In briefLA-CPD: Locating AI-written spans in human-LLM co-written text
SummaryThis paper proposes LA-CPD, a method for finding which parts of a document were written by an LLM and which by a human. It takes sentence-level scores from a frozen detector and combines segment consistency with contrast around candidate cut points. Dynamic programming and an AIC-style criterion select the final partition. On a held-out human-LLM co-authored test set, sentence-level accuracy rises from 0.747 to 0.796 over the WCP+AIC baseline.
In briefIs the uncertainty map wrong or its target? Exact diagnostics and a gradient-free alternative for Tweedie diagonals
SummaryDiffusion models can predict follow-up medical scans, but clinicians need per-voxel maps showing where predictions are trustworthy. The authors compute the exact diagonal of the Tweedie posterior covariance across six checkpoints and fourteen conditions, finding that in four real-image conditions it is anti-correlated with denoising error, down to -0.13, so the target rather than the estimator limits these maps. They also introduce T-PT, a gradient-free residual probe offered as an instrument, and compare it with Monte-Carlo ensembles on full-resolution brain MRI and lung CT.
In briefWhy Transformers Miss the Difference Between Causative Sentence Forms
SummaryThis paper tests whether encoder transformers such as DeBERTa apply the M-Heuristic, the principle that marked linguistic forms imply marked meanings. Using natural language inference on 188 conditions with 15 ambitransitive verbs, it compares lexical and periphrastic causatives. DeBERTa, RoBERTa, and BART show no sign of capturing this pragmatic distinction, and DeBERTa predicts "Neutral" in every case.
In briefEntropy-guided credit assignment that rewards surprising successes and penalizes repeated failures in LLM reasoning
SummaryThis paper introduces Entropic Advantage Policy Optimization (EAPO), a reinforcement learning method for improving reasoning in large language models using verifiable rewards. EAPO combines normalized policy entropy with the sign of response advantage to assign token-level credit without extra supervision. It strengthens high-entropy decisions in successful responses, penalizes low-entropy decisions in failed responses, and softens penalties at uncertain positions. The authors report that EAPO achieves the best overall performance across reasoning tasks and promotes broader exploration and more diverse answers.
In briefDiverse reasoning-route SFT data improves post-RL generalization
SummaryThis paper studies route diversity, the variation in reasoning-step sequences within supervised fine-tuning (SFT) data, and proposes a lightweight rule-based fingerprint to select diverse examples. Under matched conditions, selecting diverse routes improved post-reinforcement-learning problem coverage on puzzles and mathematics. The authors present reasoning-route diversity as a practical criterion for choosing SFT data that better prepares models for RL.
In briefEvidence-Inference Reconstruction: Correct Evidence, Flawed Reasoning
SummaryThis paper introduces Evidence-Inference Reconstruction (EIR), which uses structured state to guide one retrieval trajectory and accumulate source evidence. The authors report that once relevant evidence is collected, EIR can produce the correct answer in a single final model call even if erroneous evidence was mixed in. On 1,000-question subsets of HotpotQA, 2WikiMultiHopQA, and MuSiQue, EIR improves Answer F1 over the baselines, Agentic SSR, and Reflexion, while averaging 4.85 model calls per question.
In briefWho Warmed the Archives? LLMs Overestimate Historical Warmth
SummaryThis paper examines whether large language models can extract temperature indices from historical archival text to extend the climate record backward. Comparing methods on five centuries of German text against the Pfister temperature index, lexical methods outperformed all fine-tuned historical transformers. All six LLMs tested showed a warm bias that increases with calendar year, and the authors argue that correlation alone is insufficient to vet LLMs for this task.
In briefLearning Strategies to Break AI Judges: Using Agents to Find Weaknesses in Evaluators
SummaryThis paper proposes an agent-guided method for finding weaknesses in agentic judges that evaluate model outputs. Focusing on mathematical reasoning, adversarial agents insert errors into sound proofs, and the successful mutations are distilled into a small set of strategies that reveal how judges fail. Tested on GPT-5.6-sol and Claude Opus 5 with their agent orchestrators, the strategies consistently bypass every judge, and judge reliability drops at the frontier, with research-level flaws most likely to escape detection.
Guruprerana Shabadi, Aaditya Naik, Rajeev Alur et al.
In briefSkill2Env: Capability-oriented environment synthesis from skills for general agents
SummaryThis paper presents Skill2Env, a framework that starts from reusable agent skills and uses capability demands to synthesize challenging tasks with executable environments. It turns difficulty patterns into task blueprints and jointly builds instructions, environments, workspaces, and rubric-based evaluators. An iterative hardening method uses solver execution evidence to strengthen insufficiently challenging tasks. Supervised fine-tuning on 1.5K high-scoring trajectories from these environments reportedly yields consistent gains across a range of agent benchmarks.
In briefM3-Score: Three Separate Measures for Judging AI-Generated Radiology Images
SummaryThis paper proposes M3-Score, an evaluation framework that uses RadioDINO-s16, a vision transformer pretrained on radiology images, to assess generated medical images. It reports three separate axes: fidelity, memorization, and coverage. On BraTS brain MRI, the fidelity axis ranked five comparison sets of increasing severity correctly, and an unconditional diffusion model covered only 38% of the real distribution. The score varied much less with sample size than the Fréchet Inception Distance.
Sathiyamohan Nishankar, Pubudu Sanjeewani, and Asanka Perera
In briefDEALS: Decentralized load balancing that routes tasks among AI agent teams by expertise
SummaryThis paper proposes DEALS, a decentralized framework that lets multiple large language model agents self-organize and route concurrent tasks among themselves. Each agent keeps a local task queue and decides, based on backlog and success rate, whether to process a task or forward it to a neighbor, and partially solved tasks can be resumed by others. Experiments report improved answer accuracy and task throughput in both homogeneous and heterogeneous agent pools, with expertise and workload balanced in a self-organized way.
In briefBeyond Fixed Features: How GNN Architectures React Differently to Node Representations on Heterophilic Graphs
SummaryGraph neural networks often perform poorly on heterophilic graphs, where connected nodes tend to have different labels. This paper builds parallel feature versions of two large heterophilic benchmarks, Roman-Empire and Amazon-Ratings, using representations from static fastText vectors to contextual Transformer embeddings. Seven GNN architectures are evaluated, and the gains from richer representations differ substantially across them. Architecture and representation therefore cannot be treated as independent evaluation factors, although the relative ranking of architectures stays largely stable.
Priyanath Maji, Sidharth Gaur, and Rajavinoth Paul Durai
In briefEfficientAgent: When KV cache offloading helps concurrent LLM agents
SummaryLLM agents resend their full conversation each turn, so serving systems cache key-value state and offload it to host memory when GPU memory runs short. The paper finds that offloading's benefit varies across deployments and depends on whether the host tier can hold the reusable context of the whole agent pool, called the reuse working set. EfficientAgent sizes the host tier using a stack-distance model and limits writes when the tier is small, while writing everything when it is large enough. On SWE-bench Verified coding agents, a host tier sized to the estimated working set cut recomputed prompt tokens by 93% and end-to-end time by 39%.
WhySizes KV offloading tiers via a stack-distance working-set model and adds a write-filtering policy, a substantive improvement; validated on SWE-bench with ablations.
In briefMarginal calibration can hide differences between multivariate time series generators, tested on ERA5 weather data.
SummaryThe paper shows that pooled distributional metrics computed after marginal calibration carry little information about a generator when evaluated on held-out data. It audits StatD2GAN against re-implemented RCGAN and TimeGAN baselines on ERA5 weather sequences under a pre-specified protocol with a 168-hour embargo. Results use Holm-corrected Wilcoxon signed-rank tests over 25 matched location-seed pairs. The authors recommend floor-referenced marginal evaluation, matched-pair testing, and dependence and sequence-level metrics.
Mustafa \"Ozayta\c{c}, \"Ozge Karada\u{g} Ata\c{s}
In briefComputationally efficient online classification with random inputs and adversarial outputs
SummaryThe paper studies binary prediction where contexts are drawn i.i.d. from an unknown distribution and losses are chosen adaptively. It shows that a Follow-the-Perturbed-Leader algorithm adding a Gaussian perturbation to each observed context achieves Õ(√(T log N)) regret for N experts, using one optimization-oracle call per round without enumerating the class. For infinite hypothesis classes, the regret is Õ(√(T·VC(H))). The authors apply this to contextual bandits with K actions via uniform exploration, obtaining Õ(K^{2/3}T^{2/3}(log N)^{1/3}) regret.
WhyA simple Gaussian-perturbed leader algorithm with one oracle call per round resolves an open problem, improving regret bounds and extending to contextual bandits.
In briefPositions Are Not Facts: KV Cache and Memory Mismatch
SummaryThis paper examines how a language model's key-value (KV) cache should update stored history when a fact changes. The authors compare hiding whole records, hiding only replaced values, and deleting old text with recomputing the cache. They find that masking only the value can lose necessary details such as units, while earlier cached states still hold useful information, and that recognizing what to preserve is a separate challenge.
In briefENet-GP: a unified method that fixes warping and lighting distortions in document images together
SummaryPhotos of documents often suffer several degradations at once, such as page warping, uneven illumination, and blur. The authors introduce GutenDoc, a large dataset of dense-text documents with physically rendered compound degradations. They also propose a single network, trained once, that jointly corrects geometric and photometric distortions. Experiments report competitive results on single-distortion benchmarks and substantially better robustness under compound degradations.
In briefYuE2: A model that generates music by planning a readable score before producing audio
SummaryYuE2 is a music generation model that first writes a readable score specifying melody and harmony, then renders it as a full-song audio recording. In expert comparisons, the version using symbolic planning was preferred for overall quality (49.3%) over the version without planning (34.6%). On WildSongBench, its best-of-8 result reached 6.96, and the paper also introduces MERT2 and SheetSage2 as supporting representation and transcription models.
WhyUnifies symbolic planning with full-song audio generation, with expert listening tests and multiple benchmark wins supporting its claims.
In briefReal-time neural beamforming with a Transformer for speech enhancement on low-power hearing devices
SummaryThis paper presents a neural-network-based MVDR (minimum variance distortionless response) beamformer that runs in real time on microcontrollers in smart hearing devices. It uses six microphones and combines a CNN that estimates beamforming weights with a lightweight Transformer that applies per-frame correction, using mixed-precision arithmetic. The authors report STOI of 97.65%, SI-SNR of 20.26 dB, and wideband PESQ of 3.676 at an average power of 45.9 mW. A speech activity detector bypasses processing during silence to save battery.
Luca Bompani, Marco Fariselli, Giovanni Oltrecolli et al.
In briefShared synthetic data to detect financial fraud while protecting privacy across organizations
SummaryFinancial fraud cases are rare and data is split across organizations under privacy limits, which hurts detection. The authors examine CollaFuse, a collaborative diffusion-based method, and test it on five fraud datasets against oversampling, local generative models, and centralized diffusion. CollaFuse does not achieve the highest local fidelity but more consistently improves downstream fraud detection across most datasets.
Simeon Allmendinger, Domenique Zipperling, Burhanettin Bahadir Kibar et al.
In briefConcurrent coded signal-multiplexing ranging for asynchronous half-duplex networks
SummaryThis paper addresses signal-multiplexing network ranging (SM-NR), whose ranging cycle grows linearly with network size because its operation is sequential. It proposes CC-SM-NR, which uses binary transmit-listen codewords to coordinate concurrent transmissions in asynchronous half-duplex networks. The authors show that the minimum number of transmit-listen rounds grows logarithmically with network size, and they give conditions for constant-weight codeword feasibility under a concurrency limit. A low-complexity scheduling algorithm and a greedy extension for higher observation redundancy are presented, and simulations demonstrate effectiveness.
In briefResDiffFRG: A residual diffusion framework for generating multiple appropriate facial reactions
SummaryThis paper addresses generating varied, appropriate facial reactions of a listener in speaker-listener conversations. The proposed ResDiffFRG framework anchors the diffusion process to the speaker's facial behaviour, using the residual between the speaker anchor and an appropriate reaction as the diffusion target. Experiments report large improvements in correlation-based appropriateness over existing methods.
In briefTask-aware discretization for differentiable logic gate networks, explained
SummaryDifferentiable logic gate networks train efficient Boolean networks by relaxing discrete gates during training and discretizing them for inference. The paper shows that standard local discretization can be suboptimal for the task, even with globally optimal relaxed solutions and high gate confidence. It studies when first-order downstream task information reliably guides progressive discretization. Experiments on convolutional networks show these scores are accurate for local argmax decisions but unreliable for nonlocal interventions.
In briefPQ-HSA: A hybrid attention method reusing product-quantized scores to speed up long-context decoding
SummaryDuring decoding, a language model reads the key-value cache of all earlier tokens, which at long context is limited by memory bandwidth. PQ-HSA uses an inverted-file product-quantization index to estimate scores for unselected tokens and folds them into the same softmax as the exactly attended tokens. The paper reports higher accuracy than Quest and SnapKV at 128K context with a 1-2% retrieval budget, and a 1.6x faster decode attention call than FlashAttention-3 on one NVIDIA H20.
WhyReusing IVF-PQ approximate scores as a background term for unselected tokens is a substantive improvement, validated with multi-model and vLLM measurements at 128K.
In briefDuraS2ST: Duration-aligned speech-to-speech translation using chain-of-thought planning and reinforcement learning
SummaryThis paper proposes DuraS2ST, a framework in which one speech language model first writes an explicit chain-of-thought to plan target wording and phonetic length, then generates the speech. The authors build DuraSet-440K, a duration-aligned chain-of-thought corpus for supervised initialization, and optimize the model with multi-modal reinforcement learning using a Duration Margin Reward. Experiments on CVSS-T reportedly show a good balance between translation quality and duration consistency.
In briefMulti-marginal inverse optimal transport for contrastive learning using explicit anchor-positive-negative triplets
SummaryThe paper argues that existing inverse optimal transport methods for representation learning ignore negative pairs, which leads to dimensional collapse and weaker downstream performance. The authors propose Neg-MMIOT-CL, which concentrates a multi-marginal optimal transport coupling on anchor-positive-negative triplets. They report reduced dimensional collapse on a latent class model and significant gains over existing OT-based contrastive methods on synthetic and real datasets.
Ngoc-Hai Nguyen, Thuan Nguyen, Prakash Ishwar et al.
In briefA neural method for estimating dose-response effects of continuous treatments from observational data
SummaryThis paper introduces the Weighted Spline-Expanded Network (WSENet) for estimating causal effects of continuous treatments in observational studies. It first uses Distance Covariate Optimal Weights to make covariates and treatment distributionally independent, then fuses covariate representations with a spline-expanded treatment input to model the dose-response relationship. A weighted targeted regularization based on efficient influence functions is added to reduce residual bias, yielding a doubly robust estimator. The authors report that WSENet outperforms existing baselines in accuracy and stability on semi-synthetic and real-world high-dimensional genomic and environmental health datasets.
In briefIncremental Instructions and LLM Creative Writing: How splitting writing requirements across multiple turns affects story quality.
SummaryThis paper examines how large language models' creative writing changes when users provide requirements progressively over multiple conversation turns rather than all at once. Using 160 human-authored tasks across six genres, the authors compared upfront and 5-9 turn delivery across six open-weight model families, yielding 960 matched pairs. Progressive delivery reduced explicit constraint adherence and caused the largest drop in structure and coherence. The authors introduce a Creative Integrity measure, under which models retained 71.2% of their full-delivery score.
WhySystematically compares incremental versus upfront instructions on creative writing, quantifying structural degradation with a human study validating automated scores; useful for the subfield.
Anshuman Singh, Abrar Eyasir, Haseeb Yaqoob et al.
In briefConstrained Edit Fields for Training-Free Flow Editing: a method that changes only the intended region of an image while keeping unrelated content intact
SummaryThis paper proposes Constrained Edit Fields (CEF) for text-guided image editing with pretrained rectified-flow models, requiring no additional training. CEF assigns each spatial location a continuous edit responsibility that measures its relevance to the desired edit, and at each step splits the edit field into prompt-induced and trajectory-induced parts to suppress unintended changes. Evaluated on all 700 PIE-Bench examples, CEF achieves state-of-the-art Structure Distance, background LPIPS, and background MSE with Stable Diffusion 3.5 Medium and FLUX, while keeping competitive instruction alignment.
In briefYorùbá Character Encoding Problem in Unicode
SummaryThis paper documents a long-standing problem in writing Yorùbá online and on computers: Unicode does not encode the precomposed characters the language needs, so writers rely on combining sequences that display inconsistently, corrupt under font substitution, and fail in search. The author presents evidence from books, web platforms, and mobile keyboards. It identifies Unicode's NFC normalization stability policy as the structural barrier to a simple fix, and proposes a formal encoding request for four core Yorùbá characters as the most durable solution.
In briefHTN planning as a coordination layer for orchestrating tools across multiple MCP servers
SummaryThis paper proposes an architecture in which a Hierarchical Task Network planner generates a verifiable cross-server plan once. A runtime middleware then executes that plan deterministically across multiple MCP servers, without an LLM handling each tool call. The authors instantiate the approach in five HTN domains. They demonstrate end-to-end execution from a browser-based plan controller against eight live third-party MCP servers that query real biological databases.
In briefALDER: Finding the rules of a world by acting within it
SummaryThe paper introduces ALDER, a method that tests and revises world models written as explicit equations by proposing new experiments. It proposes parametric equations, fits their coefficients, and verifies them on held-out data, using interventions to separate competing hypotheses and discard incompatible laws. According to the paper, ALDER discovers laws beyond its initial formula set, improves out-of-distribution prediction, and uses the validated model to select control actions for goal-directed tasks.
In briefA review framing knowledge distillation, which transfers large-model abilities to smaller models, through a statistical lens.
SummaryThis review examines knowledge distillation (KD), a method for transferring the capabilities of large models to smaller student models. It presents a unified Bayesian formulation that treats teacher predictions as prior information and links this to uncertainty quantification. The authors show that classical methods and recent generative and foundation-model extensions share these statistical principles. The paper also identifies open problems for the field.
In briefTutlAit v1: A crowdsourced Moroccan Tamazight speech dataset with Arabic transcriptions and accent labels
SummaryThis paper presents TutlAit, a speech dataset of Moroccan Tamazight paired with Modern Standard Arabic text and regional accent labels. Native speakers contributed recordings and transcriptions through a purpose-built crowdsourcing web application, and the files were converted, deduplicated and administrator-validated. The dataset contains 13,384 audio files totalling about 20.9 hours. The authors state it can be reused for speech recognition, speech translation and accent identification.
Mohamed-Amine Chadi, Ezzahra Ait El Arbi, Ismail Khayoub et al.
In briefRobust Biomolecular Design for Multiple Protein Shapes via Inference-Time Adaptation
SummaryThe paper notes that structure-based design usually optimizes candidates for a single protein conformation, so they may lose interactions or clash when the target shifts to another shape. The authors introduce FlexEvo, which adapts candidates once at inference from one target structure, without retraining or needing a conformational ensemble. Across nine binder categories, the category-balanced mean relative performance degradation dropped from 47.8% to 4.4%, adding about 1.4 to 3.1 minutes per sample.
In briefReliable Replay via Spatial Coherence for Online Continual Learning
SummaryThis paper introduces SPHERE, a replay-allocation method for online continual learning that aggregates signed prospective loss changes of related memories with a representation kernel to suppress unsupported spikes. It casts allocation as entropy-regularized transport, shifting replay toward supported high-risk regions and penalizing long-distance transfers. Experiments cover noisy-label vision tasks, language-model instruction tuning, and code-generation reinforcement learning with incomplete test rewards, reporting improved accuracy and reduced forgetting.
In briefThe Rashomon Wikipedia: How Language Editions Tell Divergent Histories
SummaryThis paper compares Wikipedia articles in Romanian, Hungarian, Russian, Turkish, and English about three contested events in Romanian history. Using human annotators and large language models, the authors classify citation stance and track narrative change from 2005 to 2024. They find that language editions rarely share citations and that narratives shift with contemporary geopolitics. The paper also proposes an adversarial prompting method that asks models to preserve disagreement rather than invent consensus.
In briefGeoShrink: A training-free method that speeds up diffusion transformers with two lines of code
SummaryDiffusion transformers are costly at inference because they evaluate the model many times along a sampling trajectory. GeoShrink is a training-free method that evaluates the model only at a preset set of anchor steps and predicts the skipped steps from the latest exact output plus a geometrically scaled fraction of the most recent observed change. The paper tests it on image, video, motion, and audio generation, plus adapted 3D backends. At about 5x acceleration, it reports a 3.10 dB PSNR gain on FLUX over the strongest listed baseline.
In briefBOReFT: Steering Language Models on a Learned Manifold for Black-box Optimization
SummaryThis paper introduces BOReFT, which learns a compact, low-dimensional space of hidden-state interventions in a frozen language model and uses it as the search domain for Bayesian optimization with an external scoring function. The method is evaluated on the word search task Semantle and on three molecular property optimization tasks. According to the paper, BOReFT finds more hidden targets in Semantle and achieves higher property scores on two of the three molecular objectives than strong LLM baselines.
Dhruv Agarwal, Rico Angell, Kavitha Srinivas et al.
In briefShared Experience, Separate Learning: A Companion Module That Calibrates LLM Confidence
SummaryLarge language models often stay highly confident even when their answers are wrong. This paper studies learning confidence alongside capability during reinforcement learning from verifiable rewards, and proposes CoCal, a lightweight companion model trained on rollout hidden states and verifier-derived correctness while task optimization stays unchanged. Experiments on Qwen3-8B and Qwen3-14B show improved confidence estimation without sacrificing task performance, outperforming RL-based concurrent methods and matched post-hoc calibration. The companion generalizes across domains and policy shifts.
In briefUsing LLMs to Find Meaningful Boundaries Between Writing Revisions
SummaryThis study tests whether large language models can group fine-grained edit operations into meaningful revision units in student writing. Using 113 draft-revision pairs with 4,344 expert-annotated candidate boundaries, it compares prompted GPT-5.5 and Qwen3-32B with fine-tuned Qwen3-32B and simple baselines. Prompted models did not beat a proximity heuristic (macro-F1 .825), while the fine-tuned Qwen3-32B scored highest (.859).
Yu Tian, Andrew Potter, Katerina Christhilf et al.
In briefSelf-Designed Evaluators and Warm Memory for Long-Horizon Agents
SummaryThis paper addresses how a tool-using language-model agent working through a long stream of tasks can judge its own success when it receives no reward. The authors propose SelfSuite, in which the agent's base model designs a weighted evaluation suite from public materials, freezes it, and uses it to gate retries and label an outcome-tracked memory. On tau2-bench and AppWorld, SelfSuite beats the plain agent without labels and matches methods given ten expert labels on tau2-bench, while trailing ACE on AppWorld.
Saeid Asgari, Emre Kiciman, Leonardo de Oliveira Nunes et al.
In briefReusing one diffusion fine-tuning for multiple target domains in adaptation
SummaryThis paper studies diffusion-based unsupervised domain adaptation, where a model trained on one labeled source domain must adapt to several unlabeled target domains. The proposed MUSE framework separates source-supervised semantic adaptation, handled by a shared branch, from target-specific style adaptation, handled by private branches for each target. According to the paper, MUSE reduces diffusion fine-tuning cost compared with repeating per-target adaptation and improves average target-domain accuracy on standard benchmarks.
In briefA neural operator for solving PDEs, designed using a mathematical theory to be interpretable
SummaryThis paper introduces AFDONet, a neural solver for partial differential equations that combines neural operator learning with adaptive Fourier decomposition (AFD) theory inside a variational autoencoder structure. It targets nonlinear PDEs on smooth manifolds. The authors report high solution accuracy and competitive efficiency on several benchmarks, with stronger results on Riemannian manifolds and on data with sharp gradients.
In briefBIRD: Distilling Decision Boundaries into Rationales for MLLM Adaptation
SummaryThis paper presents BIRD, a framework for adapting general-purpose multimodal large language models to specialized domains. It uses the model's own confusable sample pairs to locate unresolved local decision boundaries and distills the visual evidence that resolves these confusions into single-sample rationales for supervised fine-tuning. Experiments on medical and chart visual question answering reportedly show BIRD outperforming competing rationale-augmentation methods.
In briefDuoOPD: Multi-task on-policy distillation that uses joint teacher-student results
SummaryThis paper proposes DuoOPD, a method for on-policy distillation where a student learns from a stronger teacher. The student's correctness sets the feedback direction, while the joint teacher-student outcome determines how the teacher supports learning, covering all four outcome combinations with one rule. On Qwen3 and Llama, DuoOPD improves mean macro accuracy over OPD by 2.58 and 5.98 percentage points and outperforms five baselines.
In briefA Spectral Theory of Compositional Learning: a mathematical account of how compositional reasoning emerges during learning
SummaryThe paper analyzes the learning dynamics of deep linear networks mathematically, training them in structured synthetic environments. It derives a theory linking the structure of experience to compositional learning. The theory predicts when compositional inferences emerge, whether they can be identified from available evidence, and how new linking evidence can rapidly unlock new inferences. The authors say these results qualitatively explain several phenomena observed in human cognition.
In briefDoes adversarial training make multi-view robot models more general? Finding and fixing view collapse
SummaryThis paper studies whether adversarial training (AT) improves how well multi-view vision-language-action robot models handle distribution shifts, testing seven LIBERO-Plus shift axes. Direct AT mainly helps with Camera Viewpoint and Sensor Noise, while effects on other shifts are mixed or negative. The authors identify "view collapse," where the model relies mostly on the wrist view, and show that adding a View Swap intervention further improves several shifts.
In briefLearning sparse random graphs with fewer, faster edge queries
SummaryThis paper studies how to recover the edges of an unknown graph by asking whether chosen vertex subsets contain any edge, with all queries fixed in advance. For Erdős–Rényi random graphs, the authors present a scheme that uses O(k̄ log n) queries and decodes in O(k̄ log n) time, recovering the graph exactly with high probability. This improves the earlier decoding bound of O(k̄^{1+δ} log n) and extends the range of sparsity where the method applies.
In briefExplaining Activation Confabulation and Reconstruction in Language Models
SummaryNatural Language Autoencoders produce text explanations of model activations, with a verbalizer writing the explanation and a reconstructor recovering the activation from it. The paper finds that standard point-reconstruction training makes explanations more useful for predicting behavior but also more prone to unsupported details and writing defects. The authors propose a standardized evaluation framework and Flow-NLA, which models the distribution of compatible activations and trains the verbalizer with a diffusion likelihood bound. Experiments on Qwen, Gemma, and Apertus report that it keeps the utility gains while curbing confabulation and writing defects.
WhyProposes distributional reconstruction over point reconstruction for NLAs, with a new evaluation framework tested across three model families; results are summarized only briefly.
In briefPARA: prompt-anchored residual adaptation for biomedical vision-language models
SummaryPretrained biomedical vision-language models classify images well zero-shot, but downstream tasks may need subtle visual distinctions. The paper proposes PARA, which keeps the frozen prompt prediction as a semantic anchor and adds a support-set visual prediction as an anchor-relative residual computed in closed form. It also introduces a repeated-support evaluation protocol that reports average and worst-20% performance. The authors report state-of-the-art results in few-shot classification and base-to-novel generalization.
In briefSpecRead: A Benchmark Testing Whether Language Models Understand Hardware Specifications
SummaryThis paper introduces SpecRead, a benchmark that separates a language model's comprehension of hardware specifications from its ability to generate code. It contains 385 questions over 10 open-source OpenTitan IP blocks, plus 82 control items, covering retrieval, cross-section reasoning, contradiction detection, and spec-RTL consistency. An initial test with the small Ministral-3B model scored 33.2% overall.
In briefOne Latent, Many Tokens: Jointly Learning Compressed Embeddings for Efficient Language Diffusion
SummaryContinuous diffusion language models must process one latent position per token at each sampling step, which makes generation costly. Two-stage methods shorten the latent sequence but fix the compressed embedding space before the diffusion model is trained. The paper introduces JPEG-DLM, which jointly trains a compressor, a flow matching model and a decoding module. The authors report the lowest mean Gen-PPL and highest throughput among recent diffusion and flow models on LM1B and OWT, reaching Gen-PPL 34.52 at compression rate 0.5 on OWT.
In briefConstellation-Aware Transformer for Semi-Supervised Signal Equalization
SummaryThis paper introduces the Constellation-Aware Transformer (CAT), a neural architecture that builds the known constellation geometry into signal equalization over unknown channels. CAT stacks custom TransFIRmer blocks that co-process received signals with ideal constellation symbols and combine an FIR-inspired filtering branch with a geometric refinement MLP. The authors report that, in the semi-supervised setting, CAT achieves a lower symbol error rate with 64 pilots than VAE and standard Transformer baselines achieve with 128 pilots on two of three ISI channels.
In briefSeeing and Solving Are Not Enough: Composition Failures in Vision-Language Models
SummaryThis paper studies whether wrong answers from vision-language models always come from failing at visual extraction or problem solving. It finds that composition failures, where both abilities succeed separately but direct answering still fails, account for 17.7% to 75.6% of direct-answering errors. The authors propose State Realization Tuning (SRT), which trains LoRA adapters so the model outputs the ground-truth task state before its final answer, improving over standard fine-tuning by 1.7 to 14.1 points.
In briefDynamic Kuramoto-Hodge operators for learning PDEs on complex geometries
SummaryThis paper introduces the Dynamic Kuramoto-Hodge Operator (DKHO), a neural operator for learning partial differential equation solutions on complex domains. It encodes fields on their topological supports (vertices, edges, faces), evolves Kuramoto-inspired relation states via boundary and coboundary operators, and decodes responses in orthogonal Hodge subspaces. Across porous-medium Darcy flow, torus transport-diffusion, and cavity magnetostatics, the large DKHO reduces prediction error by about 61% on average versus leading baselines, while the small version stays competitive with only 11.5 to 24.3% of the parameters.
In briefWhen Do Agents Help? Comparing Embedding, LLM and Agent Methods for Aligning Classical Texts with Translations
SummaryThe study compares seven alignment systems on 452 texts in Pali, Sanskrit, Mishnaic Hebrew and Tibetan, totaling 9,833 human-aligned units. Generative workflows recovered 93-94% of reference correspondences, compared with at most 77% for embedding pipelines. Agents produced structurally valid output for all 452 texts, versus 437 for direct LLM calls, and reduced judged residual defects on short passages. On long documents the agent's recovery advantage disappeared after chunking, and independent auditing offered little measurable extra benefit on prepared passages.
In briefTopoMamba: A Multi-Directional State-Space Model Guided by Load-Support Relations for Topology Optimization
SummaryThis paper presents TopoMamba, a deep learning framework that predicts high-performance material distributions for topology optimization. It couples physical fields with load-support relations and uses a spatially adaptive fusion of multi-directional scan features, combined with the solid isotropic material with penalty method. According to the authors, results on two-dimensional benchmarks show better prediction accuracy, out-of-distribution generalization, and computational efficiency than state-of-the-art models.
In briefNE-LoRA: a parameter-efficient way to update onboard satellite hyperspectral models over limited bandwidth
SummaryThe paper addresses the difficulty of frequently updating large onboard satellite models over the limited uplink bandwidth of LEO satellites. It proposes NE-LoRA, a parameter-efficient adaptation framework that pairs a low-rank branch with a nonlinear auxiliary branch and uses differentiated training for multi-matrix adapters. Experiments on four hyperspectral datasets and three backbone models show it outperforms LoRA-based baselines and is competitive with, and sometimes better than, full fine-tuning while updating only a small fraction of parameters.
In briefT-MoXAI: A Hierarchical Explainability Framework for Temporal Multimodal Data
SummaryThis paper presents T-MoXAI, a framework that explains temporal multimodal AI models at three levels: which timepoints matter, which data modalities contribute at those times, and which features or image regions drive the decision. It uses temporal Shapley values, attention analysis, and gradient-based attribution within a transformer architecture. The framework was evaluated on IVF treatment outcome prediction and wheat yield forecasting, with ablations showing temporal modelling is essential in both tasks.
In briefSWE-Game: Can AI Coding Agents Build the Games We Want? A benchmark testing how well AI coding agents can develop video games.
SummaryThe authors introduce SWE-Game, a benchmark of 247 tasks built on 41 executable Godot reference games across 13 gameplay categories in 2D and 3D. Tasks cover building games from a brief, implementing from design documents, completing skeleton code, repairing injected faults, and porting Godot to Unity. Evaluation combines engine-state checks, reference-input replay, playability demonstrations, and vision-language rubrics for presentation. Opus5 scored highest across all five task types, but the best overall scores on the three construction tasks stayed below 60 out of 100.
In briefAuditing AI agent actions by tracing query-specific causes
SummaryLLM agents increasingly take consequential actions through users, policies, and external tools, so auditing them requires tracing each action back to its historical basis. This paper introduces query-conditioned agent action attribution, a task that takes a natural-language audit question and recovers the source and ordered intermediate evidence for the queried aspect of an action. The authors release A^3Bench, a benchmark of 1,396 audit queries, and use small open-weight models that rank history units by combining query-conditioned gradient saliency with query-semantic relevance. They report better source and evidence rankings than open-weight baselines at lower cost, and a source accuracy of 64.5% versus 60.4% for the strongest frontier-model baseline.
Yifan Liu, Praveen Venkateswaran, Abdulhamid Adebayo et al.
In briefBenign Overfitting for General Norms and Distributions: a study of why linear models that perfectly fit noisy data can still generalize, extended beyond the standard Euclidean norm and Gaussian data.
SummaryThis paper studies "benign overfitting" in linear regression, where a model interpolates noisy training data yet still generalizes well. Prior work mostly analyzed minimum-2-norm solutions, but optimizers like Adam and Muon favor other norms. The authors develop an analysis for general norms and general sub-Gaussian distributions. They show that minimum-p-norm interpolation with p>1 can benignly overfit even for non-Gaussian data under suitable conditions. For the 1-norm, benign overfitting does not hold in general for well-behaved non-Gaussian distributions.
WhyExtends benign overfitting analysis to general norms and sub-Gaussian distributions, proving p>1 cases and showing 1-norm results depend on Gaussianity.
In briefReachability Is Not Enough: Diagnosing Long-Range Behavior in GNNs
SummaryThis paper proposes a framework that measures how strongly inputs at each graph distance influence a GNN's predictions, separating limits from architecture, finite approximation, training, and numerical execution. It finds that local message passing spreads influence slowly, so finite implementations may rely mainly on nearby inputs. In controlled tasks, models with similar architectural reach use distant information very differently, and low average error can hide failures on distant interactions.
In briefReset Is Not Recovery: Testing if Models Recover from Misleading Dialogue Pressure
SummaryThis paper studies whether language models return to correct, clean-context behavior after a user repeatedly pushes a wrong answer and then withdraws that pressure. Across seven open-weight instruction-tuned models and two factual benchmarks, ordinary resets reduce but do not eliminate the bias. Only methods that delete or truncate the pressure-bearing history fully recover, while history-preserving repairs are much less effective.
In briefClosing the Cross-Dialect Gap: Query Plans as a Portable Text-to-SQL Interface
SummaryThis paper observes that text-to-SQL systems are usually trained and tested only on SQLite, yet real deployments use PostgreSQL, MySQL, ClickHouse and others, and accuracy drops across dialects. The authors propose having the model generate a dialect-agnostic relational algebra query plan, which a deterministic compiler then renders as SQL for each backend. Across thirteen models, this restores cross-dialect portability nearly uniformly, and with fine-tuning on plans it outperforms SQL supervision. The paper also introduces MetricName, a result-set comparator for fair cross-dialect evaluation.
In briefRSD-Poker: A Certification Framework for Residual Game-Playing Policies Under Distribution Shift
SummaryThis paper introduces RSD-Poker, a framework that certifies the risk and utility of residual policies that modify a strong reference policy in imperfect-information games. It learns a policy-visible partition on an independent split, freezes it before calibration, and issues upper risk and lower utility certificates for each candidate-group pair. On stratified held-out states, the learned-partition dual selector raises weak-response utility and lowers violation rate compared with global certification.
In briefRobust Pseudo-Labeling Under Noise and Long-Tail Imbalance via Feature-Threshold Calibration
SummaryThis paper examines why pseudo-labeling for semantic segmentation degrades when strong imaging noise and long-tailed class distributions occur together, since tail-class predictions are filtered out by fixed high thresholds. The authors propose FTC-Seg, which uses Orthogonal Prototype Reconstruction to purify pixel features and Adaptive Threshold Calibration to recover low-confidence tail-class pseudo-labels. Experiments on four public benchmarks across three noise types show strong results against state-of-the-art methods, with especially large gains on tail classes.
In briefMeasuring each training example's effect on model behavior and adjusting weights during training to steer the model toward a target behavior.
SummaryThe paper addresses how to quickly determine each training example's contribution to model behavior before the next update, using mutual information to account for interference among examples in the same batch. It shows this mutual information is a logarithmic function of Behavioral Gradient Uniqueness (BGU) and introduces the BS-Ghost algorithm, which computes scores in batch space without storing model-sized gradients. On a 1,000-example Qwen2.5-7B-Instruct workload, the implementation adds 8.0% (27 seconds) to training time, and removal-and-retraining experiments indicate BGU identifies data that causally shapes final behavior.
WhyLinks mutual information to Behavioral Gradient Uniqueness with efficient batch-space computation, a clear novelty; evidence rests on a single 7B workload.
In briefMultiple AI Models Generating Varied Online Comments, Guided by Discussion Aspects
SummaryThis paper examines whether comments generated by large language models can capture the diversity of human online discussion. The authors propose combining multiple LLMs with aspect-conditioned generation and measure diversity along dispersion, coverage, and alignment across semantic, linguistic, and socio-pragmatic features. In a large-scale study of over 2 million YouTube comments, the approach aligns better with human comment distributions, but human diversity remains unmatched.
Nafis Irtiza Tripto, Delvin Ce Zhang, Mahjabin Nahar et al.
In briefCompoWorld: Compositional Environment Scaling for General Agents
SummaryThis paper proposes CompoWorld, which expands the task space by composing a finite library of reusable services so agents can practice workflows spanning several services. Coding agents convert tool specifications into verified services, and a random-walk procedure links services through dependency graphs to generate tasks that pass information across them. The authors report an average gain of 9.17 points over their backbone model across eight benchmarks, and say it surpasses Claude Opus 4.6 on AutomationBench.
In briefLieDiscover: adaptive symbolic library construction for discovering open-form symmetries
SummaryThis paper introduces LieDiscover, a method for discovering the unknown infinitesimal generators, i.e. the underlying symmetries, from data. It formulates symmetry discovery as a joint optimization over a function library and its coefficients, using an encoder-decoder to generate symbolic expressions and expand the library. Reinforcement learning with step-wise rewards accelerates the search of the symbolic space. Experiments show it uncovers open-form generators with high-order polynomials or transcendental functions, and the discovered symmetries improve downstream PDE solving and discovery tasks.
In briefAudit-First VAPO: Risk-Certified Selective Updates Under Imperfect Verification
SummaryThe paper proposes Audit-First VAPO, which separates discrete admission of update directions from continuous control of update magnitude. A policy that accepts, appeals, or abstains uses a limited secondary-verification budget, and its actions are frozen before clean labels are joined. Finite-sample bounds certify selected harmful risk, coverage, and verifier-call rate, and a bounded trust-clip-KL actuator then controls update size. On Qwen3.5-0.8B and GSM8K at target risk 0.08, RC-VAPO reports 74.1% accuracy and compares favorably with several baselines.
In briefLearning Multimodal Embeddings with Evidence-Aligned Readout
SummaryThis paper introduces EviAlign, which organizes task-relevant evidence from a shared multimodal large language model into five semantic units and reads the contextualized state at each unit boundary. These states are aggregated into a single normalized embedding, trained jointly with generation and contrastive retrieval objectives. Controlled experiments show that matching semantic organization with boundary readout improves results beyond either factor alone. Across 12 MMEB retrieval tasks, it reaches 76.9 average Recall@1 with 500K training pairs while keeping single-vector indexing.
In briefAgentBoundary: Counterfactual Safety Testing for Tool-Using AI Agents
SummaryThis paper introduces AgentBound, a framework that tests whether AI agents that use tools refuse or act appropriately by separately varying how risky a task looks and whether the action is actually permitted. The authors built a human-validated suite of 4,000 tasks and evaluated 17 model and harness configurations. They found that high safety often came with low completion of authorized tasks, and they trained a lightweight runtime calibration module that improved authorized-task completion and unsafe-action blocking on average.
In briefReducing Catastrophic Forgetting in Curriculum Learning for Math Reasoning
SummaryThis paper studies catastrophic forgetting in curriculum learning, arguing theoretically that it arises from divergence between task optima, which is tied to faster convergence and so cannot be fully eliminated. The authors propose IV-EWC, which adds Elastic Weight Consolidation to the curriculum objective and uses influence functions to build a representative validation set for dynamic regularization. Experiments on three backbone models and three benchmarks report an average 162% reduction in forgetting relative to vanilla curriculum learning, along with positive backward transfer.
In briefLearning Dense Rewards Without Demonstrations for Robot Policies
SummaryThis paper proposes a reward learning method that needs no expert demonstrations, deriving dense rewards from sparse task outcomes and the policy's own rollouts. It introduces eVTA₀, which learns success probabilities from mixed-quality rollouts via temporal-difference-style bootstrapping, and RLER, a closed-loop framework that updates rewards as the policy evolves. The authors report the best average policy performance across LIBERO task suites and improved real-world manipulation success rates.
In briefTsubame: a two-pass tree replay method for diffusion-based speculative decoding
SummaryThis paper introduces Tsubame, a two-pass tree speculative decoding framework for diffusion-based drafters. The first pass plans and freezes a context-aware tree topology using draft path scores, and the second pass replays that fixed topology by sampling tokens to fill its nodes for verification. The authors prove the method is lossless under compatible sampling and verification strategies. Experiments across three drafters and six datasets report higher acceptance length and throughput than deterministic trees.
In briefFuseAlign: Word-Level Forced Alignment for Real-World Audio
SummaryThis paper introduces FuseAlign, a transformer-based model that estimates when each transcript word occurs in an audio recording. The authors also propose improved evaluation metrics, a scoring protocol for real ASR transcripts, and AlignBench, a benchmark covering diverse speaker, acoustic, and text conditions. On AlignBench, FuseAlign substantially outperforms all baselines and remains robust when using real ASR transcripts.
Mithilesh Vaidya, Stephen Bailey, Sumukh Badam et al.
In briefFrom Distributions to Stochastic Processes: Neural Approximation of Distribution-Valued Maps
SummaryThis paper studies learning mappings whose inputs and outputs are probability distributions, and extends the approach from distributions to stochastic processes. The authors prove that continuous operators on compact families of finite-dimensional laws can be uniformly approximated by neural networks in the 2-Wasserstein metric, and extend this to laws on separable Hilbert spaces. Experiments predict first-passage-time distributions for an Ornstein–Uhlenbeck process and response-path laws for a Duffing oscillator, where the proposed models outperform a fixed-feature MLP and distribution-space kernel regression.
In briefProbe to Act: Improving Browser Agents with Active Visual Probing
SummaryThis paper proposes Probe to Act (P2A), a framework that lets browser agents actively probe the page before acting, aligning DOM structure with screenshot pixels at decision time. Before a state-changing operation, the agent issues lightweight probes to map DOM handles to visual evidence and keeps only probed, acted-on, or committed observations in memory. Across three browser-use benchmarks, P2A improves task success rates for both proprietary and fine-tuned models, for example raising Gemini-3-Pro from 54.1% to 61.2% on VisualWebArena.
In briefLearning From Context: Training Models on Perturbed Public Documents
SummaryThis paper proposes a synthetic training pipeline that uses lightly perturbed public documents, without human annotators, to generate context-dependent questions, rubrics, and answers for training a student model. The authors report that SFT raises a Qwen3.6-35B-A3B model's CL-bench score from 13.7% to 22.8%, and rubric-reward RL reaches 24.6%. They also note gains transfer to long-context understanding, instruction following, and reasoning, while code generation and knowledge stay mostly flat.
In briefScalable, data-driven decision support for maintenance, repair, and overhaul
SummaryThis paper proposes the Smart Expert System (SES), which uses machine learning to support engineers throughout the repair of components such as cast gas turbine parts. It describes the system's IT architecture and the machine learning models used for representative MRO use cases. Evaluated on real industry data from a leading gas turbine company, the SES is reported to meet its requirements for decision support suitability and IT architecture stability.
In briefSafeMol: Dual-Modality Safety Alignment for Molecular Multimodal Models
SummaryThis paper examines safety risks in molecular multimodal models, which can be jailbroken when handling hazardous molecules, in both text-only and graph-conditioned settings. The authors build SafeMolBench, a benchmark of 3702 samples covering 618 hazardous molecules, for training and evaluating safety, over-refusal, and utility. They propose SafeMol, a parameter-efficient alignment framework that uses MMD to align representations across modalities. Experiments reportedly cut attack success by several tens of percentage points while keeping over-refusal low and preserving molecular-task utility.
In briefTrajectory Unlearning for LLM Agents: Making Agents Forget Undesired Action Paths
SummaryExisting LLM unlearning work mainly removes specific knowledge, but this paper argues that agents should also stop reproducing undesired action trajectories, and formalizes this as trajectory-level unlearning. The authors propose Group-injected Relative Policy Optimization (GiRPO), which injects forget trajectories into the policy rollout group with penalized rewards and isolates normalization statistics. Experiments on ALFWorld and WebShop show that GiRPO unlearns target trajectories while preserving task success rates, outperforming knowledge-unlearning baselines.
In briefMeasuring how often black-box language models show severe behaviors
SummaryThe paper introduces RareTrap, a framework for estimating how likely black-box large language models are to exhibit severe behaviors. It uses a surrogate model to build a reproducible distribution over input prompts, and applies sequential rare event simulation to focus on increasingly severe responses. Tested on 10 open-weight models and two frontier models, it induced severe resource consumption behaviors and estimated their probability with as few as 200 evaluations.
