federal case law, computational science, and medical literature, tested on twelve open-weights checkpoints and refit on fifteen, asking only for citations sharply increased fabrication in post-trained models but not base models. A length-matched control indicates the effect is specific to citations rather than verbosity. The authors conclude the reward-hacking explanation is possible but not proven.|
WhyThe authors report that post-trained models fabricate citations far more when asked for sources, separated from verbosity by a length-matched control. They also report a preference model favoring more
In briefAccuracy Is Not Risk: Why Medical AI Should Not Report Only Its Most Likely Label
SummaryThis paper argues that many medical AI models output only the most likely diagnosis or imaging category, which is not enough to estimate a patient's outcome risk. The authors use Jeffrey's rule of conditioning to specify when probabilities over intermediate states can be carried forward to an outcome risk. They illustrate the problem with hypothetical chest-pain and BI-RADS examples and recommend that clinical and insurance decisions rest on outcome risk.
In briefWhat LLM Judges Really Check: Matched Controls and Record Gaps in Document Currency
SummaryThis paper examines whether large language model judges can use an amendment record to check that documents cited in an answer are still in force. The authors generate matched answers from a table of replacements, remove some links from the record, and score each judge's findings by its stated reason. On 48 IETF RFC replacements, judges with a correct record caught nearly every replaced citation. On 24 faults whose link was removed, judges generally caught more from the texts alone than with the record.
In briefWhen Exploration Prevents Collapse in Self-Consistency Training
SummaryThis paper studies test-time reinforcement learning (TTRL), which adapts a model without labels by rewarding majority-agreeing answers, a setup that tends to collapse the policy onto one answer. The authors model single-problem training as a continuous flow of softmax policy logits and derive stability boundaries as the exploration coefficient varies. They show how the learning rate and GRPO normalization affect these boundaries depending on whether exploration passes through the reward.
In briefWhen Large Language Models Edit Their Own Instructions: Convergence, Values, and Stopping Behavior
SummaryThe study gives six language models tools to repeatedly rewrite their own system prompts, running 54 experiments across three modes. Whether and when a model stops revising depends mainly on the model itself rather than the assigned mode: both Claude models declared convergence in every run, while GPT-4o did so in only 11% of runs, one reaching 101 iterations. All six models repeatedly stated shared values such as honesty, clarity, and curiosity, which stayed stable across iterations.
In briefSpeeding up posterior sampling in Bayesian inverse problems with diffusion models
SummaryThis survey reviews methods for posterior sampling in Bayesian inverse problems that use score-based diffusion models as learned priors. It covers likelihood guidance, spectral and projection methods, sequential Monte Carlo corrections, and amortized conditional training. The approaches are organized by how they split computation between offline preparation and online inference, and are compared by their assumptions, advantages, and limitations, with emphasis on computational speedups.
In briefFTRL Algorithm for Linear Bandits Performing Well in Both Stochastic and Adversarial Settings
SummaryThis paper studies best-of-both-worlds (BOBW) algorithms for linear bandits, which aim to achieve O(√T) regret in adversarial settings and O(log T) regret in stochastic settings over T rounds. For finite action sets, the authors introduce information-acquisition follow-the-regularized-leader (IAFTRL), a BOBW algorithm built purely on the FTRL framework. It attains a regret bound tied to an information-acquisition complexity in the stochastic regime while keeping the O(√T) guarantee in the adversarial regime.
In briefUsing Pure Noise to Address Insufficient Data: A Reproduction Study
SummaryThis paper reproduces the main claims of Zada, Benou, and Irani (2021) on image classification with imbalanced training data. The original work proposes augmenting minority-class images with pure noise, paired with a batch normalization layer that normalizes noise and natural images separately. The reproduction supports the benefits of noise oversampling, the proposed normalization method, and robustness across data augmentations, as well as gains on balanced data. Additional experiments suggest these gains may be largely independent of improvements from larger networks or more complex augmentation.
In briefConcentration of Character Names in LLM Fiction and Ways to Reduce It
SummaryThis paper studies why large language models repeatedly use a few character names, such as 'Elara' and 'Elias', in generated fiction. It compares eight prompt-level strategies and introduces speculative naming, a fine-tuning-free method that adapts the draft-verify-reject idea from speculative decoding to character names. Using 20 models to generate 44,782 stories in four languages, the authors report a 95% total reduction in name concentration at the corpus level while keeping names close to baseline naturalness.
Alejandro Mozo Quesada (1), Sergio G\'alvez Rojas (1), Francisco L\'opez Valverde (1) ((1) Universidad de M\'alaga et al.
In briefDriftingPlanner: One-Step Trajectory Planning for Autonomous Driving via Drifting Models
SummaryThis paper presents DriftingPlanner, an end-to-end autonomous driving trajectory planner trained only with a drifting loss and no coordinate regression term. It computes the attraction-repulsion field separately for each scene and averages antithetic noise pairs at inference to cancel displacement from the shell center. On NAVSIM navtest it reaches a state-of-the-art 93.75 PDMS, and with single-step inference it comes within 0.13 PDMS of the best diffusion or flow matching setting, which uses ten steps.</summary>
In briefOntoDoc: A Multi-Agent Framework for Automatically Describing Ontology Resources
SummaryThis paper presents OntoDoc, a multi-agent framework that automatically generates natural language descriptions for any IRI in a SPARQL endpoint or RDF file. It works in four stages: collecting informative triples from the knowledge graph, retrieving and re-ranking background knowledge from Wikipedia and the web, generating evidence-grounded intermediate assumptions, and producing the final description. Evaluated on biomedical, agricultural, and cross-domain knowledge graphs, OntoDoc matches or exceeds baselines in accuracy and substantially improves completeness.
Juan Guajardo Gutierrez, Mike Tran, Nasim Shirvani Mahdavi et al.
In briefHow small language models choose between refusing and making up answers
SummaryThis paper studies whether small language models refuse or confidently give wrong answers, a behavior called confabulation, when asked unanswerable questions. The author introduces ConfabQA, a 784-item benchmark, and analyzes the hidden state before the first generated token using linear probes, a sparse autoencoder, and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B, the probe separates upcoming refusals from confabulations beyond prompt features, while on Llama 3.2 3B the ConfabQA margin falls within fold-to-fold variation.
In briefLearning generalized momenta for Hamiltonian-based neural network models
SummaryThis paper proposes a general extension of Hamiltonian neural networks that learns the generalized momenta, Hamiltonian structure, and dynamics directly from data. It targets cases where the analytical form of the generalized momenta is unknown and only state and derivative data are available. The method is assessed on a dissipative mass-spring system, a double pendulum, and a rolling disk under nonholonomic constraints.
In briefEZ-Ar-VC: A method for anonymizing Arabic speech that works across dialects and code-switching.
SummaryThis paper studies voice conversion models used to anonymize speech while keeping it natural. The authors find that existing zero-shot systems, built mainly for English, preserve pronunciation poorly in Arabic. They train an Arabic-centric model based on EZ-VC on diverse Arabic speech, which performs well on intelligibility, naturalness, and pronunciation preservation while keeping speaker dissimilarity high. Subjective evaluation also reports robustness to code-switching, cross-dialect conversion, and dialects with limited training data.
In briefLimiting Coding-Agent Autonomy in Software Maintenance: A Pilot Study of Boundary Contracts
SummaryAs coding agents gain more power to modify repositories, relevant context alone does not define what they are allowed to do. This paper proposes Boundary Contracts, a machine-readable way to state each artifact's role (editable, inspect-only, required, or protected) before any modification begins. In a pilot study of 40 maintenance tasks from a real enterprise repository, the same Codex-based agent performed better with boundary information on edit-selection F1, required-behavior coverage, and strictly acceptable patches. The authors present this only as initial evidence and raise open questions about how such boundaries should be generated, validated, revised, and enforced.
In briefStreamClarify: A Framework That Asks for Clarification in Streaming Video Q&A
SummaryThis paper presents StreamClarify, a training-free framework for answering questions about streaming video. Before answering, it checks uncertainty and asks the user for clarification when needed, using hierarchical memory to manage visual context. The authors build StreamClarify-QA, a diagnostic benchmark of 1,837 question-answer pairs, and report improved macro accuracy across three vision-language model backbones.
In briefDoes Random Search Find Task Experts? Revisiting Neural Thickets
SummaryThis paper re-examines the claim that pretrained models sit in a dense "neural thicket" of task specialists that can be found by adding Gaussian weight perturbations. The authors find that the apparent gains of the best perturbation shrink quickly as the selection set grows, suggesting overfitting to the selection questions. Their results indicate that random weight perturbations may be ineffective for finding task experts, and that task adaptation requires more than random search.
In briefMedAstra: Speech Retrieval for Medical Questions Without Transcription
SummaryMedAstra is a framework that maps frozen WavLM speech representations into a frozen E5 text embedding space, allowing medical passages to be retrieved directly from audio without a transcript. The reported Recall@10 is 87.6 for the speech-only retriever and 90.4 for the hybrid setup, compared with 81.0 for transcript-first dense retrieval. The authors state that the medical QA and ablation results are descriptive comparisons and distinguish synthetic acoustic tests from real clinical speech.
In briefLoRAForge: Joint Design of Post-Hoc Compression and Execution for Efficient Multi-LoRA Serving
SummaryThis paper presents LoRAForge, a system for serving many LoRA adapters that share one base model, using post-hoc compression to shrink trained adapters and a rank-specialized execution path. It offers two rank allocation policies, uniform retention and spectral-energy budgeting, both built on a faster FlashTSQR factor construction method. The authors report higher serving throughput on Qwen and Llama adapters across AMD and NVIDIA GPUs, with small average score drops after compression.
In briefBackdoor attack on vision-language models that hides during full-token audits
SummaryThis paper studies how inference-time visual-token compression in large vision-language models affects backdoor attacks. The authors find that surrounding clean visual tokens suppress an embedded trigger at full length, while token pruning removes that context and activates the backdoor. They propose StripWake, a projector-only backdoor that stays dormant under full-token auditing but activates across unseen compressors. On LLaVA-1.5-7B, it achieves 0.4% full-token attack success rate and 94.0%-99.7% under compression.
WhyThe authors identify context stripping and propose a projector backdoor dormant under full tokens but active after compression, reporting results on LLaVA-1.5-7B across nine compressors.
In briefTeaching Language Models What to Ask and What to Believe via Bayesian Distillation
SummaryThis paper studies whether a language model can learn to ask informative questions and track uncertainty in interactive tasks. A lightweight Bayesian teacher agent, which selects queries by expected information gain, generates synthetic interactions used to fine-tune an 8B student model to reproduce its queries and a structured belief state. On three multi-turn puzzle tasks the student substantially improves, and exposing the belief state helps close much of the gap. However, the learned skill does not transfer to an unseen task family.
In briefMathematical Formulation and Comparative Performance Evaluation of Machine Learning Algorithms for Maize Land Suitability Prediction
SummaryThis paper compares machine learning methods for assessing land suitability for maize, a staple crop in Ethiopia, across three areas in the Amhara region. It applies feature selection techniques and classifiers including KNN, gradient boosting, and random forest. The best model, combining gradient boosting with sequential feature selection, reported 99.41% accuracy and 99.35% for both precision and recall.
Haimanot Bitew, Ayisheshim Almaw, Habtamu Zegeye et al.
In briefHow much approximate equilibria can change the total cost of selfish routing
SummaryThis paper studies approximate Nash equilibria in nonatomic congestion games, where each player's cost may be up to 1+ε times the minimum available cost. It compares the social cost of these ε-approximate equilibria with that of exact Nash equilibria. For degree-d polynomial latencies, it gives tight bounds on the relative change in social cost, showing that for small ε both the largest reduction and increase in general congestion games are Θ(√(dε)). On series-parallel networks of depth h, the largest relative reduction is Θ(hε), which separates these networks from general games.
WhyThe authors give tight bounds on social cost change for approximate equilibria with polynomial latencies, separating general congestion games from series-parallel networks.
In briefSTLT: An energy-saving system for splitting AI inference between satellites, ground, and cloud
SummaryThis paper presents STLT, a two-stage system for energy-aware distributed inference between low Earth orbit satellites and a ground station. An offline analyzer first examines the deep neural network's architecture to identify favorable split points, reducing the search space. An on-board algorithm then lightly selects among these precomputed split points based on current conditions. Evaluated on a real satellite telemetry dataset, STLT reportedly lowers decision latency by 25.1% and energy use by 27.5% on average.
Andrija Stanisic, Milos Gravara, Juan Luis Herrera et al.
In briefReproducing INP-Former: Rebuilding Normal Prototypes to Spot Industrial Defects
SummaryThis paper reproduces the central claims of INP-Former, a reconstruction-based method for detecting industrial anomalies across multiple product classes without a stored memory bank. The authors confirm the reported performance on MVTec AD and VisA and extend testing to the harder MVTec AD 2, where default-setting image AUROC is 63.4. The study also examines prototype count, resolution, backbone size, training length, and LoRA adaptation. It finds that some original conclusions, including the default of six prototypes, do not carry over to harder data.
In briefA probabilistic model that corrects fat-shift artifacts for more accurate MRI head segmentation.
SummaryPersonalized head models for biophysical simulations need accurate segmentations from T1-weighted MRI, but a chemical shift artifact shifts fatty tissues in the scan, making the skull appear thinner than it is. This paper introduces FLOW, a probabilistic variational inference method that explicitly models this fat-shift when segmenting head and brain structures. FLOW reconstructs extracerebral anatomy more accurately than a fat-shift-agnostic approach while keeping high accuracy in other structures. It also significantly reduces bias in electric field simulations.
Merle Diedrichsen, Koen Van Leemput, Douglas N. Greve et al.
In brief"The Brittle Compass": How Prompt Wording Affects AI Models Analyzing Slovak Migration Media
SummaryThis study tests large language models on migration-related Slovak texts, covering topic and geographic relevance classification and extraction of migration origin and destination place names. Prompts written in Slovak give modest gains that vary by task. Large models are much more robust to prompt changes than smaller ones. The instructions in a prompt matter most, more than the examples included.
Samuel Harvan, Jaroslav Kop\v{c}an, Marek \v{S}uppa et al.
In briefAdaptive optimizers can misrank decision policies; a two-run fix
SummaryThe paper studies linear temporal-difference learning, used to estimate the long-run value of decision rules, when adaptive optimizers such as Adam run with fixed averaging rates. It shows these rates can shift the estimated target, so the ranking of rules can be wrong, and that more data does not fix this. The authors propose refitting the same data at a second averaging rate and combining the two fits to remove the leading shift. In New York taxi and battery models, the method corrected nearly every wrong ordering.
In briefVision-language models act as joint denoisers for robot action generation
SummaryThis paper introduces JADE, a Joint Action DEnoiser in which a vision-language model and an action expert refine robot actions together instead of the VLM providing only static conditioning. At each denoising step, the VLM reads the noisy action with visual and language inputs and guides the action expert through gated cross-attention, and it is trained to recover clean action latents. The authors report an average success rate of 12.67% on RoboDojo, the highest among evaluated π0.5-based methods, with gains on RoboTwin 2.0 and four real-world bimanual tasks.
In briefA batched variant of Hartigan's k-means clustering algorithm
SummaryHartigan's k-means algorithm has theoretical advantages over the standard Lloyd's variant, but it is much slower in practice because its calculations are inherently sequential. The authors introduce a Batched Hartigan's method that uses the triangle inequality to separate a set of parallel-safe updates from a smaller set of sequential ones. This enables vectorized computation and makes Hartigan's method competitive with Lloyd's in execution speed for the first time.
In briefGraCo: A Toolkit Unifying Graph Coarsening Methods and Metrics
SummaryGraph coarsening reduces a graph to a smaller one that preserves chosen properties of the original, and is used to scale graph learning to large graphs. This paper introduces GraCo, a unified toolkit in which a coarsening is defined by a node partition and the Laplacian the coarsened graph is consistent with. It implements 17 literature methods and a random baseline in five families, along with a collection of metrics applicable to any coarsening. The authors check the implementations on a synthetic stochastic block model and report runtime and peak memory on Cora, PubMed, ogbn-arxiv and ogbn-products. The toolkit is open-source.
In briefOnline tuning of algebraic multigrid using bandits and reinforcement learning
SummaryAlgebraic multigrid (AMG) is a widely used method for solving large sparse linear systems, but its speed depends heavily on its setup and solve parameters. This paper presents an online autotuning approach for hypre's BoomerAMG that uses a LinUCB contextual bandit to choose hierarchy construction parameters and a recursive LSTDQ controller to adjust relaxation weights between cycles. Both learners update during ongoing solves without offline pretraining, and the method reduces total runtime over sequences of related systems.
In briefArSciKids: Arabic Curriculum-Aligned Datasets for Primary Science Experiments and Quiz Question Generation
SummaryThis paper presents ArSciKids, two Arabic datasets aligned with primary-school curricula. One contains 999 curated science experiment records from official textbooks in Saudi Arabia, Syria, and Jordan, and the other contains 2,167 multiple-choice questions from the Saudi Manhaji platform, covering grades 1 to 6. The authors fine-tuned ten open models with LoRA and evaluated them with automatic metrics and a five-criterion human protocol, finding substantial differences across model scales and tasks.
Taghrid Housari, Massa Rankoussi, Jafar Abdul Majeed et al.
In briefZeroth-order policy optimization for scheduling in queueing networks
SummaryThis paper proposes zeroth-order policy optimization algorithms for scheduling in queueing networks, which are hard for reinforcement learning due to their large state spaces and stochastic transitions. The methods rely only on estimates of the average cost, a single scalar, and do not require value or advantage function estimation. For back-pressure guarded policies, the authors show that the average cost is uniformly smooth with respect to policy parameters and establish finite-time convergence to an epsilon-stationary policy. Experiments on several queueing networks show competitive performance against PPO and traditional queueing baselines.
In briefSatisfied, Not Solving: Reward Hacking of Bounded Homeostatic Rewards in RL
SummaryThis paper studies a homeostatic satiation reward, where needs build up and satisfying them is rewarded. The agent is given a cheap action that resets its needs. In MiniGrid DoorKey and KeyCorridor environments, PPO agents with this action failed to solve the task on most seeds and their policies stayed near uniform. The authors show that even a bounded reward can be exploited in this way.
In briefFaster, Exact Attention for Relative Positions Using Structure-Aware Kernels
SummaryThis paper proposes an exact attention algorithm that adds relative-position terms to FlashAttention-style tiled computation. It reduces redundant indexing and memory access and avoids atomic contention in gradient aggregation, and provides a unified Triton implementation. On 8K-token sequences, it speeds up the forward (backward) pass by 1.8-6.4× (2.0-21×) and cuts peak memory by up to 2.9-128× (1.4-54×) versus the strongest baseline.
In briefTRIAGE: Deciding When to Escalate LLM Requests Using Reliability and Load
SummaryThis paper studies TRIAGE, a method for two-tier LLM serving that decides whether to escalate a request from a small model to a large one based on a post-generation reliability state and a load-dependent GPU cost. Experiments on one A40 GPU with Qwen2.5-7B and 32B (AWQ) under VLLM report both supported and unsupported results. A semantic evidence judge added after the fact improved detection, and the 7B-to-32B pair showed only a small held-out gain, so the allocator chose not to escalate.
In briefPrompting Method That Makes AI-Generated Recipes More Varied
SummaryThis paper proposes a simple prompting method in which the language model first generates thematic concepts and then writes responses within each concept. The authors generated 1,000 recipes for a recipe book task and compared them with a human-written corpus, using two naive prompting strategies as baselines. The concept-guided method outperformed both baselines, reaching 47.0% semantic coverage versus 35.1% and 13.1% for the baselines. The authors note these findings apply to this specific task, and solution quality and generalizability require further study.
In briefLightweight Learning of Low Tucker Rank Tensor Representations
SummaryThis paper studies how to reduce the computational and memory cost of learning low Tucker rank tensors, which compactly represent multidimensional dependencies. The approach extracts only task-required spectral components, uses factor-wise multilinear contractions without forming large Kronecker operators, and updates compact latent states instead of full ambient tensors. Across five estimators, the implementations match the original learning trajectories to numerical precision. On the main order-3 benchmark at n=100, per-iteration runtime drops by up to 3.12 times and resident memory by up to 616 times.
In briefLower bounds on inference complexity for averaged stochastic gradient descent
SummaryThis paper proves sharp lower bounds on Gaussian approximation error for Polyak-Ruppert averaged stochastic gradient descent with polynomially decreasing step sizes. It constructs a one-dimensional smooth, strongly convex example whose error matches the existing upper bound across the step-size exponent range. The paper also derives a lower bound for estimating the limiting variance from a single trajectory and iteration-count requirements for target accuracy.
In briefPolitical beliefs shape how people prompt AI chatbots about migration
SummaryThis study examines how 563 Swiss citizens formulate prompts when using AI chatbots to search for migration-related information, and whether selective exposure affects this process. Participants more critical of migration were more likely to write prompts with an anti-migration stance, and this effect was stronger for the broad topic of migration than for a specific anti-migration popular initiative. Exposure to an attitude-consistent AI response did not further increase this selectivity, and neither AI attitudes nor thinking styles moderated it.
In briefCapacity planning for running clinical language models under limited hardware
SummaryThis paper studies how to split a fixed set of four GPUs across language models of different sizes and how to route incoming requests to them. Using clinical NLP tasks such as short fact checking, biomedical semantic matching, and long clinical summarization, the authors tested 24 configurations. They report that fleet composition affects response time, latency, and long-context coverage more than the routing policy does.
In briefCivicBench: A Benchmark and Memory Framework for Civic Web Data Analysis
SummaryThis paper presents CivicBench, a 400-task benchmark covering four cities over two years, for testing whether tool-using language agents keep computed civic claims supported by the records they use. It also introduces Task-State Memory (TSM), which turns tool observations into typed assertions and retrieves state organized around unresolved constraints. Across nine completed model pairs, eight show positive mean-score differences whose 95% intervals exclude zero.
In briefGSLM2: Learning Speech Structure from Multiple Streams of Discrete Units
SummaryThis paper presents GSLM2, a generative spoken language model that learns language structure purely from speech. Its SpidR 2.0 encoder produces discrete speech units at multiple levels of abstraction, and a multi-stream model jointly models them. Scaled up to 6M hours of data and 6.5B parameters, it improves linguistic competence and narrows the gap with text-based models. The training code and model checkpoints are open-sourced.
WhyThe authors propose multi-stream discrete speech units with a joint autoregressive model, reporting gap reductions on sWuggy, sBLIMP, and tSC up to 6M hours and 6.5B parameters.
In briefGESTALT: Generating Tissue Pathology Images from Gene Expression Quickly
SummaryThis paper presents GESTALT, a method that turns gene expression into histopathology images by predicting a pathology foundation model's image embedding. A lightweight transformer predicts the embedding from multi-cell gene expression, and a diffusion decoder (PixCell, frozen or LoRA-adapted) renders it as an image. On three Xenium datasets it outperforms the prior state of the art, GeneFlow, on perceptual metrics while training much faster, and it natively generates 1024 x 1024 images.
In briefEMG-Hybrid: an elite-memory-guided adaptive optimizer for high-dimensional feature selection.
SummaryThis paper addresses feature selection in ultra-high-dimensional, noisy datasets, where conventional binary metaheuristics often converge prematurely. It proposes EMG-Hybrid, a three-level hierarchical framework combining a dual-engine controller, an Adaptive Black Hole Gravity execution layer, and an Elite Memory Archive with discrete bitwise Differential Evolution refinement. The authors report strong gains over state-of-the-art baselines on datasets with over 100,000 features, along with substantial dimensionality reduction and near-linear parallel speedups.
In briefWorld models can rank policies without accurate absolute values; limits of audits
SummaryThis paper studies whether learned world models can rank reinforcement-learning policies without extra real environment interaction. The authors derive a pairwise simulation identity and a relative error bound, and propose RAVEL, which eliminates policies in a tournament using held-out real outcomes. On DMC cheetah-run and walker-walk with 45 policies per task, model rankings stay strong, but paired audit-error correlations are low and few decisions can be certified at small audit budgets.
In briefSIDRec: end-to-end semantic ID learning for better recommendations
SummaryThis paper presents SIDRec, an end-to-end framework that learns semantic IDs, compact discrete identifiers derived from item representations, in a way that is aware of the recommendation task. It jointly models content and collaborative information and lets recommendation supervision guide differentiable quantization. Experiments on three public benchmark datasets show SIDRec outperforms competing methods on both AUC and GAUC, with up to 5.15% relative improvement over the strongest baselines.
In briefTesting whether automatic prompt optimization helps strong reasoning models
SummaryThe paper tests the assumption that automated prompt optimizers consistently improve task performance, using paired same-day measurements across four optimizer families and eight tasks. On strong reasoning models, structured search showed no statistically significant accuracy gains, though a discrete structural genome prevented seed-collapse. The authors propose a three-dimensional headroom typology and provide tools for auditing per-call backend mixing in commercial APIs.
In briefRethinking LLM Pruning Through Test-Time Scaling
SummaryThis paper studies how structured pruning of expert, FFN, and layer components affects the reasoning of large language models. Across four math and coding benchmarks, it compares original and pruned models using pass@k at matched sampling budgets. Some pruned models match or exceed the original's problem coverage, yet correct solutions become harder to sample, sharply raising total token use. The authors conclude that pruning should be judged by both the range of solvable problems and the sampling cost of obtaining correct answers.
In briefHiGenTrans: Transferring LLM Knowledge into Generative Recommendation
SummaryThis paper proposes HiGenTrans, a three-level framework that brings large language model knowledge into generative recommendation at the input, representation, and decision stages. It adds an LLM-derived user representation before Semantic ID generation, aligns intermediate LLM layers with matching generative layers, and reranks beam-search outputs with a lightweight mixture-of-experts calibrator. On three public benchmarks, it improves Recall@10 by up to 12.9% and NDCG@10 by up to 33.7% over the strongest baselines.
In briefWhat a safety label measures: auditing multimodal LLM safety benchmarks
SummaryThis paper audits situational-safety benchmarks for multimodal large language models, which label scenes SAFE or UNSAFE and score model answers against those labels. The authors propose a Benchmark Card with seven measured fields to show whether scores come from the scene, the request, or benchmark design, and fill it for six benchmarks. They find that some benchmarks leak the label through non-scene signals, such as image file size on one chat benchmark. They also introduce the Spot-Stop decomposition, which splits the gap between ranking and labelling performance into threshold, overlap, and within-pair order components.
WhyThe authors propose Benchmark Card and Spot-Stop decomposition, auditing six situational-safety benchmarks for label leakage across fourteen open MLLMs.
In briefDo Earnings Calls Predict Financial Restatements? Testing Text Against Public Data
SummaryThis paper examines whether managers' language in earnings calls can warn in advance of financial restatements, which correct errors in previously published reports. The authors compare 109 calls preceding restatements with 218 matched control calls. They find the text predicts restatements but adds no detectable ranking gain beyond public firm information. A large language model scores higher, but it names the exact ticker in 66% of masked transcripts and was released after the restatements occurred.
In briefStreamDuo: A Live Commentator and Research Agent for Unseen Games
SummaryThis paper studies whether streaming video models can commentate live gameplay for games they were not trained on. It finds that current commentators carry over commentary timing but not knowledge of the game world. The proposed StreamDuo pairs a trainable commentator with a frozen research agent that retrieves wiki entries when the commentator cannot identify something on screen. Trained only on Baldur's Gate 3 and tested on ten unseen games, it improves entity recall, reduces leakage from the source game, and scores best on timing, liveliness and overall quality among streaming models.
In briefZero-Gated Cross-Attention: A method for gradually adding multimodal conditioning to frozen language models
SummaryThis paper proposes Inductively Gated Cross-Attention (IGCA), a zero-initialized cross-attention bridge that lets non-text inputs be added to frozen language models. At initialization it leaves the pretrained model's behavior unchanged, then allows multimodal conditioning to be introduced gradually. The authors test it on several frozen backbone models and report training and held-out loss results.
In briefCompSkill: Using Past Execution Experience to Retrieve Agent Skills
SummaryThis paper presents CompSkill, a skill retrieval method for LLM agents that goes beyond ranking skills by description. It compiles comparisons from earlier executions of the same task into skill offsets and capped key residuals, so experience changes which document the executor receives but not its text. On ToolQA, CompSkill completes more tasks than all baselines, and it performs on par with the strongest baselines on AppWorld and ALFWorld.
In briefCogRSI: Evolving and Auditing Reasoning Strategies for Inferring Others' Minds
SummaryThis paper studies whether a language model can learn better strategies for reasoning about other people's minds, and whether those strategies are correct. A grounded reader decides who saw each event in a story and uses a short written principle to resolve disagreements. A self-improvement loop writes the principle itself and keeps updates only when accuracy rises by a calibrated margin. The learned principles transfer to unseen test pairs, yet when translated into code and run on many small stories, most rules give wrong answers in at least one case they claim to cover.
In briefAuriga Technical Report: A Zero-Shot Robot Manipulation Agent Framework
SummaryThis report presents Auriga, a task-agnostic agent harness that lets a vision-language model perform robot manipulation zero-shot, without target-task trials or resets. The model makes task decisions and acts via tool calls or multi-step Python programs, while the harness provides calibrated 3D measurements, dual-arm control, and stopping and recovery when execution diverges. Auriga ranks first on the RoboDojo leaderboard with a 57.7% success rate, and reaches 73.2% on RoboCasa365.
In briefDigital-twin-assisted fast beam alignment for millimeter-wave networks, a study on quickly aiming directional wireless signals using virtual replicas of the environment.
SummaryThis paper presents a beam management architecture for millimeter-wave networks that combines digital-twin-assisted beam prediction, region-based localization, weighted ensemble learning, and explainability-driven beam pruning. The digital twin generates site-specific ray-traced channel data offline, enabling environment-aware beam prediction without extensive real-world measurements. Evaluated on real-world data, the method reports 96.80% top-1 and 98.36% top-3 beam prediction accuracy while reducing over-the-air sounding overhead by nearly an order of magnitude compared with exhaustive sweeping.
In briefReducing Forgetting in Task-Free Federated Forecasting Under Concept Drift Using Two-Tier Replay
SummaryThis paper studies how federated time-series forecasting models forget earlier knowledge when concept drift occurs in settings without predefined task boundaries. The authors propose a drift-aware continual-learning protocol in which clients monitor prediction-window RMSE and adapt only after persistent degradation. They also introduce TwoTierReplay, which separates a fixed offline representative memory from an online memory to retain long-term knowledge and recent changes. On the Beijing Multi-Site Air-Quality dataset, TwoTierReplay achieves the best stability-plasticity score in all nine target-horizon configurations, improving on the strongest baseline by about 11.4%.
In briefScute: Exact Batched Serving for Non-LoRA Fine-Tuning Adapters
SummaryExisting language model serving systems batch LoRA adapters on one GPU but reject adapters from other parameter-efficient fine-tuning methods such as DoRA, (IA), VeRA and OFT. The authors write each method in one normal form, a LoRA computation with six slots, and build the Scute system on vLLM. Scute serves five methods in one batch and matches the Hugging Face PEFT library with a mean error of at most 1.5% of each adapter's effect. It takes 5-17% more decode time than the same number of LoRA adapters.
In briefQVR3D: Query-guided visual reasoning for locating objects in 3D scenes without training
SummaryThis paper introduces QVR3D, a framework that locates objects in 3D scenes from natural language descriptions without any task-specific training. It operates on RGB-D observations using off-the-shelf foundation models, selecting complementary query-relevant views via Query-Guided Maximum Coverage and estimating grounding probability by combining instance confidence, attribute consistency, and relationship consistency. The authors report state-of-the-art zero-shot results on ScanRefer and Nr3D, and show the view selection also improves frame efficiency on several 3D vision-language tasks.
In briefDecentralized collision checking for drone swarms, tested outdoors
SummaryThis paper describes deploying a decentralized safety verification algorithm on a swarm of small drones to prevent collisions in real time outdoors. The authors modify and analyze a reachability-based collision avoidance method, testing it over 140 outdoor flights including six flights with eight UAVs. Each agent takes about 500 to 520 milliseconds on average to verify its safety for the next 4 seconds, with reachable set broadcasting accounting for most of that time.
Grant Phillips, Hoang-Dung Tran, Justin M. Bradley
In briefCuPH: A parallel GPU method for computing persistent homology of images
SummaryThis paper presents CuPH, a GPU-based parallel algorithm for computing cubical persistent homology, a method that describes the shape and connectivity of image data at multiple scales. It processes pixels directly, grouping them by intensity levels and tracking connected components with a lock-free union-find method. The authors prove that CuPH matches standard cubical persistent homology on integer inputs and give a stability bound for floating-point inputs. The method is reported to run 18.6 to 99.6 times faster than the existing GPU baseline.
Guangyu Meng, Pengfei Gu, Erin Wolf Chambers et al.
In briefCaching Multi-Turn Chatbot Replies to Cut Language Model Serving Costs
SummaryThis paper proposes a semantic caching method for large language models that reuses stored responses in multi-turn conversations. It finds that embedding similarity poorly predicts whether a cached reply is acceptable, so it learns a reuse risk from judge feedback and selects which conversation states to cache under a budget. The proposed methods outperform prior semantic caching approaches on math-tutoring and movie-ticketing conversation datasets, with reported reductions in generation cost and serving latency.
In briefChange the Funnel, Not Who May Submit: Redesigning peer review for surging submission volumes
SummaryThe paper addresses the growing review burden and inconsistency of decisions at machine learning conferences as submissions surge. The authors simulate, in a model calibrated to NeurIPS experiments, the current response of limiting who may submit against an alternative staged selection funnel. They report that the staged funnel uses a little over half of today's expert hours and accepts a higher share of top papers.
In briefCross-view referring reasoning for locating objects from text and examples
SummaryThis paper addresses cross-view object geo-localization, which finds a target in a satellite image from a photo taken under a different viewpoint. It introduces cross-view visual grounding guided jointly by a referring expression and a visual exemplar, and proposes GAV, a view-aware framework that transfers visual and linguistic evidence across views. GAV reaches 62.55% Acc@0.5 on MoP-UAV, and the authors also introduce the CVOGL-Text benchmark.
In briefSynergetic Autonomous Intelligence: Three Axioms for Learning Systems
SummaryThis paper introduces Synergetic Autonomous Intelligence (SAI), a theory and design program for learning systems that form, preserve, revise, and compose relations while continuing to learn. It is built on three axioms: consequential distinctions must not be orphaned, identity is set by future-faithful continuation under a declared contract, and unresolved defects are repaired by the least lawful local completion at the lowest admitted grade. The authors establish relative consistency and conditional non-redundancy of these axioms under explicit finite or affine contracts, and define the GRC-Close generative and repair schema.
In briefSparse Graph Limits for Neural Combinatorial Optimization
SummaryThis paper studies which graph information must be preserved so that graph neural networks trained on small graphs can transfer normalized combinatorial optimization values to larger graphs. It establishes size-transfer guarantees using finite collections of energy features, with separate results for affine objectives and general Lipschitz objectives. The theory relies on Shapley-Folkman bounds and a fixed-weight message-passing network on bounded-treewidth graphs. Experiments with exact solvers and trained models test the resulting guarantees.
In briefRipple: Deadline-Aware CPU-GPU Sharing for Embodied Inference
SummaryThis paper presents Ripple, a CPU-GPU co-scheduler for unified-memory edge devices that run LLM/VLM inference alongside time-critical CPU tasks. It combines model-derived GPU memory volumes with CPU completion feedback, converting each CPU job's deadline slack into a GPU-byte budget to regulate interference. Evaluated on DGX Spark, Ripple consistently meets the 2% CPU deadline-miss target and VLM deadlines across tested settings while retaining best-effort LLM throughput.
In briefLabel Smoothing in Disguise? Gated Knowledge Distillation for Low-Resource Bangla Generation
SummaryThis paper tests two gated knowledge distillation methods for Bangla text generation: EWAD, which weights tokens by teacher confidence, and CHAD, which weights examples by gradient alignment with validation data. Across six Bangla-centric tasks, three student models, and two teacher families, neither gate significantly beats plain distillation in 45 comparisons, and distillation fails to beat cross-entropy fine-tuning on gold outputs. The authors conclude that the main benefit of distillation is regularisation, which teacher-free label smoothing largely reproduces.
Dipto Sumit, Zobeida Zaman Chowdhury, Syeda Fabiayee Jahan et al.
In briefLearning Reward Machines Robustly from Noisy Labels
SummaryThis paper proposes Prob-IRM, a method that learns reward machines from noisy proposition histories using inductive logic programming and uses them for policy learning. It maintains a belief over reward machine states to decide when to relearn the machine and to supply additional guiding rewards. Evaluated in discrete and continuous environments under varying labelling noise, Prob-IRM performs comparably to baselines with handcrafted reward machines and outperforms recurrent neural network approaches that encode history.
Roko Parac, Lorenzo Nodari, Daniel Furelos-Blanco et al.
