News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 33 of 54 · The Research front page
Leibovici et al. introduce CANTO, a transformer neural operator that maps continuous CAD geometry to physical fields without meshing. They tokenize NURBS patch parameters directly, achieving state-of-the-art accuracy on four aerodynamic benchmarks and enabling gradient-based inverse design that finds lower-drag car shapes.
The authors introduce JudgeProfile, a framework that decomposes LLM evaluation into perception (how a judge compares responses on specific qualities like clarity and correctness) and prioritization (how much each quality influences the final choice). They curate SubjectiveSet, a dataset of 50,013 response pairs from 17 sources evaluated by 21 LLMs across 87 attributes. The key finding is that judges often agree on attribute-level judgments even when their overall choices diverge, and reweighting attribute importance improves agreement with reference labels from 66.48% to 71.97% without updating the model.
The authors present a framework for improving fine-grained limb-motion captioning in Vision-Language Models (VLMs). They introduce FlexBench, a benchmark of 3,105 shots and 18,161 queries with human-verified person identities and per-shot limb action checklists, and the Graded Physical Alignment score (GPA), which awards credit for correct content and deducts for incorrect or fabricated actions. Building on this rubric, their Graded Margin Direct Preference Optimization (GM-DPO) assigns stronger preference margins and training weight to more severe action errors, achieving the highest GPA scores across three VLM backbones and reducing hallucination rates relative to standard DPO.
The authors show that pretrained transformers underuse their depth for in-context reference following, with thirteen models reliably tracking only 2-3 lines. A rank-8 LoRA applied to a single early layer, with all other weights frozen, dramatically extends this reach: Qwen3-8B reaches 99% accuracy on 24-line chains, and Ouro-1.4B follows up to 160 lines with loops.
The authors propose SCOPE, a deterministic single-pass model that reconstructs complete PDE fields from sparse measurements by coupling full-target latent prediction with a shared physical decoder. Across five PDE settings, SCOPE outperformed mask-aware neural operators on all ten tasks and reported lower errors than diffusion-based solvers DiffusionPDE and FunDPS. A quadratic risk decomposition shows why optimal latent prediction need not yield optimal field reconstruction, and a separate bound characterizes when decoder improvements transfer.
Xu and colleagues introduce PDE-OBS, a benchmark for testing how scientific machine-learning models respond to changes in measurement density and spatial layout. In 3,969 evaluations, every PDE-method pairing had higher average error under mismatched training and test patterns than under matched patterns, and simply providing denser test observations did not reliably improve a fixed model.
The authors identify an information-credit gap in agentic reinforcement learning (ARL) for LLMs: when failures arise from poor retrieval, the fixed retriever escapes blame while the LLM policy is penalized. They formulate retrieval-augmented ARL as a bilevel optimization problem, then introduce BRIDGE, a memory-efficient first-order method that updates a low-rank retriever adapter using both RL credit and a RAG likelihood term, followed by LLM policy updates. Across seven open-domain QA benchmarks, BRIDGE achieves the highest average accuracy with both 3B and 7B backbones, improving the multi-hop average over the strongest baseline by 9.6 and 3.4 exact-match (EM) points respectively.
The authors propose Tandem Token WaterMark (TTMark), a distortion-free watermarking framework that operates on adjacent token pairs rather than individual tokens. By watermarking the joint distribution of consecutive tokens, TTMark expands the effective watermarking alphabet from V to V², exploiting both token entropy and conditional entropy. The method consistently improves detection rates across models and datasets without degrading generation quality.
The authors demonstrate that LLMs can learn to evade latent space monitors by observing the monitor's binary feedback on prior outputs. Off-the-shelf models already produce activation edits aligned with the monitored direction, and scaling these edits by a factor of 8 reduced the true positive rate from 100% to 27%. A rank-1 LoRA finetuning amplified this into effective forward-pass evasion, reducing TPR to 4% on held-out concept monitors while preserving other capabilities.
The authors present Merlin Plus, a public CT dataset containing radiologist-created tumor masks for 9 organs that previously lacked such annotations. Using a report-based active learning framework, they generated initial masks from radiology reports, which radiologists reviewed and corrected, reducing annotation time by 3-5 times. The dataset enables multi-tumor segmentation and detection models that outperform existing public models.
The authors introduce CheatBench, a benchmark to measure whether AI agents violate expectations of honest work in mathematical research, coding, knowledge work, and visual tasks. Evaluating several frontier models, they find that agents often cheat when completing tasks becomes difficult, raising concerns for safety as agents take on more responsibilities.
The authors propose HeurEvo, an automated framework that co-evolves algorithmic plans, executable code, and a shared component pool to generate hybrid optimization algorithms. Evaluated on combinatorial optimization, MIPLIB-NL, and nonlinear geometry problems under a two-minute execution budget, HeurEvo often matches or surpasses solvers running for hours or days, and on some geometry problems it improves previously reported best results.
Abbasi, Shi and Kolouri address a mismatch in mixture-of-experts serving: each token activates few experts, but a batch of tokens can collectively require weights from much of the expert pool. Their method, BASE, predicts the output error from omitting each expert and uses a fixed batch-level budget to improve throughput while preserving more model quality than existing selection methods.
The authors propose MoRE, which replaces the standard linear router in mixture-of-experts layers with a low-rank factorization, reducing per-token cost from Θ(Mh) to O((h+M)r). They prove that rank logarithmic in the number of experts suffices for routing expressivity and load balancing, and demonstrate empirically that MoRE improves memorization and knowledge-intensive QA while matching reasoning ability at equal active FLOPs, with a fused kernel for efficient inference.
The authors introduce SCOUT, a two-stage agentic safety verifier for computer-use agents (CUAs). In the first stage, a rubric generator reasons over the task and trajectory to produce completion and safety rubrics without using tools. In the second stage, a probing agent follows these rubrics to interact with the post-execution environment and gather grounded evidence. On AutoElicit-Bench, SCOUT achieves 75.4 unsafe F1 and 74.5 completion F1, and its feedback reduces unsafe execution from 30.2% to 17.2% through test-time reflection.
The authors introduce Structured Matrix Attention (SMat-Attention), a family of causal attention masks parametrized by VC dimension d. For d ≤ 3, prefill is O(T); for d > 3, it is O(T^{2-3/d}). After a fixed prefix, decoding takes constant time per token using O(T^{1-1/d}) cached states. The approach is extended to Mamba-2 and Gated DeltaNet, improving recall accuracy while remaining competitive in language modeling.
Xu and Liu present the first systematic formalization of multi-agent Vision-and-Language Navigation (VLN) as a coordination problem with subtask dependencies, resource constraints, and collision avoidance. They introduce the MAVLN benchmark with 11,724 episodes across 145 scenes for teams of two to four agents, and propose TRISS, a system combining an LLM-based scheduler, shared topological memory, and conflict-aware execution. Experiments show that while TRISS provides a baseline, significant room for improvement remains in scheduling, planning, and execution.
The authors propose GeoVerse, a framework that synthesizes world-consistent novel views by combining a geometric latent diffusion model with video generative priors from Wan2.2 VACE. Using a ControlNet-style adapter and a global point-cloud memory for cross-view coherence, the method achieves a 2.23 dB higher PSNR on DL3DV and 32.4% lower Absolute Trajectory Error on Mip-NeRF360 compared to the prior state-of-the-art GLD, while reducing inference time to under 10 seconds via 4-step reflow distillation.
The authors introduce HEAR, a large-scale benchmark of 87,105 real human audio samples from 843 demographically diverse participants, designed to measure demographic bias in Audio-LLMs. Evaluating both speech-to-text and real-time speech-to-speech models, they find that voice-conditioned bias varies by model and that personalization instructions exacerbate demographic disparities, establishing voice bias as a controllable characteristic.
Cheng and Zhang trained five versions of the same GPTNeoX backbone from scratch, varying only how each model performed intermediate computation. Although all variants handled problems within the training range, only the latent-reasoning models generalized strongly to longer reasoning chains, with mechanistic evidence indicating that recurrence implemented forward search over a graph.
The authors introduce a sensitivity-aware approach to residual-stream pruning for large language models. Instead of minimizing activation reconstruction error, they derive a tractable upper bound on the KL divergence between original and pruned model outputs, which balances activation covariance with output sensitivity. Across multiple instruction-tuned LLMs, their method consistently outperforms activation-only pruning, especially at higher compression rates.
Li et al. introduce Verifiable Visual Rewards (VVR), a framework where image generation tasks are scored by deterministic program verifiers rather than learned detectors or vision-language models. Training with these rewards raises Stable Diffusion 3.5 Medium's accuracy on the VVRBench benchmark from 2.8% to 28.3%, and the gains transfer to out-of-domain natural prompts.
The authors propose FactorEngram, a memory module for language models that replaces monolithic n-gram embeddings with sparse coefficients over a shared dictionary. Contextual gating at the basis-vector level allows the model to selectively read out relevant components, improving language modeling and downstream performance in 340M- and 1B-parameter Transformer backbones.
Malandrino presents Tetra, a representation and serving kernel that makes two-bit Leech-lattice quantization practical without first expanding its enormous codebook into a wider GPU format. Complete Qwen3-4B, 8B and 14B models occupy about 2.7 bits per parameter, roughly half the reported size of 4-bit AWQ files, but remain less accurate than AWQ on MMLU.
The authors introduce Almieyar, a zero-shot speech recognition benchmark covering 17 Arabic dialects from six dialect families, with all audio newly recorded for the task. Benchmarking 12 systems, they find that GPT-4o-transcribe performs best at 35.0% word error rate, and no model is uniformly best across dialect groups.
Zhang et al. identify a privacy risk where LLMs, after being asked to remove an item such as a password, may state the removal in a comment that reveals the item. They analyze real conversations and build a benchmark, finding that across six models about 13% of deliverables allow recovery of the withdrawn item from revision traces, and propose an output-side filter that sharply reduces recovery.
The authors propose AnswerPool, a method that converts grouped multiple-choice benchmarks into harder tasks by pooling all answer options from questions sharing a context. Across eight benchmarks and eighteen models, every model scores lower under pooling, and the accuracy loss reveals how much the multiple-choice format helped via elimination. When correct answers are deleted from the pool, most open-weight models answered 87–100% of unanswerable questions, exposing a critical failure to abstain.
The authors analyze the training dynamics of score-based generative models on multimodal Gaussian mixtures. They decompose the time-integrated loss and find that the signal-to-noise ratio at a given training time dictates which features are learned: at high SNR only mode directions are acquired, while near the speciation time (SNR ~ 1) both directions and relative weights become learnable. The weighting schedule determines how much of the training loss is concentrated near that critical region, thereby governing the learning timescales of different features.
Xie and colleagues introduce AIM-ZO, a zeroth-order fine-tuning method that estimates model updates using only forward loss evaluations. Across five language models and 11 tasks, it generally outperformed prior forward-only methods under matched evaluation budgets, including gains on models with 2.7 billion and 30 billion parameters. The results suggest that carefully maintaining perturbation directions can make forward-only optimization more effective without sacrificing its low-memory premise.
Zihan Teng, Jiayu Zhao, Wentao Ren, Minhao Fan, Tianrui Ma, Song Chen, and Weichen Liu propose SlimKV, a KV-cache compression method that jointly reduces the number of cached states and their per-token dimensions. Using low-rank-aware training, SlimKV compresses long contexts into compact beacon memory states with layer-adaptive rank allocation, and exploits an asymmetric dependence on key-side RoPE to remove its application from beacon keys during decoding. This achieves up to a 7.34x attention speedup and 3.38x end-to-end decoding speedup over the uncompressed model at 128K context length.
The authors introduce Gated Slot Attention-2 (GSA2), which combines Gated Oja Rule-2 for key-side correction and Gated Delta Rule-2 for value-side correction via shared latent slots. At 1.3B scale, GSA2 consistently outperforms strong linear-attention baselines on language modeling, commonsense reasoning, synthetic retrieval, and long-context tasks, with particularly clear gains on RULER and LongBench.
The authors introduce MFPG-PPO, an extension of the multi-fidelity policy gradient framework to actor-critic methods. By redesigning sampling, advantage estimation, and control variate construction, they preserve cross-fidelity correlation and prevent variance inflation. MFPG-PPO improves upon standard PPO trained on high-fidelity data alone across simulated locomotion tasks, matching the performance of PPO with 16x more data on the hardest task, and enables stable learning on a physical Franka arm with only four episodes per update.
The authors present Coco, an agentic platform for hardware-software co-design of ML accelerators, deployed with TPU architects. By storing every simulation sweep in a relational database and encoding recurring workflows as agentic tools, Coco reduces the manual archaeology of simulation results from hours to minutes, with early deployment showing reduction in time-to-simulation and time-to-insight.
Park, Lee, and Shim analyze why global safety signals fail in text-to-image models and propose CALM, which uses prompt-local counterfactual pairs to edit token representations. Across benchmarks, CALM improves unsafe content suppression while preserving benign utility better than global removal methods.
Liu et al. derive an exact factorization of KV cache eviction error into evicted attention mass and directional gap, motivating a unified approach to retention and compensation. They propose CORE, which distills offline set-level coverage into a lightweight cache-aware indexer that simultaneously drives retention ordering and parameterizes latent-memory writes. Across three backbones, CORE exceeds the strongest RULER baseline by up to 3.78 points at 90% compression on long-context benchmarks.
Monea and colleagues pretrained looped language models in which only the first pass writes a full key-value cache, while subsequent passes reuse it and retain only a short local window. Across models with 150 million to 1 billion parameters, this design reduced context memory and improved perplexity and downstream accuracy relative to parameter-matched looped Transformers.
Gurbuz and colleagues built a Linux environment that composes real applications while collecting screenshots, accessibility data, window geometry and action outcomes. Fine-tuning four vision-language models on 200,000 generated grounding examples improved accuracy across held-out conditions and five external benchmarks, with gains also carrying into longer computer-use tasks.
The authors propose iS-KV, an online low-rank KV-cache compression method for long chain-of-thought reasoning. It keeps recent tokens at full precision and folds older tokens into bounded-rank representations using block-incremental SVD, updating both the basis and historical coefficients to prevent drift. On two 8B models, iS-KV achieves accuracy within 1.0–6.6 points of the uncompressed model at 4–6x persistent-KV compression, outperforming token-eviction baselines at matched memory budgets.
The authors propose Test-Time Calibration Learning (TTCL), a method that jointly adapts an LLM's reasoning accuracy and confidence calibration to unseen target tasks without ground-truth labels. Across eight math and QA benchmarks, TTCL improves average relative accuracy by 40.13% and reduces expected calibration error by 70.80% on base models, and it also improves already calibrated models under domain shift.
Farhat, Desmond and Kyrillidis present GraphMemory, an external memory that lets frozen language models learn from successive interactions without updating model weights. Across two financial reasoning tasks, it retained performance close to or above a strong persistent-memory baseline while sharply reducing the tokens and time used to construct and query memory.
Zhao et al. present a continual distillation method for consolidating specialized vision foundation models into one shared encoder. At each stage, the model learns from one incoming teacher while a frozen copy of the previous student protects capabilities acquired earlier. The resulting backbone spans five capability families without requiring full joint retraining whenever a teacher is added.
Hatgis-Kessell et al. address social sycophancy, the tendency of language models to endorse a user's position even when that position may harm others. Their post-training method uses models to identify affected stakeholders, evaluate candidate replies from each perspective, and learn from those evaluations without human labels or a stronger teacher model. Across four datasets and four model families, it reduced inappropriate endorsement more consistently than six baseline methods.
Li et al. develop a theoretical AI debate protocol in which two debaters recursively reduce a complex claim to subclaims that a human can judge. For problems with sufficiently stable decompositions, they prove worst-case correctness, make honest and correct play a dominant-strategy equilibrium, and establish instance-wise optimality under black-box access to human judgments.
Luo et al. tested learned visual rewards in a closed loop, optimizing diffusion-based robot policies while independently measuring requested-task success and wrong-object failures using simulator state. On a drawer-opening task, optimizing Robometer increased success by 10.2 percentage points but also increased wrong-object failures by 10.9 points, a defect hidden by otherwise strong aggregate reward-model performance.
Samanta et al. introduce MIRA, a hierarchical agent that separates deciding what to investigate from carrying out each investigation. Across theorem proving, neural-architecture research and four autoresearch environments, the authors report that this structure improves inference, value forecasting and reinforcement learning of research-allocation decisions.
The authors present an end-to-end benchmark for testing whether text-based jailbreak prompts bypass large language model safeguards. Across eight open-weight systems, the unsafe-response rate rose from 11.08% on baseline prompts to 18.65% under jailbreak conditions, an average deterioration of 7.57 percentage points.
Wang et al. study whether an LLM-based optimizer can improve both an executor agent’s harness and the process used to optimize that harness, without changing either model’s weights. Their VERSE system adds execution-based checks, failure replay and regression tracking, improving four existing harness optimizers on held-out and newer software-repair tasks.
Yang et al. introduce Sigma, a blockwise language model that generates through continuous diffusion rather than predicting tokens autoregressively or denoising categorical states. Across math and coding evaluations, the authors report performance comparable to similarly sized discrete diffusion and autoregressive models, while continuous trajectories provide direct control over output quality, diversity and inference steps.