In briefRate-distortion adaptive primitive selection for omnidirectional Gaussian splatting codec (OIC-GS)
SummaryThis paper presents OIC-GS, an omnidirectional Gaussian splatting codec for 360-degree images, built on a hierarchical HEALPix primitive grid with primitives anchored at predefined spherical locations. It uses a lightweight entropy model and a spherical rate-distortion objective, automatically removing primitives whose quantized opacity becomes zero. According to the paper, a single bitstream supports full-sphere, viewport-dependent, and progressive decoding, and on a 100-image benchmark OIC-GS reduces WS-PSNR BD-rate by 49.6% over GaussianImage++ and 68.6% over SGI.
In briefWhen Harness Beats Scale, and When Reading Beats Both: A Paper on a Document-Grounded Quantitative Reasoning Shared Task
SummaryThe paper describes the authors' system for DocSem, the document-grounded quantitative reasoning shared task at DocInsights 2026, and analyzes why it succeeded on labeled data but failed on the test set. The pipeline combines hybrid block retrieval, Program-of-Thoughts generation executed in a sandboxed interpreter, self-consistency sampling, and entity enrichment from chunk-level knowledge graphs. The authors report that application architecture moved metrics far more than model scale, while on the raster, watermarked test PDFs the system reached only 13.58% joint accuracy, ranking 149 of 163.
In briefSyncRA: Teaching Omni-Modal Models to Match Sounds with the Right Video Moments
SummaryOmni-modal models can perceive audio and video but often fail to link what they hear with what they see at the same time, sometimes answering from mismatched audio-visual pairs. The authors diagnose this with controlled temporal swaps and propose SyncRA, a lightweight method that contrasts intermediate audio-visual representations within each video to align matching moments and separate mismatched ones. It uses supervision from existing input timing, needs no extra annotations, and leaves inference unchanged. Across four open models and five public video benchmarks, SyncRA consistently outperforms answer-only fine-tuning.}
In briefCoeF-SFL: a communication-efficient split federated learning method that keeps end-to-end training
SummarySplit Federated Learning lets resource-limited clients train collaboratively with a server, but exchanging data and gradients every batch is costly. Prior methods use an auxiliary network, which makes clients optimize a local objective different from the end-to-end one. CoeF-SFL avoids auxiliary networks, exchanges data and gradients once per round, and compensates stale gradients with a curvature-based correction in activation space. It reportedly outperforms auxiliary-network methods at the same communication frequency, with the largest gains on vision tasks.
In briefImproving Code LLMs by Training Runtime Program-State Reasoning
SummaryThis paper studies whether training language models to reason about runtime program states improves software-engineering abilities. It introduces two tasks: generating inputs that expose behavioral differences between a buggy program and a correct one, and characterizing bug-triggering preconditions and postconditions as executable regression tests. Using these tasks in a staged post-training pipeline, the authors build Comet-9B, a 9B model based on Qwen3.5-9B Base, and report gains on patch generation, regression-test generation, and security PoC benchmarks.
In briefFORGE: A router that picks evidence form and reasoning depth per query for frozen LLMs
SummaryThis paper proposes FORGE, a framework that adapts frozen closed-weight language models by choosing, for each query, both the form of supporting evidence and the amount of reasoning. A lightweight 269K-parameter router is trained in three stages: offline arm enumeration, KL distillation from a Boltzmann target, and GRPO refinement using host feedback. The authors report that across 5 knowledge-intensive benchmarks and 8 frozen backbones (7B to 671B), FORGE improves accuracy at 42-45% lower token cost on the two main hosts.
In briefPredictive Semantic Safety: Using vision-language models to anticipate physical hazards for safer robot control
SummaryThe paper introduces Predictive Semantic Safety (PSS), a framework linking vision-language model predictions of physical events to backup-based safety filtering. Split conformal prediction calibrates object position errors, and predicted occupancy is checked against a backup maneuver to adjust robot inputs. In MuJoCo experiments with a Unitree Go1 on falling fixtures and support loss, PSS reached a 99.3% safe episode rate versus 43.3% for a current-geometry-only baseline.
In briefFAST-Brain: A flow-aligned model that simulates resting-state brain fMRI signals
SummaryThis paper proposes FAST-Brain, a surrogate model for resting-state functional MRI (rs-fMRI) data aimed at understanding brain-wide neural activity. It combines a flow-aligned generative framework that directly predicts the clean BOLD signal, a graph convolutional network for anatomical spatial structure, and a Transformer for long-range temporal dependencies. Experiments on synthetic data and Human Connectome Project datasets reportedly achieve state-of-the-art results in recovering functional connectivity, effective connectivity, and the low-dimensional signal subspace.
In briefSemRD-V2X: Communication-Efficient Cooperative Perception via Closure-Guided Bounded Inference
SummaryThis paper studies communication efficiency in vehicle-to-everything cooperative perception, asking which remote evidence is indispensable given the receiver's own observation, using a closure-fidelity perspective. It proposes SemRD-V2X, which combines exact-budget BEV support selection, pointwise channel compression, and masked shared-weight reconstruction before standard fusion. On the simulated V2XSet dataset, compared with a locally reproduced V2X-ViT-v1 baseline, it reduces the analytical feature payload by 26.6 times and improves AP@0.5 and AP@0.7 by 4.13 and 8.57 points, with 3.81% additional mean compute latency.
In briefFuzzy-Set Modeling for Interpretable Synthetic Tabular Data with Causality Preservation
SummaryThis paper proposes a method for generating synthetic tabular data based on fuzzy set theory. Feature distributions are represented with fuzzy sets, and dependencies between features are modeled with Fuzzy Cognitive Maps, yielding a low-parameter, interpretable representation. Samples are generated by sampling fuzzy concepts rather than raw values, which natively handles mixed data types, missing values, and domain constraints. Experiments on benchmark datasets report competitive utility, fidelity, and privacy compared with state-of-the-art methods, with substantially better interpretability.
Michael Vasilakakis (1), Dimitris K. Iakovidis (1) ((1) Department of Computer Science, Biomedical Informatics et al.
In briefCommutator Memory: a readout of training-order path dependence in language models
SummaryThe paper studies how training a language model on two data sources in opposite orders leaves different weights, and asks whether this difference forms a parametric memory of training history. The authors approximate the weight difference with a Lie bracket of the two gradient fields and project it onto vocabulary tokens to get one score per token. They report that these scores are localized, that downweighting top-scoring tokens closes part of the measured loss gap in Qwen-3-4B SFT, and that the trained weights can identify which order produced them in 92% of cases.
WhyDefines training-order memory via a Lie bracket projected to tokens, validated across models with localization and causal interventions, a notably novel and well-tested approach.
In briefSAIL: Spatial Audio Intelligence with LLMs Using Separate Acoustic and Spatial Encoding
SummaryThis paper presents SAIL, a framework that lets large language models reason about spatial audio, such as where sounds come from. It separates Mel-spectrogram features from interaural phase difference features into distinct streams and learns event, direction, and distance information for each source. Compared with an early-fusion baseline, SAIL reports consistent gains on dual-source sound event detection, direction and distance estimation, and spatial reasoning.
In briefE-WAVE: Correlation-free continuous optical flow from event cameras
SummaryThis paper presents E-WAVE, a framework for estimating high-temporal-resolution optical flow from event camera streams. Instead of building all-pairs correlation volumes, it uses global attention for long-range feature dependencies and Bézier-curve trajectory-guided feature warping. Iterative updates predict trajectories that can be queried at arbitrary timestamps without repeated inference. On MultiFlow and DSEC-Flow, it reports 25% lower trajectory error and comparable endpoint flow accuracy versus baselines, and a head-mounted prototype evaluation suggests robustness in real-world conditions.
In briefCRISP: Cultural Reward Modeling for Implicit Situated Propriety
SummaryThis paper introduces CRISP-RM, a reward model that scores how culturally appropriate a language model's responses are in open-ended social scenarios. It also proposes Norm Grounding Supervision (NGS), which helps the policy better attend to relevant cultural norms during optimization. Experiments report that CRISP-RM outperforms strong general reward models in best-of-N selection and improves culturally situated behavior under GRPO, with NGS adding further gains.
In briefSteering vectors reveal the geometry behind RL gains in language models
SummaryThis paper studies reinforcement learning with verifiable rewards (RLVR) for large language models by using vector steering to find a low-dimensional effective manifold in activation space linked to RL-induced gains. The authors describe two geometric properties, effective manifold capacity and control manifold separation, supported by experiments on 5 LLMs and 6 verifiable-reward tasks. They also propose Alpha-Stabler, a plug-and-play framework that monitors principal-subspace intrusion and removes principal-subspace gradient components, which they report stabilizes training for 2,000 steps and improves RL gains.
In briefComposition Gap in Dataset Distillation: Combining Separately Distilled Sets Need Not Match Joint Distillation
SummaryThe paper asks whether merging separately distilled synthetic datasets reproduces training on the union of the real data. It shows the combination can fail even when each source is distilled exactly, because compressing training trajectories is nonlinear. For quadratic objectives the authors derive the exact composition error, and for learned synthetic sets they decompose the error into a local discrepancy and an aggregate source residual.
In briefSAGE: Structured Reasoning Framework for Efficient LLM Game Playing
SummaryThis paper proposes SAGE, a training-free inference-time framework that structures a language model's strategic reasoning into three steps: anchor to an equilibrium policy, adapt deviations using a soft belief about opponent tendencies, and recalibrate using counterfactual lessons from related past interactions. SAGE is evaluated on three repeated imperfect-information games: Leduc Hold'em, Liar's Dice, and Goofspiel. It achieves up to a 127.6% payoff improvement over reasoning-intensive LLM agents in Liar's Dice, while reducing input and output token usage by up to 80% and 90%. Code is publicly available.
In briefSkillPE: A Creativity-Oriented Method That Evolves Cinematic Skills to Write Better Text-to-Video Prompts
SummaryThe paper introduces SkillPE, a prompt engineering framework that helps non-experts get more cinematic results from text-to-video generation. It evolves reusable filmmaking skills starting from expert-written seeds and retrieves movie references grouped into three categories. On StoryEval and VBench, it improves over the strongest external baseline by up to 1.40 points and over seed skills by 0.51 points on a 7-point four-dimensional evaluation.
In briefMiCo: A training-free method that prunes visual tokens to speed up multimodal model inference
SummaryMultimodal large language models incur high computational costs when processing many visual tokens. This paper proposes MiCo, a training-free two-stage pruning method derived from a semantic erasure model and a mutual information coverage objective. It first selects a representative candidate pool using visual signals, then performs task-aware subset selection by greedily optimizing a monotone submodular coverage function under a token budget. On LLaVA-NEXT-13B, MiCo uses only 5.6% of visual tokens, retains 97.5% of baseline performance, and achieves a 3.8-fold inference speedup.
WhyDerives a training-free two-stage visual token pruning method from a mutual information objective, validated across many MLLMs and benchmarks.
In briefKnowing When Extra Thinking Isn't Enough: Small Models Learn to Query Stronger Models
SummaryThis paper examines whether more test-time thinking always helps small reasoning models, finding that self-refinement mostly concentrates probability on solutions already reachable rather than creating new ones. Based on this, the authors propose FlyBy, which reasons first, diagnoses unresolved parts, and selectively queries stronger models at knowledge bottlenecks, trained with supervised fine-tuning and cost-aware reinforcement learning. On 1,158 hard problems, FlyBy-4B reaches 45.96% pass@8, surpassing Qwen3-14B (41.64%) at 2.7 times lower serving cost, and FlyBy-8B reaches 51.81%.
WhyIntervening on intermediate reasoning states distinguishes execution from knowledge bottlenecks, motivating FlyBy's selective querying; evaluated on 1,158 problems across six benchmarks.
In briefRouting LLM queries with regex features instead of neural embeddings
SummaryThe paper finds that scaling Qwen2.5 encoders from 0.5B to 72B parameters barely improves LLM routing accuracy. The authors introduce REGEXROUTE, which uses sparse autoencoders to discover interpretable features and has an LLM turn them into regex extractors, removing the need for a neural encoder at inference. Across four benchmarks, 128 fixed features reach 76.43% average routing accuracy, close to the strongest neural text encoder baseline (76.41%), with much lower latency.
In briefDORA: A Reinforcement Learning Agent That Adaptively Removes Image Tokens in Vision Transformers
SummaryVision Transformers have computation that grows quadratically with the number of tokens. This paper proposes DORA, which learns an input-adaptive policy for a frozen ViT that decides when to prune, how many tokens to remove, and which ones. On ImageNet-1K with DeiT-Base, DORA reduces FLOPs by 38.4% with less than one percentage point of accuracy loss.
In briefBudgeted multi-attribute checking to certify better LLM answers at test time
SummaryThis paper studies how to decide which candidate LLM answers to check and which attributes to evaluate when verification is limited by a shared computational budget. The authors propose BMA-GAI, an algorithm that pairs cost-aware candidate selection with adaptive attribute sampling, and it provides an asymptotic coverage guarantee plus a matching information-theoretic converse. Experiments on synthetic benchmarks and an LLM answer-verification task show it certifies more high-quality candidates than competing methods.
In briefOne Rollout Is All You Get: Fully Test-Time Adaptation for GUI Agents
SummaryDeployed GUI agents keep frozen weights and discard what they experience on the job. The paper defines fully test-time adaptation under these constraints and proposes SOLO, a minimal method in which auxiliary models select successful episodes and relabel failed prefixes, and a small adapter is updated by self-distillation. On recurring task streams built from WebArena, VisualWebArena and MobileWorld, SOLO improves success rate over the frozen agent and outperforms two in-setting memory methods on the web streams.
In briefCertified automation of LLM agent evaluation: letting an automatic judge score part of agent trajectories with a guaranteed error rate
SummaryThe paper asks what share of LLM agent evaluation an automatic judge can take over while keeping the error rate among auto-decided trajectories below a set budget. It introduces a task-level bootstrap certificate that stays valid where naive certificates overstate safety. A 4B judge certifies 0.30 to 0.59 of evaluation on tool-use and web corpora at alpha=0.1.
WhyProposes a task-level bootstrap certificate fixing error underestimation from correlated agent trajectories, validated across multiple corpora, models, and a self-training application.
Chengguang Gan, Yunhao Liang, Qinghao Zhang et al.
In briefTraining-Free Token Caching for Faster Vision-Language-Action Robot Control
SummaryThis paper presents TVCache, a training-free framework that speeds up inference in Vision-Language-Action robot models. It filters attention heads by how well they focus on text-relevant visual regions and selects reusable layers using text-vision entropy differences. Experiments across four VLA models, two simulation benchmarks, and real-world robot tasks show higher task success than existing VLA caching at matched token retention ratios.
In briefDynamical Parameters: An Interpretability Framework for Time-Series Foundation Models
SummaryThis paper examines a gap in time-series foundation models: a dynamical property can be read from a hidden state, yet the forecast may not respond correctly when that property changes. The authors define dynamical parameters such as trend slope, oscillation frequency, and autoregressive dependence, and compare how accessible each is in hidden states with how the forecast responds. Across nine frozen models and thirteen laws, most cells showed high accessibility, but the median forecast response was only 0.46 relative to the reference. Causal geometry analysis indicates that parameter interventions move the hidden state in a different direction than required.
WhyProposes dynamical parameters separating representation accessibility from forecast response, explained via causal geometry across nine TSFMs and thirteen laws.
In briefPlaylistEval: Are Video-Language Judges Reliable on Day-Long Videos?
SummaryVideo-language models are increasingly used to judge video understanding, but their reliability on very long videos was unclear. PlaylistEval is an agentic framework that builds judge benchmarks over 100-hour playlists without human annotation, using causal degradation to control answer-pair differences. The benchmark has 630 pairs across seven domains, and on a 152-pair subset it agrees with human judgments 93.0% of the time. Evaluating 17 models, the strongest judges reach only 75.4% pairwise accuracy, with open-source judges performing worse.
WhyProposes an annotation-free framework for hundred-hour video judge benchmarks, validated against human agreement across 17 models.
In briefControlScope: How Much Language Model Agents Should Revise Running Workflows
SummaryThis paper studies how much a language model agent should revise a workflow while it is running. It compares three options: continuing the generated code, editing only the data arguments of the next tool call, and replacing the unfinished workflow from the same public execution state. The experiments cover filesystem tasks, ALFWorld, and AppWorld, and report how these revision permissions affect task completion and cost. The authors conclude that repair access should be tied to the agent's actual choices and subsequent execution, rather than one option being best in all cases.
In briefRiemannian Difference-of-Convex Optimization for K-Means Clustering: A paper reformulating K-means clustering as a constrained problem on a manifold and proposing a new solver, RADA-DC.
SummaryThe paper studies K-means clustering as a cardinality-constrained problem on a compact embedded submanifold. It replaces the constraint with a difference-of-convex penalty and shows that both formulations share the same global minimizers once the penalty parameter exceeds a finite threshold. The authors propose RADA-DC, a Riemannian alternating descent ascent method, which finds an ε-Riemannian critical point within O(ε^-3) iterations under standard assumptions. Experiments on synthetic and real datasets report better solution quality than the tested baselines, including K-means++, when the number of clusters is large.
In briefMaLiang-Harness: A Framework for Generating Images and Videos by Programming
SummaryThis paper presents MaLiang-Harness, a framework that uses multimodal large language models to generate executable programs for building images and videos. It organizes generation into a persistent cycle of construction, inspection, and revision, so that the program, its history, and its verification stay linked. The authors evaluate 11 closed-source models on MaLiang-IBench and four on MaLiang-VBench, and report that general capability scores do not reliably predict visual generation performance.
In briefCAT-Free: Multi-camera pedestrian location finding that needs no calibration, labels, or scene training
SummaryThis paper presents CAT-Free, which locates pedestrians across multiple cameras using only synchronized RGB video, without camera calibration, position annotations, or training on the target scene. It estimates camera setup from the video and uses two adaptive geometric filters to remove unreliable position estimates. It reaches 82.5, 84.5, and 65.7 MODA on WildTrack, MultiviewX, and GMVD, and transfers to new installations without retuning.
In briefOne Model, Many Drug Design Tasks: Masked Molecular Inference
SummaryThis paper introduces CAGenMol-2, a masked diffusion molecular language model that places molecules, continuous property values, and 3D protein pockets in one sequence. By choosing which regions are observed or masked at inference, a single checkpoint performs property prediction, conditional generation, and partial-constraint design. The authors also propose AdaFO, a gradient-free mask-and-refill search, which raises the success rate on CrossDocked2020 from 30.2% to 70.8%.
In briefWhen World Models Lie: Adaptive Safety Filtering Under Wrong Predictions
SummaryThis paper addresses the risk that robot world models can make biased or overconfident predictions, which can cause latent-space safety filters to give unreliable safety estimates. The authors propose an adaptive filter that uses Adaptive Conformal Inference to build online uncertainty sets from differences between predicted and observed latent states, and evaluates safety pessimistically over these sets. Simulation and hardware experiments reportedly show significantly fewer failures than state-of-the-art latent safety filters while preserving task completion. The paper also provides a finite-time coverage guarantee for the adaptive uncertainty radius.
In briefPSM: Difficulty-Based Precise Statistical Matching for Dataset Distillation
SummaryDataset distillation compresses a large dataset into a small one that still trains models well. The authors note that existing statistical matching methods supervise all distilled samples with statistics from the whole dataset, ignoring differences in sample difficulty. PSM ranks images by a Global Precision Score, splits each class into difficulty groups, updates teacher BatchNorm statistics per group, and initializes samples from matching-difficulty originals. Experiments on multiple datasets and architectures report broader difficulty coverage and better downstream performance in most settings.
In briefDr.Credit: Rubric-Grounded Process Credit Assignment for Deep Research Agents
SummaryThis paper addresses reinforcement learning for open-ended, rubric-based tasks, where rubric scores only reward final answers and do not distinguish intermediate decisions. The authors propose Dr.Credit, which uses task requirements as a shared reference to assess how much new support each tool return provides for each rubric, relative to prior accepted support. Evaluated on four benchmarks, Dr.Credit outperforms the open deep research baselines it was compared against, and its 8B-parameter agent is competitive with evaluated frontier proprietary models.
In briefPre-registered tests of physics-inspired compression for small language models, with mostly negative results
SummaryThe authors ran a three-month autonomous research-agent program testing five solid-state-physics-inspired compression methods on pretrained language models. Predictions were committed to git before pilot data, and a 3-sigma gate decided PASS or SHELVE. Most pre-registered predictions were falsified, and the attention-head decay pattern inverted the expected behavior, though this holds only at the 350M-parameter scale or below.
In briefReScraper: a small model that unifies web scraping and cleaning for LLM pretraining data
SummaryThe paper introduces ReScraper, a unified 0.6B-parameter language model that replaces the hand-written heuristic stack used to extract and clean web text for LLM pretraining. It first extracts main content from raw pages, then chooses to keep, edit, delete, or rewrite each page. Pretraining 400M, 1.4B, and 2.8B models on its curated data improves the DCLM Core score by a relative 3.8–4.7% over the strongest baseline at each scale.
In briefDexterous Tactile World Model: predicting future hand-manipulation frames from video and touch
SummaryThe paper introduces the Dexterous Tactile World Model (DTWM), a video model that predicts future frames of egocentric manipulation using both video and tactile signals from gloves worn on each hand. Compared with a vision-only model matched in architecture, parameters, and training, DTWM cuts hand-motion underestimation from 23% to 9% and reduces hand-region perceptual error by 7.4%. Ablations indicate that both the magnitude and spatial location of force contribute to the gains.
In briefLearning from imprecise labels by predicting sets of distributions instead of one
SummaryThe paper proposes epistemic learning from credal supervision, which uses convex sets of plausible label distributions as training targets and learns sets of predictive distributions. It instantiates this with the pessimistic-optimistic credal classifier (POCC), whose two heads minimise worst-case and best-case losses, and whose output spread gives an uncertainty score. Experiments on human annotator disagreement, teacher predictions, and label smoothing report a favourable balance of accuracy, calibration, and selective classification against baselines.
In briefOver-personalization as a decision failure: generation-induced Apply bias in LLMs
SummaryPersonalized LLMs must decide whether to apply or suppress each stored preference based on context, but they often apply preferences the context rules out. The paper splits preference handling into three stages and finds that most failures come from wrong decisions that the model then faithfully follows. The proposed ABIDE method shows that asking the model to generate an answer shifts its decision scores toward Apply while sensitivity is largely preserved. Subtracting a single bias value at decoding time reduces leakage while largely keeping preference fulfillment.
In briefAgentic high-dimensional Bayesian optimization guided by hypotheses and evidence
SummaryThis paper examines whether LLM-based and agentic Bayesian optimization methods, previously tested mainly on low- and moderate-dimensional problems, remain reliable when the number of variables is large. The authors find they do not, and introduce HERA, an agent that uses task context, optimization feedback, and structural diagnostics to revise search hypotheses and choose strategies, paired with the numerical engine PRISM. HERA is reported to be competitive with strong numerical baselines and to achieve the best mean final objective on most of eight real-world benchmarks.
In briefMoSPR: Predicting gene expression from tissue slide images using spatial tissue-pattern groups and low-rank gene basis
SummaryThis paper introduces MoSPR, a linear method that predicts gene expression from histopathology whole-slide images. It clusters image patches into morphology microstates, groups them into shared macrostates based on spatial adjacency, and maps slide features to a low-rank gene-expression basis learned from training data. The authors report the highest mean gene-expression prediction scores across three TCGA cancer cohorts.pathway
WhyCombining adjacency-informed morphology macrostates with a low-rank molecular basis substantially improves histology-to-expression prediction. Evaluation on three TCGA cohorts with baselines and ablat
In briefDirect Self-Evolving Optimization (DEO): Evolving Language Models Without Training a Challenger
SummaryThis paper introduces Direct Self-Evolving Optimization (DEO), a method that replaces challenger parameter updates with solver-guided task sampling. A frozen LLM generates and mutates tasks, the solver scores them, and an approximate Metropolis selection rule refines the training pool, with only the solver being trained. The authors report reasoning performance competitive with R-Zero while using over 50% less wall-clock training time. They also present theoretical results for an idealized variant under stated regularity and initialization conditions.
In briefA study on teaching vision-language models to see, measure, and reason accurately in pathology images
SummaryThis paper addresses the weak fine-grained visual perception of vision-language models in histopathology. It proposes ASPECT, which explicitly supervises cellular appearance and counts through training stages and reinforcement learning. The authors also introduce PathoVernier, a benchmark of 759 expert-reviewed questions that evaluates intermediate measurements as well as final answers. The paper reports relative accuracy gains over the strongest baseline and its backbone, and improvements on three external pathology benchmarks.
In briefBIABench: Testing AI agents on real bioimage analysis tasks
SummaryBIABench is a benchmark of 16 bioimage analysis tasks rebuilt from published biological studies, keeping their scientific questions, imaging data and ground truth. Submissions get an outcome score comparing output files to ground truth, and a process score in which a vision-language model judges method choice and quality control against an expert rubric. Routine two-dimensional tasks were solved well, but on tasks adding a third dimension or time axis no agent scored above 0.19. The agents were also unreliable across repeated runs.
WhyBuilds a 16-task end-to-end bioimage benchmark from published studies and shows agents fail on 3D and time-lapse tasks, with repeated-run evaluation supporting the findings.
In briefStudy of how query and key projections diverge geometrically in Transformer attention training
SummaryThis paper examines how the query and key projections in Transformer attention change geometrically during training. The authors trained small GPT-like models on character-level WikiText-103 across several depths, initializations, and random seeds, measuring effective dimensionality with participation ratios. They report that query dimensionality expands while key dimensionality shrinks, and that shrinking keys sharpen attention, with direct spectral control suggesting this link is causal.
Vidit Gupta, Siddhesh Nadkarni, Mihik Chaudhari et al.
In briefWhy BF16 FlashAttention gradients blow up late in training, and a fix
SummaryThe paper reports that pretraining a 450M-parameter transformer with FlashAttention-3 in BF16 caused gradient norms to grow sharply after about 25B tokens, leaving the loss above that of FP32 attention. The authors trace the problem to a fused multiply-add in the forward softmax and to BF16 rounding that breaks the zero-sum property of softmax score gradients in each row. They introduce GProj, which applies two rank-one corrections per row to restore the zero sum, reducing gradient errors substantially and matching FP32 attention's training loss at a 4.7% per-step time cost.
WhyIdentifies a broken zero-sum conservation law in BF16 softmax gradients, proposes GProj correction, and validates it in matched pretraining runs against FP32.
In briefScaling 3D Asset Editing with a Million-Scale Dataset
SummaryThis paper presents Alchemy3D, a framework for training and evaluating general-purpose 3D asset editing models, covering data, architecture, and benchmarks. The authors build Alchemy3D-1M, a dataset of 1.25M assets and 1.38M editing pairs across seven editing types, and train generative flow models on it. The models support image- and text-conditioned editing, few-step inference, and transfer to multi-view 3D part segmentation. The authors also introduce GEdit3D-Bench, an open-world benchmark, and report better results than prior methods on most metrics.
In briefBayesian active learning to resolve ambiguous robot task intentions through clarification questions
SummaryThis paper addresses ambiguous or incomplete natural language instructions given to robots by framing clarification as an active learning problem over Signal Temporal Logic (STL) task specifications. Large language models propose candidate formal specifications and clarification questions, while Bayesian optimization estimates uncertainty over user intent and selects queries that maximize information gain. The inferred STL specification is then passed to a formal planner to produce a verifiable robot trajectory. Across four simulated and real-world domains, the approach generally achieves higher task satisfaction with fewer clarification rounds than LLM baselines.
Huao Li, Carson Sobolewski, Augustinos Saravanos et al.
In briefSAGE: A Symbolic Gate and Local Editor for LLM Household Task Planning
SummaryThis paper presents SAGE, a single-LLM planner for household robots that combines a lightweight symbolic gate, which blocks actions violating preconditions before execution, with a local edit that regenerates only the failed sub-goal's remaining steps. The authors evaluate it on a 75-task AI2-THOR benchmark using five open-weight models under a leak-free protocol. They report results on the standard benchmark, a harder multi-goal composition task, and settings with injected mid-execution failures.
In briefZeroCode: Reinforcement-learning method that builds error-correcting code parity-check matrices from a zero matrix on demand
SummaryThe paper proposes ZeroCode, a reinforcement learning approach that sequentially builds parity-check matrices for error-correcting codes starting from an all-zero matrix, to meet varied application requirements. For the (32,16) code at a bit error rate of 10^-4, it gains about 1 dB over the prior RL-based method and outperforms genetic, differentiable, and classical design methods in its experiments. Action masking allows structural constraints such as maximum degree, 4-cycle-free, and quasi-cyclic structure to be added.
In briefA new optimizer for low-rank adaptation that respects factorization symmetry
SummaryThis paper proposes a new optimizer for low-rank adaptation (LoRA) that explicitly accounts for the gauge symmetry of low-rank factorization. It extends recent full-parameter matrix optimizers to the fixed-rank matrix manifold by treating them as normalization in a rotated basis. The authors show how rotation and normalization can be combined with the fixed-rank manifold efficiently. The optimizer reportedly reaches lower held-out loss faster and performs better or comparably on supervised finetuning and reinforcement learning tasks.
In briefKeeping Benchmarks Valid as AI Agents Improve: Detecting and Fixing Unearned Passes
SummaryThe paper presents a process-verification framework that audits passing agent trajectories, separates evidenced reward hacking from verifier weakness, and locates exploitable benchmark surfaces for repair. Across 3,810 passing trajectories from 29 model-benchmark cohorts, confirmed violations often rise with model generation but not monotonically. Repair case studies show that blocking a recorded exploit is insufficient, so the authors combine minimal patches, exploit replay, and fresh agent evaluation.
In briefRoboICL: In-context robot control with GPT-6 Astra, without robot-specific training
SummaryThis paper introduces RoboICL, a framework that lets a general vision-language model control robots through in-context learning, without robot-specific parameter updates or a learned VLA. It keeps demonstration examples separate from the model's own action-outcome memory, using bounded anchored memory to retain experience across task stages. Across 30 RoboDojo tasks it improves on official zero-shot GPT-6 Astra, and on three real-robot tasks mean progress rises as demonstrations are added.
In briefQuantaSpike: Short-Window Spike-Driven Quantization for Large Language Models
SummaryThis paper proposes QuantaSpike, a spike-driven quantization framework for large language model inference that uses short firing windows. It is built on Logarithmic Ternary Integrate-and-Fire (LTIF) neurons and handles outlier activations with group-adaptive gain and selective outlier admission. On OPT and Llama-2, it reports perplexity and zero-shot accuracy that are state-of-the-art or competitive among spike-driven methods, and it stays close to FP16 on Llama-3-8B and Qwen3-8B with a four-step window. Analytical projections estimate about 80.0% energy reduction on OPT and 67.1% on Llama-2 for one linear transformation relative to SpikeQuant.
In briefExamining Human-AI Differences Through Multiple-Solution Problems
SummaryThis paper asks whether frontier AI models and humans produce different distributions of answers on reasoning puzzles that have several valid solutions. The testbed has 270 puzzles across five families, each with 3 to 8 valid solutions. Model distributions resemble each other far more than human ones and are less diverse within every puzzle family.
In briefDegradation of language models trained recursively on their own outputs in biomedical question answering
SummaryThis paper studies what happens when language models are repeatedly trained on data they generated themselves, using biomedical question answering on PubMedQA with Qwen2.5 models of 0.5B and 3B parameters. It compares a recursive synthetic-data condition against a condition that reuses the original human training data, across four generations and two random seeds. The recursive condition showed larger declines in disease and chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity, with answers becoming longer, and the effect was larger for the 3B model. The authors note these results do not establish clinical hallucination rates or universal model collapse.
In briefDreamingGoose: Staged Distillation from Autoregressive Transformers to Bidirectional Recurrent Diffusion Language Models
SummaryThe paper presents DreamingGoose, which converts Qwen3 autoregressive teacher models at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially, while in-context retrieval does not transfer, with converted students scoring 0.000 on a multi-query recall probe where teachers scored 0.34 to 0.58. A retrieval curriculum that gradually lengthens the key-value gap restores retrieval only stochastically under a fixed schedule, but advancing the gap based on a running accuracy threshold succeeds for all three seeds and carries to 8B, where two of three seeds succeed.
In briefWAM-OPD: On-policy distillation to sharpen robot world action models
SummaryThis paper proposes WAM-OPD, which applies on-policy distillation to pretrained world action models so they reach expert-level performance on target robotic manipulation tasks. To avoid repeated environment rollouts, it introduces prefix-weighted trajectory replay (PWTR), which reuses a fixed trajectory pool and reweights distillation losses. Simulated and real-robot experiments show improved target-task performance while tasks excluded from adaptation stay near their initial performance.
In briefAdapting Reward Rubrics to Shifting Support Needs in Empathetic AI Dialogue
SummaryThe paper argues that empathetic reinforcement learning uses fixed reward specifications, even though the priorities for support change with the dialogue. It proposes CARE, which generates a context-adaptive rubric each turn by adjusting the weights and criteria across cognitive, affective, and proactive empathy. Integrated with RLVER and MICA, CARE reports state-of-the-art results on SentientBench, EQBench3, and EMPA, improving EPM-Idx on EMPA by at least 13 points over the strongest baseline under all three judges.
In briefSALMONN-duo: A voice agent pairing a fast real-time speaker with a slower reasoning backend
SummarySALMONN-duo is an adaptive voice agent with two parts: a fast, always-on full-duplex speech model for real-time conversation, and a slower asynchronous LLM agent for tool use and deliberate reasoning. The fast system learns when to answer directly and when to delegate to the slow system, integrating returned results into the dialogue. Experiments on spoken question answering and multi-turn conversation show that adaptive delegation improves accuracy on knowledge-intensive and multi-hop questions, and cost-aware reinforcement learning further balances performance against backend usage.
In briefCasEm: A Cascade Architecture for Long-Horizon Neural Emulation
SummaryAutoregressive neural emulators can drift or diverge over long rollouts even when short-term predictions are accurate. The paper introduces Cascaded Emulation (CasEm), a one-way architecture that adds an independently evolving model of physically defined aggregates to an existing full-state backbone, using its forecasts to correct full-state predictions without feedback to the aggregate model. Across four ODE/PDE benchmarks, CasEm reduces long-horizon rollout errors. In global climate emulation, adding a regional total-water subsystem cuts 10-year full-state time-mean error by 66.6% for frozen ACE and 46.3% for Spherical DYffusion, with under 3% added inference time.
In briefStashbird: a conversational agent memory system that keeps derived state consistent through updates and deletions.
SummaryThe paper presents Stashbird, an episode-grounded memory system that uses provenance to reconcile derived graph and retrieval state during updates and deletion. Repeated evidence extends support, corrections supersede prior active relations, and deletion removes unsupported state. Under fixed semantic decisions, reconciliation produced correct final states in all 20 update sequences and satisfied the checks across 263 deletion interventions. Across four long-term memory benchmarks, it reports competitive question-answering accuracy and lower token use than Graphiti and Hindsight on LoCoMo.
Chidera Biringa, Lucas Yannul, Xiaowen Wang et al.
In briefAdaGuard: A Safety Guard Model Adapting to User-Defined Rules
SummaryThis paper presents AdaGuard, a guard model that judges whether language model agent behavior violates rules supplied by users at inference time, rather than a fixed risk taxonomy. The authors build AdaptiveSafety, a dataset of 10,939 training and 1,000 test examples with policies of 1 to 100 rules, and propose SafePO, a reinforcement learning method for refining violation identification. A 4B model reaches 89.30% binary accuracy on AdaptiveSafety and 71.82% on DynaBench.
In briefCoherence-aware distributional metric (CHORD) for evaluating open-ended text generation
SummaryThe paper argues that existing open-ended generation metrics can miss global coherence failures, such as passages that are locally fluent but globally contradictory. It introduces CHORD, which encodes generated and human texts in the hidden states of a frozen LLM using a coherence-eliciting prompt and compares them with RBF-MMD. On a counterfactual evaluation suite, CHORD selectively detects relation, discourse, structural, and mixture failures, and its model rankings align strongly with human judgments.
WhyProposes a frozen-LLM hidden-state RBF-MMD metric for global coherence, validated with a counterfactual suite separating coherence failures from benign rewrites.
In briefDecFlowEdit: Background-preserving image editing that decouples guidance settings for localization and editing
SummaryThe paper finds that FlowEdit's default classifier-free guidance setup, with asymmetric source and target scales, causes substantial background leakage during edits. It proposes DecFlowEdit, which separates the guidance scales best suited for localization from those best suited for editing, extracting an edit-relevant prior from guidance-free velocity differences to reweight updates. The method is training-free and inversion-free, without external masks or attention manipulation. On PIE-Bench with FLUX, SD3, and SD3.5, it reduces structure distance by about 61 to 73 percent and background LPIPS by 68 to 80 percent relative to FlowEdit at comparable editing fidelity.
In briefSegBanana: Adapting Unified Multimodal Models for Training-Free Medical Image Segmentation
SummaryThis paper proposes SegBanana, an agentic visual generation framework for medical image segmentation that requires no training. It uses a frozen unified multimodal model as the core generator, enhanced with anatomy-aware knowledge retrieval and comparative quality critique, and a state-aware controller that iteratively refines predictions. Across eight medical segmentation datasets, it achieves an average mDice of 77.45%, outperforming the listed generalist and medical-specific baselines.
In briefMAS-OPD: On-Policy Distillation for Multi-Agent Systems
SummaryThis paper addresses how to jointly post-train multi-agent systems, where several specialized roles split a task. It proposes MAS-OPD, which applies on-policy distillation so a teacher gives token-level supervision on trajectories sampled by the student, using role-advantage specialization and privileged attribution of cross-agent conflicts. Experiments on code and mathematics benchmarks report the highest mean score at both student scales, with agents developing clearer role specialization and more effective collaboration.
In briefTrustworthy synthetic visual media: evidence across the media lifecycle
SummaryThis review examines how to judge whether images and videos can be trusted when generative systems can create or alter realistic media without a stable record. It argues that trust must come from evidence about an asset's path and use, some recoverable from the media itself and some recorded during production and preserved as it circulates. The authors contend there is no single universal marker of authenticity, so evidence should fit the question, reach the person judging, and be open to correction.
In briefReGDiff: Guided Diffusion in Regulated Latent Space for Generating Metamaterial Voxel Shapes
SummaryThis paper introduces ReGDiff, a generative framework that combines voxel representation, latent space regulation, and guided diffusion to generate metamaterial geometries. It uses a repel-and-sink mechanism to smooth the latent distribution of plausible shapes and short-range repulsion guidance to avoid outputs too close to known samples. The authors also release a voxel benchmark for truss and shell metamaterials with an evaluation module, reporting average gains of 8.9% in plausibility, 46.4% in novelty, and 128.6% in diversity over voxel-based baselines.
In briefSleuthBench: a benchmark testing LLM statistical analysis using hidden signals in tables
SummarySleuthBench injects controlled data-quality problems and feature effects into public tabular datasets, so reference answers can be computed automatically and verified. Six state-of-the-art LLMs were tested on 1680 graded responses using a Python tool. Models detected data-quality problems with 83.8% accuracy but identified feature contributions with only 41.9% accuracy. Providing precomputed statistical artifacts, called the Empirical Layer, raised feature-contribution accuracy to 68.0%.
Jingyun Jia, Antoine Remond-Tiedrez, Aaron Alvarez et al.
In briefWhen Does Selection Replace Extraction? A Pre-Registered Agent Memory Test
SummaryThis paper asks whether conversational memory needs LLM-extracted facts or whether selecting the right raw dialogue turns is enough. In a pre-registered study on held-out LoCoMo and LongMemEval conversations, raw turns chosen by one call to Jev, a typed decision model, were non-inferior to an LLM-extraction memory at a tight budget, at far lower write cost. The authors also find that reranking's benefit shrinks as the budget grows.
WhyA pre-registered comparison of raw turn selection against LLM extraction yields a fresh finding, with solid evidence on LoCoMo and LongMemEval and useful guidance for conversational memory design.
In briefUSA: Update-aware SAM for cross-domain on-policy distillation of language agents
SummaryThis paper studies on-policy distillation for language agents, noting that a single domain saturates early and that merging task vectors from several domains can cause negative transfer. The authors attribute this to cross-domain update coupling, where both domains update many parameters similarly. They propose USA, which sets per-coordinate perturbation radii from update magnitudes measured in a brief warm-up. Experiments on mathematics, science and code at two student scales report that USA performs best in all six transfer directions and reverses negative transfer for conflicting pairs.
In briefOrdinal-Matching Bias in Audio-Visual LLMs Explained
SummaryThis paper studies whether audio-visual large language models can match speech to the correct visible speaker in scenes with several people. Using a synthetic diagnostic dataset, the authors find that three open-source models tend to pair utterances with faces by their left-to-right, top-to-bottom order instead of audio-visual cues like lip sync, which they call ordinal-matching bias. They propose Ordinal-Decoupled Fine-Tuning, trained on 400 synthetic videos with randomized speaker positions and speaking order, which reduces this bias and improves performance on real-world audio-visual benchmarks.
In briefWorldWeave: Building Expandable 3D Worlds for Consistent Video Generation
SummaryWorldWeave is a framework that separates maintaining a world's structure from rendering its visuals. It uses diffusion-based outpainting to generate continuous elevation maps and an agent to organize and stitch them into an expandable explicit 3D world. During video generation, camera trajectories only read this geometry to produce depth sequences, without writing results back, so structure stays consistent across viewpoints and repeated visits.
In briefmmHRI: Privacy-preserving human-robot interaction using millimeter-wave radar
SummaryThis paper presents mmHRI, a framework that uses millimeter-wave radar to sense human actions and 3D poses without capturing identifiable images. The estimated states are converted into textual instructions that guide a vision-language-action policy for delivery and retrieval tasks. Under a privacy curtain setting, it reports 85.09% action-recognition accuracy and stable task performance under visual occlusion.
In briefLoop Dropout: Regularizing Shared Updates in Looped Language Models
SummaryThis paper studies how to adapt looped language models, which repeatedly apply the same transformer block. The authors find that standard LoRA gives limited adaptation at early loop positions. They propose Loop Dropout, which randomly masks adapter applications and rescales them to keep the expected update strength, improving adaptation across loops. Experiments report gains in mathematical reasoning, instruction tuning and code generation, with inference using standard LoRA and no extra trainable parameters.
In briefExplainable, generalisable LLM-based detection of cognitive decline from spontaneous speech
SummaryThe study proposes a bilingual speech large language model that processes raw speech directly to learn acoustic-semantic representations for automated cognitive screening of Alzheimer's disease and mild cognitive impairment. Using a newly collected PUTH-AD dataset and open-source corpora, it achieved the highest average accuracy and AUROC across six dataset/task conditions versus three baselines. It also maintained classification accuracy on an unseen cognitive task without task-specific fine-tuning, and clinicians judged its natural language explanations clinically relevant and largely consistent with the speech evidence.
In briefSame Winners, Different Success Rates: How LLM Agents Recover from Failures
SummaryThis paper argues that existing checkpoint-based benchmarks measure LLM agent recovery only through set agreement, an ordinal measure that ignores absolute success levels and can look stable when all actions fail. The authors prove that different success probabilities can yield identical best-action-set distributions, so pooled success is needed to resolve this ambiguity. Experiments on 864 RecoveryBench episodes and two planning cohorts totaling 3,456 responses support these predictions. The authors recommend reporting four diagnostics: agreement, all-zero fraction, held-out success, and pooled success.
In briefGlyphBench: A reinforcement learning testbed for language-model agents using Unicode grid games
SummaryGlyphBench is an environment suite with over 360 tasks across diverse games, designed for reinforcement learning post-training of language-model agents. It renders spatial observations as two-dimensional Unicode grids and provides a unified interface for training, evaluation, and trajectory replay. In experiments, glyph observations outperformed native text and pixels on Craftax, and RL on 100 tasks lifted Qwen3.5-4B to 63.48% accuracy on held-out Reasoning Gym problems, above the base, math-trained, and code-trained baselines.
Roger Creus Castanyer, Marc-Alexandre C\^ot\'e, Matthew James Sargent et al.
In briefX-MoD: Practical scaling laws for sparse-depth routing beyond Mixture-of-Depths
SummaryThe paper introduces X-MoD, a scalable sparse-depth Transformer that decouples token sparsity from anchor stride, so total parameters can grow while active-equivalent capacity stays nearly fixed. To train deep sparse routing, it combines dense anchors, variance-scaled layer-wise gating, and depth-wise token balancing. The authors also frame choosing a routing configuration as a design problem and fit an interpretable scaling law, relative to FLOP-matched dense baselines, that predicts validation loss.
In briefBehavior-grounded semantic enrichment for modeling and reasoning about financial fraud
SummaryPublic real-world financial datasets often lack rich semantics because of privacy constraints, and synthetic alternatives sacrifice behavioral realism. This paper proposes a multi-agent semantic enrichment framework that generates interpretable financial semantics grounded in original transaction behavior, with consistency refinement. It also releases MS-FFSD, a multimodal fraud dataset with structured and textual semantics. Results show statistical fidelity and generalizability, and that richer semantics benefit fraud modeling and context-aware LLM reasoning.
In briefAmplitude-phase functional autoencoders: unsupervised learning that separates shape from timing in functional data
SummaryThe paper proposes Amplitude-Phase Functional Autoencoders (AP-FAE), an unsupervised framework for univariate and multivariate functional data. It factorizes the latent space into amplitude and phase embeddings, using a shared monotone, endpoint-preserving warp to capture phase variation. Across synthetic data and six real-world benchmarks, the authors report that AP-FAE outperforms state-of-the-art baselines on most clustering and alignment metrics and on all reconstruction metrics.
In briefWorldGuide: Learning success and failure boundaries in latent world models to improve robot vision-language-action policies
SummaryThis paper proposes WorldGuide, a framework that learns to tell successful from failed robot interactions in a latent world model. It combines predictive pretraining on success and failure trajectories with contrastive learning on matched pairs, then uses the learned predictor as a differentiable reward for training the policy and visual encoder. The predictor is discarded after training, so deployment needs no extra world-model inference. The authors report 96.8% on LIBERO 100 and 72.0% on SimplerEnv.
In briefTraining LLMs to model optimization problems by self-play with a solver
SummaryThis paper proposes OPT-Zero, a self-play framework for optimization modeling that uses no external training data. A single LLM plays two roles: a Proposer that generates optimization problems with their formulations and solving code, and a Solver that answers from natural-language descriptions alone. Both roles are trained alternately with reinforcement learning, using feedback from external optimization solvers. The authors report that it matches state-of-the-art data-dependent methods while generalizing substantially better.
In briefFixed Judges, Changing Agents: Version-Dependent LLM-Judge Errors in Agent Evaluation
SummaryThe paper studies whether language-model judges can fairly compare new and old versions of AI agents, finding that a fixed judge can make version-dependent errors. The authors analyze 35 public coding-agent submissions, two customer-service agents, and expert-labeled trajectories, and find that most judges fail a task-conditioned error invariance test and sometimes accept upgrades that execution-based results do not support. They conclude that paired audits of current outputs are preferable to judge-only release decisions or transported calibration.
WhyShows LLM judges make version-dependent errors across benchmarks with multiplicity-adjusted tests; moderate novelty, solid evidence, and practical relevance to agent evaluation.
In briefConvCue: Adding Convolutional Visual Features to Vision-Language Models
SummaryThis paper introduces ConvCue, which adds final-stage features from a frozen, pretrained convolutional network to an existing vision-language model without replacing its native visual encoder. A learnable adapter and gated cross-attention fuse these features with the original visual tokens, trained in two stages. Across 13 multimodal benchmarks on Qwen3-VL-2B, Qwen3-VL-4B, and LLaVA-OneVision-7B, the authors report higher average performance than both the original models and matched fine-tuning controls. On Qwen3-VL-4B, the average rises from 75.00 to 78.82 relative to the matched control.
In briefPainterBench: A Benchmark for Creative Drawing by Tool-Using Language Models
SummaryPainterBench adapts the human incomplete-drawing test into an agent task, where models draw on a canvas via tool calls and must build an original picture around an uneraseable starting shape. The authors evaluated 14 multimodal models, collecting 2,700 drawings and human creativity and recognizability ratings. An automated scorer, ViDrA-adapted, predicts human creativity ratings with r = 0.85. Performance varies widely across models, and agent drawings rate higher in creativity but lower in recognizability than human drawings.
In briefDecomposing basket-size-stratified associations to reveal pairwise and triple item structure
SummaryThis paper introduces CSID, a framework that splits log-odds contrasts, stratified by the number of remaining items in a basket, into item-set-specific and cardinality-common components without fitting a global joint distribution. In a simulation with zero pair effects, ordinary Ising couplings became spuriously negative under small-basket cardinality potentials, while CSID pair estimates stayed near zero. Across three grocery datasets, high-information triple components were reproducible over time and transferred better across periods than a comparator method.
In briefMotionSpaceFlow: Flow-based text-to-human-motion generation directly in motion space
SummaryThe paper introduces MotionSpaceFlow (MSFlow), a flow-matching framework that predicts clean human motion directly in continuous motion space without a learned encoder or decoder. It proposes representation-aware noise scaling and a RA-MMDiT transformer that handles frame-change features and absolute joint features with different attention schemes. The authors report state-of-the-art text-to-motion results, and their global representation variant supports zero-shot joint or frame control at inference via projection sampling.
In briefAlphaPareto: LLM-guided multi-objective reinforcement learning for formulaic alpha discovery
SummaryThis paper proposes AlphaPareto, a reinforcement learning method for discovering formulaic alphas used in quantitative trading. It addresses the non-stationarity caused by an evolving alpha pool by including the pool in the state, encoded with a large language model, and replaces the single scalar reward with a multi-objective vector reward covering predictive power, temporal stability, perturbation robustness, and diversity. These objectives are optimized through a Pareto-regularized learning procedure. The authors report that AlphaPareto outperforms competing methods in empirical tests on real-world datasets.
In briefLLMs are not stochastic parrots: evidence for meaning-based abstraction
SummaryThis paper tests the strong 'stochastic parrot' claim that large language models can only do statistical pattern matching and cannot abstract or reason. The authors give several LLMs only natural-language descriptions of invented languages that break common training-data patterns, with no example outputs. The models often followed the rules, adjusted semantic relations to new constraints, and sometimes exactly matched complex translation answer keys, which the authors say refutes the strong hypothesis.
Julia Witte Zimmerman, Calla G. Beauregard, Tabia Tanzin Prama et al.
In briefEntroPack: Fast, accurate neural network weight compression at any bitrate
SummaryEntroPack is a method for compressing neural network weights to any target bitrate without calibration or fine-tuning. It combines row-normalized E8 lattice quantization with a conditional probability model and uses independently decodable entropy-coded tiles for fast GPU decoding. On the Z-Image-Turbo image generator at 4 bits per parameter, it reports about 24% lower relative L2 weight error than NF4 while using less storage.
In briefCASS: Structured sparsity to reduce interference when merging AI models
SummaryModel merging combines separately fine-tuned models into one multi-task model, but conflicting task-vector updates cause parameter interference. The paper proposes CASS, which identifies and preserves task-relevant attention heads and FFN neurons using a contribution-aware structured mask. It offers CASS-Merging as a plug-and-play filter for existing merging methods and CASS-Tuning for settings with fine-tuning access. The authors report that CASS improves several representative merging baselines across vision and language benchmarks.
In briefCRF Loss: How Networks Should Learn Boundaries in Weakly Supervised Segmentation
SummaryThis paper addresses boundary accuracy in weakly supervised semantic segmentation, which learns pixel-level predictions from image-level tags. The authors argue that DenseCRF post-processing can flip correct labels because it relies on low-level colour cues, and that using SAM to produce one-hot pseudo-labels discards the uncertainty in CAMs. They propose DS-CRF, which treats soft pseudo-labels as unary supervision and binary edge maps as pairwise supervision, each with its own loss. DS-CRF reports a new state-of-the-art of 56.5% mIoU on MS COCO.
In briefLearning dexterous robot manipulation from scalable human touch demonstrations
SummaryThe paper addresses the scarcity of tactile data for robot dexterous manipulation, noting that teleoperation gives operators little tactile feedback while human demonstrations are easier to scale. The authors built a tactile motion-capture system and released the UVTA dataset, with 1,000 human and 150 robot demonstrations per task across five contact-rich tasks. Their Unified Visual-Tactile-Action Model aligns both embodiments and jointly predicts future actions and tactile trajectories, achieving a 70% average real-robot success rate versus 29% for the strongest visual-tactile baseline and 42% for an architecture ablation.
In briefEfficient Reasoning via Constrained Optimization in Latent Space
SummaryLarge reasoning models often overthink, producing redundant steps that consume many tokens. The paper finds that efficient reasoning steps cluster in a concentrated latent-space region, and uses a quadratic program to project deviating hidden states back into it, in a training-free framework. Tested on four models from 1.5B to 14B across six math, coding, and scientific QA benchmarks, it reports accuracy gains of up to 12.1% and token reductions of 11.8% to 52.8%.
WhyProjecting hidden states into an efficient-reasoning region via quadratic programming is a substantive improvement; validated on four models and six benchmarks, influence limited to reasoning efficien
In briefComparing decision readouts for text-based video anomaly detection
SummaryThe study compares Jev typed decisions with three Qwen readouts, holding video-derived textual evidence fixed, on 40 videos and 400 target anchors from UCF-Crime and XD-Violence. On XD captions, Jev Noul reached higher average precision than the Qwen readouts, but UCF showed no corresponding advantage. The authors describe these exploratory offline results as characterizing ranking, probability quality and interface failures, not establishing causal benefit or general superiority.
In briefRAGWarrant: Auditable Governance for Deciding Whether to Release RAG Policy Changes
SummaryThis paper presents RAGWarrant, an open-source framework that decides whether a proposed change to a retrieval-augmented generation system should be released. It checks quality and hard-risk gates declared in advance, applies evidence-based claim limits, keeps negative results, and outputs PROMOTE, BLOCK, REJECT, or INCONCLUSIVE decisions. The authors evaluate it on several datasets and state that the contribution is an auditable promotion-control abstraction, not proof of optimizer superiority or production readiness.
