News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 38 of 54 · The Research front page
Agarwal et al. introduce BOReFT, a method that turns hidden-state interventions in a frozen language model into a compact continuous domain for black-box optimization. Across a semantic word-search benchmark and three molecular-design objectives, the method improved target discovery in Semantle and outperformed strong language-model baselines on two of the three molecular objectives.
YuE2 unifies symbolic and audio music generation by first writing an explicit score, then expanding it into semantic tokens and synthesizing full-song audio. Expert evaluations show that symbolic planning improves quality, and YuE2 achieves state-of-the-art results on benchmarks.
The authors propose Teacher Alignment to address the Gap Curse in reasoning distillation, where larger teacher models produce distributions too complex for smaller students to approximate. They develop TeacherGRPO, built on Group Relative Policy Optimization, which adapts the teacher via curriculum selective alignment and importance-adaptive length regularization. Experiments show consistent performance gains over existing baselines across diverse reasoning benchmarks.
A unified theoretical framework reveals that three prominent regression-based diffusion reinforcement learning methods—DiffusionNFT, FlowAWR, and RAM—are each solutions to a divergence-constrained reward-maximization problem, differing only in the convex generator defining the constraint. The authors identify approximations in prior work and propose DiffusionRFT, which uses an exact sparsemax projection, leading to faster convergence, more stable training, and top performance.
Goo and Hong prove that exact vanilla posterior sampling for reinforcement learning (PSRL) is minimax optimal in leading-order Bayesian regret for finite-horizon time-inhomogeneous tabular MDPs with unknown stochastic rewards, achieving the rate Õ(√(SAH³K)). They extend the same proof principle to linear-mixture MDPs, obtaining Õ(d√(H³K)). The key technical advance is a method to decouple the posterior-sampled model from its own continuation value using a common empirical transition reference and a Bellman-based variance argument.
de Seyssel et al. show that strengthening language discrimination during self-supervised pretraining of a bilingual English/French HuBERT model reduces the performance gap compared to monolingual models. On continuous phonetic phone discrimination error, the bilingual baseline of 11.6% dropped to 10.4% (monolingual: 10.8%), while lexical accuracy rose from 52.1% to 56.7% (monolingual: 58.5%) and prosodic lexical performance from 68.9% to 72.9% (monolingual: 72.6%). These gains are largest when the intervention is applied early in training.
Xi et al. show that the confidence reasoning models verbalize is systematically overconfident and uncorrelated with actual correctness. By training a linear probe on hidden states between chain-of-thought and answer, they obtain a confidence signal with 5 to 38 times lower expected calibration error (ECE) than verbalized scores across four benchmarks. However, this probe performs no better than majority voting at selecting correct answers, indicating internal states capture "how certain" rather than "which answer." They introduce probe-guided self-distillation (Probe-SD): using the probe to relabel confidence in model-generated traces and fine-tuning the base model, achieving in-domain ECE of 0.024 (from 0.178) and out-of-domain ECE of 0.113 (from 0.542) on Qwen3-14B.
The authors developed a behavioral test suite to distinguish whether language models rely on shallow patterns or true generalization. Tracking models across pre-training checkpoints, they found that models frequently and abruptly alternate between these modes, often within tens of billions of tokens. This mode-hopping is not an artifact of noise or metric selection and cannot be smoothed by checkpoint averaging.
Zhang et al. introduce VSA2, a trainable sparse-attention system for video diffusion transformers that selects a small, query-dependent subset of token interactions. The authors report substantially faster 720p generation than a FlashAttention-3 baseline, while maintaining comparable or better video quality.
Liang and colleagues introduce Allspark, a framework for transferring reasoning behavior from a reinforcement-learned small model to stronger models through alternating text-based chains of thought. In experiments using Qwen models and ARC-AGI-2, the trained weak teacher improved some within-family and cross-family students, although the benefit depended on the inference configuration and required additional tokens.
Zhou et al. test whether relationships learned from persona prompts predict new responses and carry across models and prompts. They find persona effects are predictable within a model but not portable: borrowed gains underperform target refitting in most transfers, and rewording consistently breaks reuse.
The authors introduce Pruned CTC, a method that reduces the memory footprint of CTC training for large-vocabulary ASR by restricting alignment computation to the subset of vocabulary tokens that appear in each batch, plus the blank token, while retaining full-vocabulary softmax normalization. They prove that this reduction is mathematically equivalent to standard CTC in both loss and gradients. With a Zipformer-M encoder and a 180K vocabulary, Pruned CTC achieves a 5.1x reduction in full-step memory with only a 17% step-time overhead, matching standard CTC accuracy across three corpora.
Guo and colleagues trained a 200 million-parameter Transformer whose computation can stop after any layer, yielding a continuum of smaller language models from one artifact. Unlike suites trained only at selected exit depths, their Telescopic Language Model remained functional at all 20 prefixes and improved the reported quality-budget tradeoff without changing the architecture or adding inference overhead.
Laguna and colleagues test whether compressed key-value caches remain safe to reuse when a user asks questions unrelated to the cached document. They find that standard Cartridges preserve document-specific information well but can contaminate unrelated answers, weaken general knowledge and disrupt instruction following. Two simple modifications largely preserve these off-context capabilities with little reported loss of document utility.
Wang and colleagues present Tex-Zero, a native 3D texture system trained exclusively on 2D images rather than textured 3D assets. They convert image patches into artificial 3D arrangements, then use them to train both a variational autoencoder and a diffusion transformer. The authors report that the resulting model reconstructs and generates detailed textures on real, previously unseen 3D geometry.
Chen and colleagues target two sources of multimodal model overhead: redundant visual tokens and unnecessary computation inside the language-model decoder. Their training-free SPIDER framework retains complementary visual information from middle and deep vision layers, then selectively skips attention or feed-forward sublayers. On LLaVA-NeXT-7B, the authors report 79 percent fewer FLOPs while retaining 96 percent of baseline performance.
Chen and colleagues trace a late-training failure in BF16 FlashAttention-3 to two numerical errors, including a previously overlooked violation of a softmax conservation law. Their correction, GProj, reduced median query and key gradient errors to below 0.4 percent and trained a 450 million-parameter transformer to the same loss as FP32 attention, with a 4.7 percent increase in step time.
Zhang, Duan and Chen present a weight compressor that can target rates between the coarse options offered by fixed-width formats. On the Z-Image-Turbo generator, EntroPack produced lower weight and denoiser-output errors than comparable fixed-width representations, including roughly 24% lower relative L2 weight error than NF4 when targeting 4 bits per parameter.
Ding, Zazo and Hensman introduce Rotated Manifold Optimization, or RoM, an optimizer designed around a symmetry in low-rank adaptation: many pairs of LoRA factors represent exactly the same weight update. Across supervised fine-tuning and reinforcement learning experiments, RoM converged faster and reached lower held-out loss than the compared LoRA optimizers, while producing better or comparable downstream performance.
Zhu and colleagues evaluate action-conditioned joint-embedding predictive architectures, or JEPAs, as driving world models without training a driving policy. Their proposed AD-E2E-JEPA compresses the latent patches used for planning and reports a 100-fold inference speedup, while retaining planning accuracy and improving representations transferred to imitation learning.
Shao et al. address a weakness of sparse attention: tokens excluded from retrieval normally receive zero weight, even when useful information is distributed across the context. Their method, PQ-HSA, retains approximate contributions from those tokens and, at 128K context, improves accuracy while making the decode attention call faster than full attention on the tested hardware.
Zeng and colleagues examine whether every token in a chain-of-thought trace contributes equally to solving a problem. They report that normalized token log probabilities can identify compact regions containing decisive reasoning, then use this signal in reinforcement learning to reduce token consumption by 76% relative to the baseline without sacrificing overall reasoning quality.
Jeddi et al. investigate why reasoning vision-language models falter when most visual tokens are removed. They find that pruning does not always erase the necessary evidence: repeated sampling from the same sparse representation often produces correct answers that greedy decoding misses. Their self-distillation method improves how reliably models use this surviving information, reaching 92.43% of unpruned performance with only 10% of visual tokens retained.
Wang et al. identify inaccurate critic estimates as a source of oversaturation and artifacts during Distribution Matching Distillation, a method for compressing diffusion generation into a few denoising evaluations. Their proposed projection filters the error-prone component of each update, improving stability and benchmark performance in four-step video and joint video-audio generation.
The authors introduce UniAdapt, a method that learns the marginal utility of residual-vector-quantization (RVQ) refinements on a frozen codec and allocates tokens under exact serialized-bit budgets. In utterance-level tests, it reduces Log-STFT distortion by 1.07 to 4.39 percent across speech, music and environmental audio without increasing the budget. A causal streaming variant improves three of four speech rates with zero budget violations across 800 evaluations and runs faster than real time.
Yin and colleagues tested a local learning method that lets network modules update independently rather than waiting for gradients to traverse the entire model. Across Transformers pretrained on 15 billion tokens, SOLO remained within one percentage point of backpropagation in average zero-shot accuracy and achieved up to 1.44 times its best measured pipeline throughput.
The authors introduce TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery in theoretical computer science. It consists of 398 theorem-level challenges from 138 papers from STOC and COLT 2026. In their evaluations, the best-performing system (GPT-5.6 Sol max with discussion) achieved 23.6% verifier-accepted coverage after 10 rounds, while an agentic planning workflow reached 25.4%.
The authors introduce d-OPD, a future-aware on-policy distillation method for converting autoregressive LLMs into block diffusion language models. It addresses a fundamental mismatch: the block-diffusion student sees future context within a block, but the standard AR teacher distribution does not. By aligning the teacher distribution with the student's visible state, d-OPD improves average benchmark scores by up to 4.0 points and reduces training time by about 1.35-1.58x on Qwen3 models from 0.6B to 8B.
Moon and colleagues address open-vocabulary 3D segmentation, in which users identify scene objects with unrestricted text rather than predefined labels. Their EviSplat system stores separate features from multiple camera views and selects among them only after receiving a query, instead of compressing potentially conflicting evidence into one descriptor beforehand.
Zhaohan Zhang and colleagues propose ConfLens, a framework that monitors how an LLM's confidence in its final answer evolves during chain-of-thought reasoning. They find that shortcut reasoning consistently shows a pattern of premature confidence, where the model commits to an answer early. Their new metric, the Distributional Answer Commitment Score (DACS), detects this reliably across math and code tasks, improving shortcut detection F1 by over 4.3% over strong baselines.
Meng et al. propose CoDrive, a cross-vehicle video generation framework that produces world-consistent observations for multiple vehicles in the same scene. By interleaving local and global self-attention and injecting camera trajectories in a shared coordinate system, the model significantly improves trajectory controllability and cross-agent geometric and instance consistency while maintaining visual quality.
Wang et al. propose BaP, a test-time adversarial defense for CLIP that implants a defender-controlled backdoor probe into an MLP layer. Adversarial inputs strongly activate the probe, enabling detection and subsequent rectification via a small learned perturbation. The method raises average robust accuracy from 1.0% to 52.3% on a suite of 16 benchmarks, with up to 5.7x inference speedup over prior test-time defenses.
Faure and colleagues introduce ActionLens, a 6,701-question video benchmark designed to test whether vision-language models can associate an action with the correct person and moment. Across 20 models, the strongest system reached 68.8% accuracy overall, while on a human-reviewed subset it scored 65.9% against a 91.0% pooled human reference.
The authors propose Strategy-Diverse Reinforcement Learning (SDRL) to train open-source LLMs as adaptive meta-solvers for industrial-scale optimization. They show that combining solver-integrated reasoning, exact algorithms, and heuristic search with a hierarchical diversity reward allows the model to select the best strategy for each problem, outperforming DeepSeek-V4-Pro and GPT-5.5 on average across benchmarks and on industrial-scale tasks.
Ma et al. introduce Beyond3D, a VQA benchmark with 9,000 questions over 135 egocentric videos that tests whether VLMs can track an object's location after it is moved and has left the field of view. The best model achieved only 42.2% accuracy, with the largest failures in recovering when an object was last visible, indicating that current models struggle with out-of-sight spatiotemporal reasoning.
The authors introduce two complementary program-state reasoning tasks—buggy input-output reasoning and precondition-postcondition reasoning—and incorporate them into a staged post-training pipeline to build Comet-9B, a 9B-parameter language model. They find that adding both tasks to supervised fine-tuning on issue resolution improves success rates by 7.25 percentage points on SWE-bench Pro and 9.70 points on SWT-Bench Verified, with further gains from sequential reinforcement learning. Despite its small size, Comet-9B achieves scores comparable to reported results from GPT-5.2 and GPT-4o-based agents.
Lin et al. introduce ATLAS, a method for fine-tuning language models that preserves behavior on retained domains while learning new tasks. On Qwen3-8B, ATLAS achieved lower mean retained-output KL divergence than all seven published baselines at matched coding performance, with fewer rewritten answers across mathematical and commonsense domains.
Zhang et al. introduce SimpleCluster, a baseline that compresses video representations by grouping similar visual tokens across frames and replacing each group with its mean feature. Their experiments suggest that preserving broad coverage of the original feature distribution can matter more than sophisticated token selection, particularly when the model must retain very few tokens.
The authors introduce ORAV Bench, a benchmark for generating audio and video from several heterogeneous references linked by a textual instruction. Testing five frontier systems shows that aggregate rankings obscure composition-specific strengths, while models frequently copy unintended source content instead of transferring only the requested attributes.
Wang and colleagues study how to choose the teacher layer and duration of representation alignment when training diffusion transformers. They find that alignment is most useful for features the student cannot already reconstruct, then use this principle to improve image generation while reducing training cost.
Zhang, Shi, Liu and colleagues present a common mathematical account of distributional training, where generators learn by matching collections of real and generated features from frozen image encoders. Their resulting method, MGFlow, represents those distributions as Gaussian mixtures and reports stronger one-step generation than prior feature-distribution objectives on ImageNet and text-to-image benchmarks.
Deschenaux et al. introduce Simplex Diffusion Models, which keep intermediate predictions as probability distributions rather than repeatedly collapsing them into categorical token choices. The authors derive closed-form reverse transitions, use ordinary cross-entropy training, and report gains over masked and uniform diffusion on code generation, including after aggressive sampling distillation.
Chen and colleagues compare multimodal language models that use a pretrained visual encoder with models that instead feed projected image patches directly into the language decoder. Encoder-free models are less efficient at the tested scales, but their multimodal loss improves faster with compute, leading the fitted scaling laws to predict parity around 10^22 FLOPs.
Que et al. study how narrow harmful fine-tuning can weaken an instruction-tuned model’s safety behavior outside the adapted domain. Across four open-weight model families, they identify geometric signatures of this shift and show that projecting a learned harmful subspace out of parameter updates reduces free-generation misalignment by up to 80.0 percent on Qwen2.5-14B-IT.
Lévy and colleagues identify slow rotary frequency bands, whose wavelengths exceed the training context, as a weakness of Rotary Position Embedding during length extrapolation. Their alternative, Data-aware RoPE, performs best or comparably to other tested encodings across language models, synthetic tasks, music, genomics and neural signals, while reducing the tendency to favor recent tokens.
Chen and colleagues examine whether the classical edge-of-stability picture for gradient descent also describes language models trained with Muon, an optimizer that replaces matrix gradients with approximately semi-orthogonal directions. Their theory and experiments indicate that Muon separates two behaviors that coincide under gradient descent: short-term loss balance and reversal between successive update directions.
The authors present two surrogate benchmarks, ScAn-Bench-LLM and ScAn-Bench-VLM, built from 4524 and 8024 checkpoints of language and vision-language model pipelines, respectively. They use these to perform the first systematic comparison of scaling analysis methodology across data modalities, covering both data acquisition and extrapolation approaches.
The authors train a multimodal language model to alternate between chain-of-thought reasoning and tool calls that crop relevant regions from a table image. Across nine table benchmarks, InterTab raises its backbone model’s average accuracy from 68.28% to 73.17%, a gain of 4.89 percentage points.