In briefTest-Time Calibration Learning: a label-free method to make LLM answers more accurate and their confidence more trustworthy
SummaryThe paper proposes TTCL, a label-free framework that adapts both answer accuracy and verbalized confidence of large language models on unlabeled target-task data. It derives self-supervision signals from multiple model-generated responses and provides theoretical analysis showing it is a bounded surrogate for the ideal calibration objective. Across eight benchmarks in mathematical reasoning and factual QA, TTCL reports an average relative accuracy gain of +40.13% and an ECE reduction of +70.80% on base models.
WhyExtends calibration learning to label-free test-time adaptation using multi-sample self-supervision, validated extensively across eight benchmarks.
In briefForensic trajectory analysis to identify faulty or falsifying lunar rovers
SummaryThe paper addresses how to reconstruct the trajectories of multiple independently operated planetary rovers from sparse telemetry, to verify compliance with operational constraints such as Lunar Safety Zones. It shows that standard outlier-robust pose graph optimisation fails when Byzantine rovers provide consistent but falsified measurements. The authors propose an attribution-aware method that judges rover credibility and estimates trajectories using only credible agents' data. Tests on synthetic simulations and planetary-analogue trajectory data show it identifies Byzantine rovers and yields more accurate trajectories than existing baselines.
Lachlan Holden, Feras Dayoub, Melissa de Zwart et al.
In briefStructured graph memory that cuts token costs for continual learning in LLMs
SummaryThe paper addresses how language models can keep learning from experience at inference time without ever-growing context costs. It presents a unified view of context optimization, treating memory updates as optimization steps over the model's context, and proposes GraphMemory, a lightweight graph that stores and connects reusable strategies. Per query, only the relevant subgraph is retrieved, and the authors report competitive performance with about 81-85% fewer memory-construction tokens than their baselines.
Yehya Farhat, Michael Desmond, Anastasios Kyrillidis
In briefLLMs Update Clinical Judgments Unreliably as Patient Evidence Changes
SummaryThis study tests whether large language models appropriately revise clinical judgments as patient evidence changes, using matched intensive-care trajectories from electronic health records. Models more often increased than reduced prediction error when their estimates changed after conditioning on a prior judgment. They responded more strongly to worsening respiratory evidence than to matched improvement, and shifting prior risk from 10% to 90% moved estimates by 26.2 percentage points. Prompting did not fix this, while the proposed EVLU method yielded fewer but more reliable revisions, showing a reliability-coverage trade-off.
In briefRAOA: Neural computing that alternates problem and mixing steps using programmable radio propagation
SummaryThe paper proposes the Radio Alternating Operator Ansatz (RAOA), a recurrent architecture that treats programmable radio propagation as computational depth. It alternates an energy-derived problem update with a mixing update on a persistent latent state. The authors evaluate it on exact discrete optimization, constrained propagation simulation, and adapting pretrained language models. The RF realization evidence is simulation-based, not a hardware demonstration.
In briefDataWeave: A Human-LLM System for Exploratory Analysis of Structured Data
SummaryThis paper presents DataWeave, a system that supports exploratory analysis of structured data by combining conversational interaction, schema grounding, analytical planning, and executable query generation. It treats large language models as interactive partners whose outputs users can inspect, correct, and steer, rather than as autonomous answer engines. A case study with professional journalists analyzing the U.S. Department of Education's IPEDS dataset informs the design principles and deployment lessons the authors report.
Raquib Bin Yousuf, Harith Laxman, Vitaliy Shkremetko et al.
In briefSpend Teacher Tokens Where They Matter: Success-Referenced On-Policy Distillation
SummaryThis paper introduces SR-OPD, a method that reduces the cost of on-policy distillation by choosing which prompts and student rollouts receive teacher supervision. It uses a successful rollout as a reference to select failed rollouts and accounts for teacher-input cost. Across three teacher-student pairs and six math reasoning benchmarks, SR-OPD uses only 3.46-5.02% of the teacher-input tokens of Vanilla OPD in the one-pass setting while keeping comparable reasoning performance.
In briefOne Photon, Many Worlds: Posteriors and Predictions with Single-Photon Cameras
SummarySingle-photon avalanche diode (SPAD) cameras output binary frames, where each pixel shows whether any photon arrived. Many different scenes can produce the same binary frame, so recovering a scene from one frame is inherently ambiguous. The paper proposes a conditional generative framework based on a Hypergeometric frame-thinning process and studies how this ambiguity shrinks as more binary measurements are accumulated. It also examines the effect on downstream tasks such as character recognition, QR code decoding, and facial analysis.
Haejoon Lee (Carnegie Mellon University), Mohit Gupta (University of Wisconsin-Madison), Vijayakumar Bhagavatula (Carnegie Mellon University) et al.
In briefAsterism: turning scattered paper findings into literature-grounded hypotheses and theories
SummaryThe paper presents Asterism, a system that extracts concept-relation triples from hundreds of papers and unifies concepts in a hierarchical ontology so researchers can aggregate observations into theories. In a field deployment with 10 researchers, participants moved from observations to theories while keeping concepts and hypotheses matching their preferences. In two case studies, immunology and agriculture teams found mechanisms beyond their standard analyses and built hypotheses worth follow-up experiments.
Joseph Chee Chang, Michael D'Arcy, Amy X. Zhang et al.
In briefLEAP: Learning Efficient Action Proposals for LLM Agents
SummaryLLM agents run slowly because they complete tasks one step at a time, reasoning and then acting. The authors build a latency framework that weighs the gains and costs of speculative action proposals, and use it to design LEAP, which trains a small 0.6B drafter on target action sequences. The paper reports up to 60% faster end-to-end wall-clock time with no systematic change in task success. It also states the drafter can be trained online without prior trace collection.
In briefCHASE-VLA: Chunk-aware post-training quantization for robot vision-language-action models
SummaryThis paper addresses the difficulty of low-bit post-training quantization for the diffusion-based action expert in Vision-Language-Action (VLA) robot models. The proposed CHASE-VLA method uses the generated action chunk, including its unexecuted future part, as causal context, together with denoising-step group information, to adapt activation scales without modifying the pretrained policy. On LIBERO, the method reports a 97.3% average success rate on π0.5 with W4A4 quantization, restoring FP16-level performance, and reduces weight storage and memory traffic of the quantized layers.
In briefLarge Language Continuous Diffusion Models: a text generation method that works in continuous latent space instead of discrete tokens
SummaryThe paper presents Sigma, a 3B and 8B parameter continuous diffusion language model that generates text along low-dimensional ODE/SDE latent trajectories. It is trained blockwise with likelihood optimization and warm-started from autoregressive model weights. On math reasoning and coding benchmarks, the authors report performance competitive with discrete models.
In briefLong-horizon agents may break earlier safety rules; this paper names the risk GHOST and proposes a defense, STAR-Guard.
SummaryThe paper argues that long-horizon agents, through extended interaction histories, may execute actions that violate safety constraints specified many turns earlier, a failure mode it calls GHOST. Experiments report an occurrence rate of 11.5% on GPT-5.5 under benign interaction conditions. The authors propose STAR-Guard, a two-layer defense combining restoration of historical safety constraints with pre-execution audit, and report no GHOST events under the GPT-5.5 setup.
In briefGeneralization bounds for flow-matching generative models on intrinsically low-dimensional data
SummaryThis paper studies how well flow-matching generative models generalize when learning an unknown data distribution from finitely many samples. It derives finite-sample error bounds in the Wasserstein-p distance, where the convergence rate depends on the intrinsic rather than ambient dimension of the data. The results suggest flow matching mitigates the curse of dimensionality and explain its empirical success under weaker assumptions than prior analyses.
Saptarshi Chakraborty, Quentin Berthet, Peter L. Bartlett
In briefMind the Refinement Gap: When Safe High-Level Robot Plans Produce Unsafe Executions
SummaryThis paper examines whether the high-level action sequence checked by a safety monitor in language-enabled robots faithfully reflects the navigation and implicit effects that occur during execution. The authors compare RoboGuard's verdicts on surface plans and graph-refined traces under the same LTL specification across 28 controlled and 14 end-to-end cases. All 12 targeted abstraction cases showed a discrepancy between surface and refined traces, and the authors propose graph-based trace refinement as a lightweight mitigation and diagnostic tool.
In briefAIGS: An adaptive system that balances learning and memory on changing data streams
SummaryThis paper presents AIGS, a lightweight framework for online representation learning on real-time data streams whose patterns change suddenly or gradually, such as in edge and IoT settings. It introduces a Shock Ratio that compares current reconstruction error with recent variation, and uses it to smoothly balance learning plasticity against memory retention. Experiments on smart city datasets covering transformer temperature, traffic, and weather report domain-dependent gains, and the authors position it as a plug-and-play adapter for resource-constrained edge monitoring.
In briefSpectralCache: Speeding Up Diffusion-Based World Models by Caching Spectral Features
SummaryDiffusion-based world models generate high-quality interactive environments but are slow because the Transformer runs repeatedly during denoising. The authors find that features at nearby denoising steps share highly stable singular subspaces, and their singular values change predictably. They propose SpectralCache, a training-free method that reuses these subspaces, estimates singular values by linear extrapolation, and skips some costly backbone evaluations. On HunyuanWorld-Voyager-13B it reports 5.22x acceleration while keeping a WorldScore of 65.90 for static scenes..
In briefFederated learning with selective state space models: architecture-aware convergence analysis
SummaryThis paper studies how selective state space models such as Mamba2 behave in federated learning. The authors derive gradient and smoothness bounds for single- and multi-layer selective SSMs, along with convergence bounds for FedAvg and FedProx, showing how recurrent stability, input-dependent discretization, and state projection norms affect optimization. The single-layer bounds are checked numerically on sequences from a teacher SSM, and nine federated algorithms are compared on Mamba2 language modeling across six text domains.
Adam Piaseczny, Md Kamran Chowdhury Shisher, Shiqiang Wang et al.
In briefFreezing a context tower preserves generation better than in-place AR-to-diffusion conversion
SummaryThe paper compares two ways of converting a 30B Mixture-of-Experts autoregressive language model into a diffusion language model under a shared data and evaluation setup. The frozen-tower method, which conditions a diffusion model on a frozen copy of the parent, scores 71.60 on HumanEval pass@10 with 1B training tokens, versus 6.19 for the in-place method. The authors also argue that comparing diffusion language models requires a common evaluation protocol, since protocol choices substantially change reported scores.
In briefHow AI agents adopt beliefs and how those beliefs spread through groups
SummaryThe paper measures how often language model agents adopt a claim as more of their peers endorse it. The adoption curve is S-shaped, a pattern of complex contagion, and its threshold depends on the claim's plausibility, the source's reliability, and the agent's disposition. These factors are roughly captured by one dimension, the claim's coherence with the agent's prior beliefs. The authors also report faster spread on clustered networks and self-sustaining consensus that is hard to remove.~
In briefEquivariant flow matching model that predicts electron density for faster DFT calculations
SummaryThis paper introduces OrbFlow, an SE(3)-equivariant generative model that uses flow matching to predict Gaussian-type orbital coefficients, providing an initial electron density for density functional theory calculations. The authors report a 13.6% lower density error than the previous best model on QM9, and 51% to 63% lower error on each molecule of the MD benchmark. The predicted density is said to reduce self-consistent field iterations by up to 68% and to match dipole and quadrupole moments within a few percent of DFT references without any SCF calculation.
In briefCapturing Motion: A 4D Dataset for Facial Expression Intensity
SummaryThis paper introduces the 4D Facial Expression Intensity Dataset (4DFEID), containing 2,869 controlled 3D mesh sequences generated with a parametric face model. Over 90,000 subjective intensity ratings were collected on a Likert scale through crowdsourcing. The authors report that spatial-temporal graph models consistently outperform traditional frame-aggregation methods as baselines.
Zesheng Wang, Alexandre Bruckert, Pierre Lebreton et al.
In briefPerformance of zero-determinant strategies in repeated games without discounting
SummaryThis paper studies zero-determinant (ZD) strategies, which let a player unilaterally control payoffs in repeated games. Previous work assumed the Cesàro average limit of the action-profile probability distribution exists. This paper explains ZD strategy performance without that assumption, yielding a stronger result than earlier studies.
In briefBatched speech decisions without decoding, letting a frozen LLM hear beyond transcripts
SummaryDuplexJev feeds speech-encoder hidden states through a small connector into a frozen LLM and reads each question as a single-token distribution over its options, without decoding. An 8-GPU node answers 80 decisions about eight utterances in about 0.1 s. With a cross-attention connector, gender and emotion accuracy both reach 90%, while spoken QA drops only from 83% to 82%. Decisions are trained with cross-entropy on the answer token, and distillation is kept for content. The authors release weights, the training recipe, a batched-inference pipeline and a bilingual spoken-QA set.
In briefCombining a CNN with adjoint optimization to reconstruct tissue stiffness in MR elastography
SummaryThis paper addresses the inverse problem in magnetic resonance elastography, where complex shear modulus is recovered from measured shear wave displacement fields. It compares adjoint-based PDE-constrained optimization, a two-dimensional CNN reconstruction, and a hybrid method. The CNN provides a rapid initial estimate that the adjoint-based optimization then refines, which the authors report yields faster convergence and improved accuracy.
In briefChecking AI answers online while keeping verification costs low
SummaryLarge language models doing multi-step reasoning need their outputs verified, but calling an expensive ground-truth checker at every step is impractical. This paper introduces OMVV, which maintains a pool of weaker verifiers with different costs and accuracies and adaptively selects one per round using an online score combiner and an exponential-weights routing policy. The authors report that on reasoning benchmarks, OMVV achieves higher accuracy at lower verification cost than any single fixed verifier across several budgets.
In briefDesigning Better User Feedback for Generative AI
SummaryThis paper studies how users can give feedback on generative AI systems after deployment. In collaboration with eBay, the authors benchmarked current industry approaches and found issues such as poor discoverability, unclear terminology, and insufficient attention to user value. Based on these findings, they proposed best-practice recommendations and designed and tested a prototype feedback-collection tool.
In briefLost in the Request: How Phrasing Changes Affect Email Agents' Retrieval and Actions
SummaryThis paper tests whether email assistants stay reliable when the requested information, evidence, and expected outcome stay the same but the phrasing or English dialect changes. The authors built variants along five communication-style axes and four rule-based dialect conditions. They evaluated them on a retrieval-augmented generation pipeline and two tool-using agents, finding that indirect requests hurt all three benchmarks and formal requests hurt both agentic ones.
In briefImagining the future in latent space to make robot action reasoning faster and more efficient
SummaryThis paper proposes IG-VLA, a vision-language-action reasoning framework that lets robot models imagine task-relevant future scene changes in visual representation space instead of generating costly pixel-level videos. A Scene Gist Memory compresses reasoning-derived scene-behavior associations into a compact Scene Gist Token, so explicit future imagination is skipped at inference. Experiments on LIBERO, LIBERO-Plus, and VLABench report higher success rates and a reduction in per-action-chunk latency from 1081ms to 169.5ms on a single A6000 GPU.
In briefCuBEs: Culture-aware behavioral testing of LLMs and why culture-blind judges fall short
SummaryThis paper argues that existing behavioral evaluations of large language models usually ignore cultural context, which limits how well results generalize to a global user base. The authors propose CuBEs, which injects diverse cultural context into test scenarios and evaluation, and build a human-labeled dataset covering 12 cultures. Testing 13 open- and closed-source LLMs, they find that cultural context significantly changes whether a behavior appears, such as political bias surfacing along religious and colonial axes in non-Western settings rather than Western ones. They conclude that culturally situated testing is necessary for global deployments.
In briefWebUIProof: Benchmarking Web Interface Code Generators Through Executable Interaction Tests
SummaryThis paper introduces WebUIProof, a benchmark that evaluates code generated for web user interfaces by running executable interaction tests. It covers general web interfaces and 3D interactive simulations, using a UI agent that performs actions in a headless browser and checks the results. Tests on eight commercial LLMs found frequent failures in interaction requirements even when pages rendered, especially for 3D simulations. The authors also show that training smaller models with rewards from these interaction tests improves functional completion and reduces build failures.
In briefVERSE: A Verified Self-Evolving Optimizer That Improves AI Agent Harnesses
SummaryThis paper studies whether an optimizer that also improves its own diagnostic and testing tools can better enhance another AI agent's prompts, tools, and workflow. The authors find that self-evolution helps only when execution-based verification is available, and introduce VERSE, which tests draft edits, replays failures, and tracks fixes and regressions before submission. Model weights stay fixed. On held-out SWE-rebench and newer multilingual tasks, VERSE's best validation-selected harness reached 42.3% and 37.7% accuracy, above the strongest baselines at 39.2% and 29.3%.
WhyProposes a verified self-evolving optimizer improving its own tools; held-out SWE-rebench and multilingual out-of-distribution tests show clear gains over strongest baselines.
In briefA new measure, effective dimension, to quantify how problem properties affect average-case learning performance.
SummaryThis paper introduces the effective dimension, a new learning measure that links learning-problem properties to average-case learning performance. Unlike the VC dimension, it can be estimated empirically and makes average-case predictions. The measure is demonstrated on several learning systems, including Backpropagation. It is then used to predict the benefit of the feature construction system FRINGE, which decreases as the complexity of the target concept increases.
In briefWhat Is Lost in Post-Training? Default Collapse and Loss of In-Context Steerability
SummaryThis paper examines whether large language models, after post-training, can still be steered in context toward perspectives they were not trained to favor. In controlled experiments, the authors fine-tune models toward one side of cultural-value disagreements and evaluate checkpoints throughout training. The trained side becomes increasingly dominant in ordinary use, while the ability to recognize and faithfully enact the opposing view declines. The authors also propose an alternative objective that maximizes reward subject to a prescribed distribution over expressed perspectives, with stance-distribution matching as a practical implementation.
In briefLearning When to Commit Translations from Partial Speech for Simultaneous Speech Translation
SummaryThis paper studies simultaneous speech translation, which outputs usable translation before the source speech ends while keeping already committed tokens fixed. The authors create prefix training data from the model's own complete and partial waveform translations, without transcripts or human translations. On FLEURS and CoVoST2 across three language directions, prefix training improves the quality-latency trade-off, and multi-turn training reduces commit-calibration error by 63–68% overall.
In briefScale-recursive rectified flows for generating rainfall ensembles in fewer steps
SummaryCoarse satellite precipitation products cannot resolve rainfall within grid cells, so this paper proposes a scale-recursive rectified flow that generates broad rainfall patterns before local details. It allocates sampling steps across spatial scales by comparing ensemble variability with prediction error. In satellite-to-radar downscaling over the contiguous United States, giving more steps to the coarse flow improved probabilistic accuracy and rain detection. The model also outperformed a nonrecursive flow using more steps, while sampling faster.
In briefSeer: Maximum-likelihood regression for predicting machine-learning learning-speed curves
SummaryThis thesis presents Seer, a system that empirically measures classification-learning performance and fits statistical models to those measurements. The models predict how many training examples are needed to reach a target accuracy and the maximum accuracy attainable with unlimited data. The author reduces model selection to an efficient nonlinear regression solved iteratively. Experiments on soybean disease, heart disease, and audiology data are reported.
In briefForecasting Benchmarks Should Stress-Test Realistic Failure Scenarios: a call for evaluations that reveal when and why time series models fail.
SummaryThis paper argues that current time series forecasting benchmarks reward only low held-out error, and robustness studies usually rely on Gaussian noise, random masking, or bounded adversarial perturbations, which miss real deployment failures. The authors propose scenario-grounded stress testing, where each test instance includes historical inputs, future targets, a semantic scenario, an explicit failure operator, and a measurable difficulty level. They contend this makes evaluation more interpretable and attributable, showing under what conditions and why a model fails.
In briefTasteBench: A Multimodal Benchmark for Predicting Taste from Molecules to Plant-Based Foods
SummaryTasteBench is a multimodal benchmark and privacy-preserving competition for sensory prediction, aimed at speeding up the search for sustainable protein foods. It includes a food-level ranking task built on over 21,000 human evaluations of 215 plant-based foods, and a molecular-level taste classification task over 15,000 flavor molecules. The best baseline model reaches 0.661 pairwise accuracy on the rated pairs, close to the median individual panelist at 0.650.
Anna T. Thomas, Sohum Patnaik, Caroline Cotto et al.
In briefFaster LLM decoding on offloaded weights via weight-approximated activation sparsity
SummaryRunning large language models on consumer GPUs with too little memory forces weights to be offloaded to system RAM or flash, making decoding slow. This paper proposes SpAx, which skips weights tied to near-zero activations, reads compressed approximate weights for smaller activations, and reads original weights for the largest ones. With weights offloaded to CPU memory, SpAx speeds up decoding by 3.86X on average with 16-bit weights and 2.06X with 4-bit weights, with WikiText-2 perplexity increasing by at most 10%.
JuneHyung Kim, Sankeerth Durvasula, Nandita Vijaykumar
In briefGRAFT: Growing a single vision backbone by continually distilling from multiple foundation models
SummaryGRAFT is a continual multi-teacher distillation framework that merges the abilities of different vision foundation models, such as DINOv2, SigLIP2, and MASt3R, into one backbone. When a new teacher arrives, the previously distilled model serves as a teacher to preserve learned capabilities, so only a single distillation is needed. The method uses teacher-specific readout tokens and a geometry-agnostic relational loss that matches image-text similarity structures. The authors state that the resulting model covers five domains: image understanding, 2D dense prediction, 3D human pose estimation, 3D vision, and vision-language.
In briefHow causal structure links measured variables to their meanings
SummaryThe paper notes that numerical measurements describe how a system behaves but often leave variable meanings unspecified. The authors align unnamed variables' semantics using the dependence relations in a causal graph, anchored by a few known names, and build a framework called CausalBridge. In evaluations on five questionnaires and three robotics scenarios with 20% to 90% of names masked, it recovers semantics more accurately than association-based existing methods.
In briefFisher-guided data selection to reduce forgetting in LLM continual pre-training
SummaryThis paper addresses forgetting when large language models are continually pre-trained on new domain data. It proposes a selector that splits each candidate's gradient into components measuring change along parameters the pretrained model relies on and capacity in less constrained directions, then picks data using a log-determinant submodular objective in a single streaming pass. Experiments on TinyLlama-1.1B and Llama-3.1-8B with medical data report better target-domain quality and bounded forgetting. The authors state that 1B selected tokens outperform a replay baseline trained on 10B tokens.
In briefOpen-Endedness Bench (OEB): Measuring How AI Agents Reason During Research
SummaryThis paper introduces OEB, a method that evaluates an AI agent's research process by reading only its execution records, without reference answers or outcome scores. It builds a graph linking the agent's stated claims to the experiments that test them, and scores evidence, experimentation, revision, and avoidance of reward hacking. Across 119 runs on 12 tasks from three benchmarks, only 16-29% of claimed improvements were supported by the logged results.
In briefOption labels override definitions in Jev-style typed decision models
SummaryThis paper studies open typed decision models, which score caller-defined options with a label and a written definition, and finds that their probabilities depend mostly on the labels rather than the definitions, a effect called option-label bias. Experiments cover four open-weight models, three answer-reading methods, eleven classification tasks and PolicyBench, a synthetic routing suite. Deleting definitions leaves accuracy unchanged, and renaming options to A and B raises it, while changing only the prompt rendering flips the effect. The authors locate the cause in prompt rendering rather than the constrained decision head.
In briefDense Mixture-of-Experts as a Wide FFN: Expert Count Sweep at Fixed Compute
SummaryThis paper studies a dense analogue of Mixture-of-Experts, where all K SwiGLU experts are active for every token and combined by a softmax gate at fixed total FFN width. Validation loss varies non-monotonically with K: K=2 beats the K=1 baseline, while K=4 and K=6 perform worse. The authors also show the architecture is equivalent to a dense SwiGLU with token-dependent, simplex-constrained scaling, and these results come from single runs.
In briefPhysical AI Smart Spaces: A Large-Scale Multi-Camera 3D Perception Benchmark
SummaryThis paper presents Physical AI Smart Spaces, a large-scale benchmark for 3D perception in indoor smart spaces such as warehouses, hospitals, and retail venues. It provides over 280 hours of synchronized 1080p footage from nearly 1,800 cameras, with automatic annotations including identities, 2D and 3D bounding boxes, calibration, and depth. The authors extend the HOTA tracking metric to 3D and report baseline results from the AI City Challenge.
In briefDeciding Whether Private External Data Helps Transfer Learning
SummaryThis paper studies how to decide whether purchasing or collaboratively using external data will improve prediction, given that only aggregated statistics may be shared. It models dataset selection with high-dimensional regression across multiple heterogeneous sources using a weighted ridge estimator, with privacy guarantees under ρ-zero-concentrated differential privacy for labels alone or for features and labels jointly. The main contribution is a deterministic equivalent of test error that captures sample size, covariance structure, model shift, regularization and privacy noise, and the authors support the theory with experiments on synthetic and real-world data.
Filip Kova\v{c}evi\'c, Edwige Cyffers, Stefano Sarao Mannelli et al.
In briefLanguage models that answer clinicians' questions over trial evidence tables and show their work
SummarySystematic reviews summarise clinical trials in evidence tables, but clinicians can usually query them only through database searches, and many questions concern attributes the table does not record. This paper presents FD-SCoPE, a language-model framework that answers these questions, exposes the query, selected trials and derivation rule behind each answer, and learns from expert corrections. On an oncology table of 159 immune checkpoint inhibitor trial records, it completed all 140 clinician-style tasks and outperformed four alternative approaches on questions requiring derived attributes.
Manan Roy Choudhury, Suparno Roy Chowdhury, Swastik Sahoo et al.
In briefDISSOLVR: A transparent, fast framework for predicting how well molecules dissolve in water and organic solvents.
SummaryThe paper introduces DISSOLVR, a transparent framework for predicting molecular solubility by mapping molecules to physically grounded descriptors. The authors also present a literature review, a benchmark against other methods, and an LLM-assisted pipeline that turns symbolic model outputs into chemical explanations. A comparative evaluation involving 22 expert chemists is included.
Vansh Ramani, Har Ashish Arora, Dhairya Kuchhal et al.
In briefImproving the energy efficiency of LLM-generated code through effective prompting
SummaryThis paper evaluates 21 prompting strategies for energy-efficient code generation and selects 8 for testing across 10 open-weight and proprietary LLMs, for both Python and C++. Compared with a baseline prompt, the selected strategies reduced energy consumption by up to 25% for Python and 17% for C++. The effectiveness of the prompts varies across models and programming languages.
In briefSLVR: Structured latent visual reasoning using human-like reasoning flows
SummaryMultimodal models often answer visual questions by relying on language patterns instead of the image. SLVR is a training framework that organizes reasoning into typed latent stages for planning, grounding, evidence selection, and integration, each supervised by a matching signal. Built on Qwen2.5-VL-7B, it improves on several multimodal benchmarks, including MMVP and BLINK Relation, without generating textual reasoning chains at inference time.
In briefReducing AI Overagreement in Personal Advice Using Multi-Stakeholder Training
SummaryThis paper proposes Pluralistic Preference Optimization (PlurPO), a method that trains language models to identify and simulate the stakeholders in interpersonal conflicts and prefer responses acceptable to all of them. It relies only on the model's own output signals, without ground-truth labels. The authors report substantial reductions in social sycophancy across four datasets and four model families compared with prior methods. A preference dataset built for an 8B model also transferred to a 32B model.
Stephane Hatgis-Kessell, Myra Cheng, Xiaoxuan Hou et al.
In briefDAGS: Steering a Frozen Image Diffusion Transformer for Stable, Controllable Rendering
SummaryDAGS is a method that controls a frozen image diffusion transformer using two small convolutional encoders, one for appearance and one for geometry, whose features are added as per-layer residuals to the image tokens. It adds a small recurrent lighting stabilizer and a training-free temporal guidance term to make per-frame generation work as a streaming renderer. The authors report PSNR gains of 8.6 dB over Intel OIDN and 10.1 dB over RGB<->X on a 1-spp plus G-buffer input, with 2.5 to 8 times lower perceptual temporal flicker. The method is not real-time.
Karthik Mohan Kumar, Damian Andrysiak, Pedro Antonio Pena et al.
In briefLearning Closure Terms of Dynamical Systems with Kernel Ridge Regression
SummaryThis paper proposes a closure modeling framework that uses Kernel Ridge Regression (KRR) to identify missing components of dynamical systems. It covers difference-equation closures in ODE and PDE settings and algebraic closures from moment closure in kinetic equations. Experiments on the Lorenz-63 system and the Kuramoto-Sivashinsky equation show accurate long-horizon predictions and substantial gains over an LSTM-based closure model. For moment closure of a one-dimensional kinetic equation, a global model performs well on unimodal initial conditions but degrades on bimodal ones, while spatially local models with appropriate inputs are more robust and more accurate.
In briefOpenGameEval: A benchmark for testing AI agents coding and exploring inside a game engine
SummaryOpenGameEval is a benchmark that runs language model agents on game development tasks inside Roblox Studio, using reproducible stateful sessions. Each run is scored with executable checks on the edited scene and a simulated play session. Tests of 13 frontier models on 84 tasks show the tasks are hard, and that inspection behavior before acting is linked to higher pass rates. The task suite, plugin, and leaderboard are released under the MIT license.
WhySeparates observation from editing tools in a stateful engine benchmark, showing exploration predicts success across 13 models and 84 tasks.
In briefOracle selection can look accurate by chance: a null-calibrated check for thermal heart rate candidates
SummaryThe paper studies camera-based heart rate estimation, where several candidate estimates are produced from different facial regions and processing settings. It shows that an oracle that picks the candidate closest to the reference in each window can reward chance agreement, with expected error falling roughly as 1/K as candidate count K grows. On 96 thermal iBVP recordings with 168 candidates per window, the oracle's mean absolute error of 0.91 bpm is far below the best fixed configuration and best quality index, and a forehead signal from another recording reaches a similar error. The authors recommend reporting candidate count, valid coverage, and matched null controls.
Mohammad Rakibur Rahman, Nhi Nguyen, Sasan Sharifipour et al.
In briefCompressing trained sigmoid networks by merging neurons using inverse-activation regression
SummaryThis paper studies cluster-based neuron merging for compressing trained neural networks. Besides a data-free contribution-weighted averaging method, it proposes mapping neuron responses back to pre-activation space via the inverse activation function and estimating representative neuron weights and biases by least squares. The authors compare data-assisted and data-free (random input) strategies. In the tested sigmoid networks, weight information proved useful for clustering and activation information for reconstructing representative neurons.
In briefA new AI debate protocol that is provably optimal for checking complex answers
SummaryThe paper studies AI debate, where two AI systems argue to help humans judge complex questions by breaking them into simpler claims. The authors design a new protocol that guarantees correctness in the worst case, where honesty and correctness form a dominant-strategy equilibrium for both debaters. They also prove black-box lower bounds showing the protocol is instance-wise optimal among protocols using only human judgment queries.
WhyImproves debate protocols with worst-case correctness and dominant-strategy equilibrium, plus matching instance-optimal black-box lower bounds proven rigorously.
In briefTest-time coordination for multi-agent learning via decomposed value gradient flow
SummaryThis paper addresses a trade-off in offline multi-agent reinforcement learning: generative policies capture multi-modal behavior but cannot identify high-value regions, while value-optimized policies collapse to a single mode. The proposed SCOUT framework trains a flow-matching behavioral prior and a decomposed value function separately, then uses Stein variational gradient descent at test time to move behavioral samples toward high-value regions. The authors report the best average performance on discrete and continuous offline MARL benchmarks and improvements in all offline-to-online configurations.
In briefTesting how well large language models generalize across dimensions when extracting time and event expressions
SummaryThis paper evaluates how prompted large language models generalize on time and event expression extraction under four dimensions of distribution shift. It compares multiple model families, architectures, and reasoning strategies. The authors find that strong base performance generally predicts better generalization, but this link weakens under large shifts. Inductive prompting is the most consistent across the tested dimensions.
Fahmid Shahriar Iqbal, Ritam Dutt, Soumitra Das et al.
In briefReward Inflation: A Healthy Stimulus for Reinforcement Learning
SummaryThis paper proposes reward inflation, which gradually scales up rewards during training, as a healthy stimulus for reinforcement learning. The authors argue it implicitly upweights recent transitions, enabling faster policy adaptation, and that sustaining gradients as the policy saturates suppresses dormant neurons and preserves plasticity. Experiments on ALE games and MuJoCo tasks support these findings, and the paper introduces Fed, an adaptive variant that adjusts the inflation level on the fly.
In briefHow to train world models: fine-tuning versus retrieval for language-model world modeling
SummaryThis paper compares two ways to build language-model world models, which predict how environments respond to actions: fine-tuning and retrieval-augmented generation (RAG). Across five environments covering embodied, web navigation, and social settings, fine-tuned models let agents earn higher rewards in 15 of 20 settings. The authors also propose a hybrid system that encodes core environment dynamics in the model while retrieving from an actively maintained memory store, which performed best across environments and models.
Dhananjay Ashok, Shantanu Agarwal, Vivek Datla et al.
In briefENCORE: Exact parallel control for diffusion generation via replica exchange
SummaryThe paper proposes ENCORE, a method that steers a pretrained diffusion generative model toward a reweighted target distribution without retraining. Each replica stores its generation trajectory, so upward moves are truncations and the intractable time reversal is never simulated, and the authors prove target invariance. The authors report competitive accuracy and diversity on synthetic targets, Boltzmann sampling of biomolecules, and image generation, and state it is the first exact parallel control method.
WhyFirst exact parallel control via stored trajectories avoiding intractable time reversal, validated on synthetic, biomolecular, and image tasks; impact mainly within generative control.
Jiahao Yu, Saifuddin Syed, Jos\'e Miguel Hern\'andez-Lobato et al.
In briefDetecting breast implants in mammograms with foundation and lightweight CNN models
SummaryThis paper compares two foundation models (RAD-DINO and MammoCLIP) with two CNNs trained from scratch (ResNet18 and a lightweight ResNetLite) for detecting breast implants on 2D mammograms. All models performed strongly on 1,000 held-out test images. MammoCLIP achieved the highest AUROC (0.999), while ResNetLite used only 1.4% of ResNet18's parameters and had the fastest inference time. The authors suggest choosing models by deployment setting: MammoCLIP for GPU-equipped hospitals and lightweight CNNs for resource-constrained or edge devices.
Vasisht Ishwar, Hari Trivedi, Young Seok Jeon et al.
In briefA neuro-symbolic method using generative syntax to resolve structural ambiguity in Arabic noun phrases
SummaryThis paper proposes a framework that combines generative syntactic theory with the AraBERT model to resolve structural ambiguity in Modern Standard Arabic determiner phrases. Ambiguity is treated as a choice among linguistically motivated candidate structures, each evaluated through candidate-conditioned inputs. On an unseen evaluation set, the model reached 96.88% accuracy and 93.94% binary F1. Recall was 99.71% for High/VP attachment but 89.26% for Low/NP embedded attachment, indicating the embedded reading was harder to recover.
Mohammed Damom, Mohammed Q. Shormani, Muneef Y. Alshawsh et al.
In briefA Differentiable Autoregressive Neural Method for Solving Integer Programs
SummaryThis paper proposes an autoregressive method for solving 0-1 integer programs, where a transformer predicts binary variables in a fixed order while staying feasible. It is first initialized on feasible solutions from a solver, then trained with a Lagrangian penalty and Gumbel-softmax to explore the feasible set. Tests on non-convex quadratic knapsack problems with up to 10,000 binary variables show consistent gains over state-of-the-art open-source solvers on dense instances.
In briefCriticHack: Checking Visual Rewards When Robot Policies Are Optimized
SummaryThe paper studies risks of optimizing robot policies against learned visual reward models. On a drawer task, fine-tuning a diffusion policy with the Robometer reward raised both task success and wrong-object failures, while the reward still looked healthy. The reward model scores wrong-object failures slightly above successes, so standard monitoring signals miss the problem. A frozen outcome verifier redirects the same optimization toward the requested task.
WhyShows learned visual rewards amplify wrong-object failures during policy optimization, explained by a tilt model. Strong evidence from 512 seeds and controls; substantial impact on robot policy optimi
In briefLearning What to Investigate Next: Meta-Reasoning for Research Agents
SummaryThe paper proposes MIRA, a hierarchical architecture that separates deciding what to investigate from carrying out the investigation. An outer-loop meta-reasoner writes work orders for an inner-loop executor based on a persistent research record, or ends the episode. The authors also introduce MIRA-AC, a generative actor-critic trained to forecast remaining return and choose the next investigation, with reported gains across four autoresearch environments.
In briefBaCP: Backbone contrastive pruning to keep representations in extremely sparse neural networks
SummaryThis paper proposes Backbone Contrastive Pruning (BaCP), a method that aligns the embedding space of a sparse network with pretrained, fine-tuned, and historical snapshot models to prevent accuracy collapse at extreme sparsity. It builds on the contrastive decomposition of the CAP framework and gives a matched-budget analysis across several pruning criteria. Across 90 evaluated settings, BaCP substantially improves accuracy where standard pruning fails, and performs close to the baseline where representations remain intact.
Mohammad Haroon Khawaja, Muhammad Haseeb, Mohammad Fatim Shoaib et al.
In briefUsing AI agents to refine human-written cognitive models to better match behavior
SummaryThe paper proposes a hybrid system that treats discovering cognitive algorithms as program refinement. Researchers' cognitive models, written as probabilistic programs, are revised by LLM agents that identify mismatches with behavior and propose code changes within set constraints. On human behavior in a problem-solving task, the revised models consistently fit the data better than the original models.
In briefSpatial Memory Intelligence: Long-Term Memory for World Models Driven by Understanding
SummaryThis paper proposes Spatial Memory Intelligence (SMI), a framework that uses a multimodal large language model to manage spatial memory in long-video world models. SMI combines four operations: spatial clustering, within-cluster sparsification, action-aware retrieval, and reliability-aware filtering. Experiments across several baselines, benchmarks, and world-model backbones report gains in memory sparsity, generation stability, and spatial consistency.
In briefRegret Bounds for Constrained MDPs with Step-Wise Safety Constraints
SummaryThis paper studies online learning in episodic tabular constrained Markov decision processes with step-wise safety constraints, where constraints define a safe subgraph that affects reward variance and learning difficulty. The authors propose Safe Variance-Adaptive Exploration (SVAE), which learns candidate safe subgraphs and performs variance-adaptive optimistic planning. The paper gives high-probability bounds on cumulative regret and step-wise constraint violation, and proves a lower bound showing dependence on these instance-specific quantities is unavoidable.
Qian Zuo, Francesco Emanuele Stradi, Leyang Xue et al.
In briefFlowBudget: Training-free estimation of sampling steps for flow matching models
SummaryFlow matching generators need numerical integration, so inference cost depends on the number of function evaluations (NFEs), which is often chosen through repeated FID-versus-NFE sweeps. The paper introduces FlowBudget, a training-free estimator that uses interpolation probes of the learned dynamics to estimate the step budget needed for a numerical tolerance, given a fixed pretrained model and solver. It derives estimators for Euler and Heun ODE solvers and Euler–Maruyama SDE sampling. Experiments on SiT, FlowDCN, and Rectified Flow models test whether the NFE budget can be estimated without generating samples at multiple budgets.
Zahra Dehghani, Pablo Piantanida, Mohammadhadi Shateri
In briefLearning Latent Graph Structure via Feature-Centric Data Augmentation
SummaryReal-world graphs are often incomplete, which limits graph learning performance. This paper proposes SelfAug, a feature-centric graph data augmentation framework that works directly in the embedding space and uses a self-supervised inverse masking process to recover unobserved structural signals. It adds a message regularizer and a bootstrap strategy to improve robustness under noisy, sparse supervision. The authors report that across ten graph datasets, SelfAug outperforms state-of-the-art methods in accuracy and efficiency in inductive and cold-start settings.
In briefStudent-guided teacher distillation for efficient LLM task routing, compared with Jev-style System-1 classifiers
SummaryThe paper studies a distillation pipeline for routing requests across 60 LLM task categories. A compact ModernBERT student predicts the category distribution in one pass and retrieves top-k candidates, which a larger DeBERTa-v3 zero-shot NLI teacher reranks. The best student matches the teacher on 77.5% of a 200-example evaluation set, with preliminary Coverage@16 of 91-100%. The authors also argue that truncated top-k teacher scores should not be used as full 60-class soft targets for KL distillation, since this introduces systematic bias.
In briefIGNITE: A World Model Architecture for Tokamak Fusion Plasmas
SummaryThis paper introduces IGNITE, a generative world model that simulates fusion plasma behavior. It is trained in a self-supervised way on over a decade of unlabeled experimental data from the DIII-D National Fusion Facility. The core is a dynamics model that simulates DIII-D discharges from actuator trajectories, which can be supplied directly or generated from a text prompt or desired outcomes. The architecture uses several spatio-temporal tokenizers to embed different input types, such as time-series measurements, image sequences, and spectrograms, collected at very different time scales.
Peter Steiner, Azarakhsh Jalalvand, Nathaniel Chen et al.
In briefFrom Fragments to Global Maps: Learning Vectorized Map Aggregation with LLMs
SummaryAutonomous driving needs vectorized HD maps, but local predictions collected along a vehicle trajectory are noisy, fragmented, and overlapping. This paper proposes MapMergeLLM, which treats map aggregation as conditional sequence generation and uses a large language model to directly predict global map polylines. The model is trained on synthetic local maps generated from clean vector maps with corruptions simulating prediction errors. Experiments on Argoverse2 and nuScenes show it outperforms heuristic and optimization-based baselines without detector-specific retraining.
In briefPost-training quantization study for autoregressive weather models
SummaryThis paper examines how post-training quantization (PTQ) affects pre-trained AI weather emulators, aiming to reduce computation and power use. The authors implement PTQ in the DLWP and FourCastNet models as a proof of concept for geophysical fluid dynamics. They systematically test the effect on short-range forecasts and evaluate configurations via simulated quantization. The results suggest qualitatively meaningful short-term forecasts and provide a first PTQ benchmark for autoregressive weather emulators.
In briefOn-premises multi-course AI tutor for business courses: hardware and software trade-offs
SummaryThis paper presents CourseChat, an on-premises tutoring system for undergraduate business courses that answers questions using each course's assigned materials and is intended to be embedded in Moodle. The authors compared several local language models; some larger ones failed a classroom speed test, while a 12B and a 7B model passed, and they kept an 8B model in production. The paper does not establish learning gains, which remain a separate evaluation need.
In briefWorld Action Modeling with Progressive Visual Planning (ProWAM)
SummaryThis paper proposes ProWAM, a progressive world action model for robotic control. It jointly predicts actions and an ordered sequence of sparse visual sub-goals to guide task execution, avoiding dense full-video generation by caching sub-goal features in a single video-backbone pass. The authors report strong results on simulation benchmarks such as LIBERO-Plus, RoboTwin and RoboCasa365, as well as zero-shot real-world experiments.
In briefMeshQuery: Agentic Seam Planning for Automatic UV Unwrapping
SummaryThis paper presents MeshQuery, a training-free agentic method for automatically unwrapping UV coordinates of production-grade quad meshes. A vision-language model plans seams using edge-selection tools guided by natural-language UV knowledge and refines them through a feedback loop. On Adobe Substance 3D and Toys4K meshes, it yields fewer charts and shorter seams than the strongest baseline, and professional artists preferred its results in 80.9% of comparisons.
Marco Schouten, Arthur Roullier, Elie Michel et al.
In briefMulti-fidelity policy gradients for stable RL with scarce data
SummaryThe paper addresses unstable policy gradient reinforcement learning when target-domain data is scarce and expensive, by combining it with abundant but biased low-fidelity data such as from a simplified simulator. It extends the multi-fidelity policy gradient framework to PPO, redesigning sampling, advantage estimation, and control variate construction, and monitoring estimator uncertainty to avoid variance inflation. Experiments on simulated robot locomotion tasks show gains over PPO using only high-fidelity data in nearly all settings, and stable learning on a physical Franka arm using 4 real-robot episodes per update.
WhyExtends multi-fidelity control variates to PPO with fixes for failure modes; substantive improvement, supported by simulation and real-robot experiments.
In briefHXAI: Privacy-Preserving Explainable AI for Household Energy Demand Management
SummaryThis paper presents HXAI, a hierarchical framework that lets grid operators gain explainable insights into household energy use while protecting privacy. A local model generates fine-grained explanations inside a private environment, and a zonal model aggregates them for grid-level analysis under a managed privacy budget. Experiments on simulated and real energy datasets show useful zonal load-management insights, with appliance-level data kept local. The authors argue that preserving the semantic structure of explanations matters more than minimizing numerical error under differential privacy.
Poushali Sengupta, Sabita Maharjan, Frank Eliassen et al.
In briefReliability of Compound AI Systems: A Failure Catalog and Resilience Patterns from 150 Incidents
SummaryThis paper analyzes 150 production incident reports from open-source compound AI projects and anonymized enterprise deployments. It builds a taxonomy of 23 failure modes in five categories: retrieval, generation, tool, orchestration, and integration. For each category, the authors propose resilience patterns and measure their effectiveness through controlled fault injection experiments. Reported results include an 89% reduction in cascade propagation with circuit breakers and a 71% reduction in mean-time-to-recovery for systems using three or more patterns.
In brief“I just assumed it would translate”: examining healthcare staff awareness of machine translation risks using medical abbreviations
SummaryPublic healthcare staff in the UK often use machine translation, mainly Google Translate, to communicate with patients who speak other languages. Using medical abbreviations as a case, the study ran French and Spanish clinical examples from authoritative corpora through Google Translate and presented them to 21 staff across roles and specialties in semi-structured interviews. The responses were analysed qualitatively and cross-analysed.
Eleanor Taylor-Stilgoe, F\'elix do Carmo, Constantin Or\u{a}san
In briefDeepStratNet: a context-aware regression framework for seismic horizon tracking with sparse labels
SummaryThis paper addresses automatic horizon tracking in 3D seismic interpretation by replacing dense segmentation with a bounded regression that directly predicts the time or depth coordinate of a target horizon at each lateral position. The approach uses a lightweight regression head that works with any pretrained vision backbone, plus an LSTM module to model context across slices. Tested on a New Zealand seismic volume, it outperforms segmentation counterparts on RMSE and PCC and qualitatively, and it is more robust when training picks are sparser.
Aniq Ahmad, Musham Ahmad Malik, Ahmad Mustafa et al.
In briefHow much compute does one token need? Measuring sufficient per-token compute with a model panel
SummaryLarge language models spend the same computation on every generated token, even though some tokens are easier to produce than others. This paper uses a Mixture-of-Agents panel of fifteen models of different sizes to measure how much compute each token actually needs, defining it as the inference cost of the smallest model that reproduces the reference token. On the 500 MATH-500 problems, routing guided by this per-token map cut projected latency from 7.59 to 5.12 seconds while slightly improving accuracy over the best confidence-routing baseline.
In briefHarnessing LLMs as Agents: What Does It Cost? A resource theory for the computational cost of agent harnesses
SummaryThe paper introduces the Language Model Agent Machine (LAM), an abstraction that fixes the underlying model and explicitly charges the computational resources used by harness mechanisms such as bounded context, persistent memory, tools, and verification. It gives results on communication, memory access, recomputation, and reliability, including bounds and a checkpoint law. Controlled and held-out experiments on GPT-6 Astra test the communication and reliability predictions.
Zelin Zhao (Georgia Institute of Technology), Xinyu Guo (Georgia Institute of Technology), Jingyuan Zhang (Georgia Institute of Technology) et al.
In briefThreshold-Aware Conformal Routing: Deciding When a Fast Surrogate Can Replace a Slow Simulator
SummaryHigh-fidelity simulators are accurate but costly to run repeatedly, while learned surrogates are faster but less precise. The paper proposes Threshold-Aware Conformal Routing (TACR), which uses the surrogate when its conformal interval lies entirely on one side of a decision threshold and sends the input to the simulator otherwise. TACR learns an input-dependent interval scale that is tighter near the threshold, while keeping distribution-free marginal coverage. Across several scientific and engineering datasets, the authors report 14-75% fewer simulator deferrals than standard conformal prediction at the same coverage target.
In briefTurning Biomedical Sentence Encoders into Calibrated Typed-Decision Models
SummaryThis paper studies typed decision models that answer schema-constrained questions about a text in one forward pass and output probabilities. It presents SBERT2S1, which converts retrieval-trained biomedical sentence encoders into decision models, along with the BIODECIDE suite and MEDLINE-S1, 243k decisions derived from NLM indexing. The authors compare decision heads and training objectives after fine-tuning, and examine calibration after temperature scaling. Code, labels, and a model are released.
In briefMEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations
SummaryThis paper presents MEA, a multi-agent framework that generates natural-language explanations of machine learning model behavior without requiring users to have explanation-tool expertise. A Proposer agent selects and configures explanation tools, while an Actor agent is optimized end-to-end for faithfulness across tabular, text, and vision data. The authors report that MEA outperforms post-hoc explainers, agentic, and closed-source baselines on six datasets.
Yuyang Cheng, Raghav Kaushik Ravi, Srivarshinee Sridhar et al.
In briefTropical Reinforcement Learning: Using Maximum Instead of Sum to Improve Compositional Reasoning
SummaryThis paper proposes Tropical Reinforcement Learning for large language models, replacing the summing of probabilities over alternative solutions with taking their maximum, so the best verified solution path can be identified and reused. The authors introduce TROPIC, a training algorithm for deterministic, resettable environments with verifiable outcomes. On four agentic tasks (Sokoban, Countdown, FrozenLake, WebShop), TROPIC outperforms the strongest on-policy baselines by up to 16 percentage points.
Arip Asadulaev, Aladin Djuhera, Karim Salta et al.
In briefAPDMem: Adaptive Long-Term Memory for Chatbots That Digs Deeper Only When Needed
SummaryAPDMem is a hierarchical memory architecture for personalized LLM assistants that retrieves sparse evidence from long conversation histories. It organizes history into four layers, from thematic summaries to raw messages, and a controller reads high-level summaries first and drills down only when a query requires finer detail. Experiments on LongMemEval report strong long-context memory reasoning while accessing only 8% of the total conversations.
In briefWindfoil: Exact-Area Coverage Calculation for Fast, Differentiable Vector Graphics
SummaryThis paper presents Windfoil, a GPU-friendly algorithm that computes box-filtered winding numbers of quadratic Bézier contours in closed form, unifying rasterisation and differentiable vector graphics. It is implemented in WebGPU and applied to real-time 2D rendering, high-resolution print rasterisation, and a differentiable renderer. Compared with Skia and Slug, the renderer matches the reference coverage more closely, at performance comparable to Slug. The optimiser reaches reconstruction quality equal to or better than DiffVG and Bézier Splatting at a lower per-step cost, scaling to tens of thousands of shapes.
In briefSD-DPC: a sparse, interpretable learned feedback controller for nonlinear systems
SummaryThe paper presents SD-DPC, a framework that learns sparse, interpretable feedback control policies for nonlinear systems from data. A prediction model is identified with rollout-based SINDy, and the policy is a sparse combination of dictionary functions trained by differentiating a constrained predictive-control objective through that model. Across three benchmark control problems, SD-DPC satisfies constraints in all test scenarios, outperforms a distilled policy on the same terms by up to an order of magnitude, and uses far less memory and online computation than an optimization benchmark.
In briefCapability Scaling-Down Laws for LLM Compression: Predictive Relations for Choosing Compression Methods
SummaryThe paper studies how capability loss in large language models under pruning, quantization, and distillation relates to model size, compression settings, and data, and builds simple predictive relations. Sharing density responses across pruning levels halves the measurements needed, and predictions hold on new model states and under a different pruning method. The authors also assess whether these predictions help select compression methods.
WhyProposes predictive capability-loss relations across pruning, quantization, and distillation, applied to method selection. Validated on multiple model families and pre-registered tests; useful for a s
In briefA composable AI-accelerated solver for thermal modeling of 3D chip packages
SummaryThis paper presents DAIST, a thermal solver that splits full-package simulation into block-level subdomains, replaces each subdomain solver with a neural operator, and couples them by iteratively exchanging interface temperature and heat flux. Evaluated on a multi-chiplet system and an advanced packaging system, it reports up to 178× speedup over traditional FEM solvers with mean temperature errors of 0.068% and 0.323%. The authors also show that block-level models can be reused across structurally different package assemblies.
In briefCUE: Calibrated User Simulators for Multi-Turn Agent Benchmarks
SummaryThis paper introduces Calibrated User Embeddings (CUE), a framework that encodes real user sessions and decodes sampled representations into persona commands that steer LLMs to act as user simulators without training. Existing simulators match human style but do not reproduce how often and how agents fail with real users. On τ²-Bench, CUE-based simulators better match real-user success rates and failure modes than other persona methods while keeping competitive user fidelity, and they generalize to other tasks and simulator LLMs.
In briefFinDialogLens: Extracting Missed Trades from Multi-Party Financial Chatroom Dialogue
SummaryThis paper presents FinDialogLens, a hybrid LLM pipeline that extracts events from multi-party financial chatrooms to identify missed trades. Small fine-tuned classifiers detect request-for-quote triggers and outcomes, and the system reports 92.1% accuracy on final price and 94.3% on trade outcome with GPT-4o. A difficulty-aware router reduces LLM calls by 85% on final price to lower cost.
