News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 32 of 54 · The Research front page
Jiang et al. introduce Guidance-TTT, a test-time reinforcement learning framework that trains a small guidance model to propose high-level strategic changes while a larger execution model remains frozen. Across four discovery tasks, the system outperforms prior methods and matches top public leaderboards at lower training cost.
The authors propose Flash-OPD, a method for on-policy distillation that stops each training trajectory individually based on observed teacher-student compatibility events rather than using a fixed rollout length. Across diverse teacher-student settings, it achieves 2.2x to 7.5x speedups over standard on-policy distillation while maintaining or improving accuracy.
Jie Wang, Shiwei Luo, Qi Zhang, and Yuanbin Wu train flat byte Transformers without specialized tokenization architectures and show that they outperform subword Transformers as model size scales. The models also learn local text abstractions similar to tokenizers, with segmentation positions that can be exploited for speculative decoding, yielding 3.4 times more accepted tokens.
The authors introduce Mulligan, a data-collection strategy for batch-online robot learning that picks each round's initial states from observed failures and untried regions. Across three real-world and two simulated tasks, Mulligan beats uniform initial-state sampling at matched budgets, and with a value function that reranks policy samples improves final real-task success by 10 to 34 percentage points. Operator-assisted collection completes 98% of episodes, so learning does not stall task completion.
Dhouib et al. study why training-based prompt-injection defenses can impair an agent even when no attack is present. Their RAISED framework uses a frozen copy of the original model as both data generator and teacher, training a student to reproduce clean behavior in clean and attacked contexts. On AgentDojo with Gemma, it cut attack success from 39.5% to 1.0% while losing one point of benign utility.
Chen and colleagues treat an agent’s memory system as executable code that an LLM can iteratively revise using task feedback. Their method, PrisMem, optimizes distinct abilities such as factual retrieval, temporal tracking and cross-session synthesis separately, outperforming the strongest reported baselines by 10.54 percentage points on BEAM-1M and 7.83 points on LongMemEval-M.
Jiang and colleagues tested whether language-conditioned robot policies actually use instructions to choose among multiple feasible actions. Across vision-language-action policies and world-action models, robots usually pursued the target associated with the scene even when the instruction validly requested another visible object, exposing a gap between encoding language and letting it control behavior.
Saini, Adsumilli and Bovik derive an image-distance metric from the Jacobian, or local sensitivity, of a frozen DINOv2-S vision encoder. Their Jacobian Lens Distance (JLD) agrees closely with human judgments across four perceptual databases, remains stable when image resolution changes, and can also score video.
Saini and colleagues introduce MEND, a reinforcement learning method for post-training text-to-image flow models without KL penalties, frozen reference models or advantage weighting. In reported experiments, MEND improved several reward benchmarks more quickly than Flow-GRPO, ReFL and DiffusionNFT while explicitly limiting how far individual samples move.
Chen et al. identify a failure in block diffusion drafting, where independently predicted positions often repeat the same token and consequently produce shorter usable proposals. Their D-Loop method runs the same diffusion drafter twice, first proposing a token block and then conditioning on part of that proposal to regenerate the remaining tokens in parallel.
The authors present PixelUMM, an encoder-free model that handles both image and video understanding and generation directly in pixel space. By representing images as spatial patches and videos as spatiotemporal tubelets, and using a Mixture-of-Transformers architecture, the model achieves competitive performance across tasks. This design removes the need for separate pretrained visual encoders (e.g., ViT for understanding, VAE for generation), simplifying the visual interface.
The authors introduce Interactive Intent Alignment, a task setting where agents must recover and track user intent despite miscommunication and evolving goals. They develop Drift-Bench++, a benchmark with controlled misalignment and intent shifts, and evaluate current LLM agents, finding that even with better interaction, performance remains far from perfect.
The authors propose Restoring without Forgetting (RwF), a continual learning method for image restoration that adapts models to sequential degradation tasks without accessing past data. RwF identifies a small set of degradation-critical filters using integrated gradients in parameter space, then generates task-specific filters from a filter bank using factorized low-rank transformations. On six restoration tasks, RwF avoids catastrophic forgetting and matches the quality of all-in-one models, while using about 10 times fewer additional parameters than LoRA-style adaptation.
The authors studied how positional encoding (PE) choice shapes the internal mechanisms language models use for in-context retrieval across 22 open-weight models from eight families. They found that standard RoPE models primarily use positional cues, while PE hybrids (SWA NoPE) shift to semantic cues, explaining observed trade-offs in long-context tasks.
Jiang et al. directly test a common premise in unified multimodal model design: that less conflict between task gradients produces a better balance between visual understanding and generation. Across 63 reported configurations and 372 checkpoints, standard directional conflict measures were weak predictors, while an intervention that progressively suppressed conflict left the final trade-off unchanged.
The authors develop an information-theoretic model showing that post-training performance scales as Θ(√min(log M, K)), where M is the number of preference comparisons and K is the divergence budget. They validate this scaling law with extensive experiments using a 70B gold reward model and smaller proxy models, achieving R² between 97% and 99%. The result provides a principled way to set divergence limits based on available preference data, clarifying the trade-offs in reward optimization.
Li et al. introduce PAMI, a hierarchical generative framework that synthesizes full-body human-object interactions from text. By representing object motion as a weighted vote from multiple body-part anchors, the model produces more accurate and contact-rich motions than previous approaches, boosting contact recall by 14.5% on the InterAct benchmark.
Lu et al. introduce Activation-Conditioned Self-Distillation (ACSD), which turns differences in a model’s own reasoning trajectories into dense next-token training signals. Across five language models, ACSD achieved the highest average accuracy among the evaluated methods on four mathematics benchmarks. It also improved code-generation results on the tested 8-billion-parameter model, while requiring no reference solutions during distillation or activation intervention at inference.
The authors introduce E2E-SWE, a benchmark of 186 whole-repository generation tasks across 11 programming languages. They find that pass@1 ranges from 11.7% to 67.7% across 13 models, providing strong differentiation and significant headroom for future progress.
Ji and colleagues introduce LIFT, an image-to-video system designed for camera movements that reveal substantial unseen areas. The model uses a sparse last-frame layout to control what appears in the future view and where, while on-policy self-distillation transfers guidance from a teacher trained with layouts for every frame.
Through controlled comparisons, the authors demonstrate that reconstruction fidelity alone is insufficient for choosing representations in world-action models, and that perceptual features benefit from calibration for dynamics modeling. Their proposed method, ReWAM, achieves 93.6% success on RoboTwin 2.0 and an average score of 12.29 on RoboDojo without relying on generative video pre-training, by using frozen DINO features processed through feature calibration and a temporal representation bottleneck shaped by action-loss gradients.
The authors propose LeapQuant, a training-free method for 8-bit quantization of recurrent states in linear attention layers. It reduces state memory traffic by 3.4x and achieves up to 1.47x end-to-end speedup on NVIDIA GPUs while matching FP32 accuracy on long-context reasoning and code generation tasks across Qwen, Kimi, and GLM models.
Puduppully et al. trained GPT-2-sized models on iGSM, a synthetic mathematics benchmark where every required quantity and dependency is known, allowing each generated reasoning step to be checked. On the hardest out-of-distribution problems, 31.6% of correct answers were paired with invalid traces, showing that answer correctness does not guarantee a valid derivation.
The authors propose SplitMoE, a video diffusion architecture that bifurcates the expert pool into semantic experts (for high-level abstraction) and generic experts (for residual visual details). Using prototype-guided routing and pull-push regularization, they avoid the uniformity trap of standard MoE—where load-balancing scatters coherent patches across experts—achieving faster convergence, lower fragmentation, and better generation fidelity on video benchmarks.
Fu et al. introduce TabFM-Auto, which pairs a tabular foundation model (TabFM) with a language model agent that iteratively edits the data pipeline. Across 51 TabArena datasets, five configurations take the top five spots, with the best raising TabFM from 1785 to 2013 Elo, and the discovered pipelines transfer to other frozen tabular models.
The authors propose EOGP, a method to compress training gradients for reusable influence function computation in LLMs. They show that PCA of half-whitened gradients is optimal for linear compression under a worst-case error criterion. EOGP approximates this using EK-FAC dimensionality reduction, PCA within the subspace, and one-bit quantization, achieving better influence prediction than baselines while using 1/16 the storage on GPT-2.
Jin et al. analyze Adam when an objective’s local smoothness may deteriorate as its gradient grows and stochastic gradients satisfy only second-moment bounds. They show that Adam’s self-normalization keeps its trajectory inside a controllable region, establishing high-probability convergence for the full range p < 2 and expectation guarantees for p < 1.
Gao et al. introduce EngiWorld, a benchmark testing whether AI agents can complete realistic workflows in professional engineering software rather than merely produce plausible-looking outputs. Evaluating seven frontier agents, the authors find substantial weaknesses in constraint satisfaction, tool coordination and verification of finished designs.
The authors demonstrate that base language models can perform complex reasoning if their output begins with specific short token sequences (e.g., "Okay" or "Alright,"). These cues, learned from training data, boost MATH-500 accuracy from 42% to 78% for Olmo-3-7B, matching its RL-trained counterpart. The work further shows that RL essentially increases the likelihood of such cues, and that causal data interventions can turn arbitrary words into effective reasoning triggers.
The authors propose BRANCH-MoE, a routing architecture that places experts at leaves of a binary decision tree. By centering branching probabilities on arrival-weighted mean scores estimated via exponential moving average, the method prevents routing collapse and provides theoretical convergence guarantees. Experiments show it matches or outperforms flat routers on four tabular benchmarks while inducing localized expert co-activation.
A team from Zhejiang University, Tencent, and Peking University introduces Programmatic Search Agent (PSA), which gives agents control over how retrieved candidates are processed and presented, not just query reformulation. Across two benchmarks with five policy backbones, PSA improves task success by 4–7.5 percentage points and reduces final-step tokens by 28–34%, confirming that evidence delivery is a critical bottleneck in search agents.
The authors formalize 'predictive sufficiency' to measure what access to the true latent context adds over history alone for next-step prediction in model-based reinforcement learning. They find that in some settings the agent's behavior implicitly identifies the context, leaving no headroom for a context encoder, while in others residual headroom persists but learned state representations recover it only partially.
The authors present RealtimeWAM, a post-training framework that accelerates World Action Models (WAMs) by enabling one-step action generation and overlapping execution of video and action experts. On multiple robot manipulation benchmarks, RealtimeWAM achieves up to 25x speedup (12.2 ms inference latency on H100) while maintaining near-lossless performance (<1% accuracy drop).
Zhang et al. combine heterogeneous human-object motion capture, robot-specific retargeting and physics-based imitation to train one humanoid controller for dexterous whole-body tasks. Their central result is a data flywheel: successful variations of existing interactions become new references, progressively expanding the range of executable motions without changing the underlying task.
Chen et al. tested whether language models track which requirements remain active as user intent evolves across multiple turns. Across 414 tasks in four domains, models lost substantial accuracy when information was distributed across dialogue and often incorporated rejected or superseded instructions into their answers.
Zhang et al. trained a hierarchy of action-conditioned Joint-Embedding Predictive Architectures, or JEPAs, which predict future representations rather than reconstructing pixels. Each level operates over a longer temporal stride and in its own latent space, allowing upper levels to represent slower, more predictable aspects of an environment. Across simulated tasks, this hierarchy improved planning success while reducing test-time search, including a rise from 18% to 73% success on Visual AntMaze.
Chen et al. analyze why sparse attention loses fidelity in diffusion transformers, separating errors caused by coarse token grouping, approximate selection and discarded attention mass. Their training-free MC-Sparse method addresses all three and reports denoising speedups of 1.80 times for video and 2.32 times for 3D generation, with negligible measured quality loss relative to dense attention.
Yang and colleagues treat reinforcement learning for joint audio-video generation as a changing credit-assignment problem: the model must determine both where each reward should update the network and how strongly that reward should count. Their Adaptive Reward Routing method tracks cross-modal influence and reward conflicts during training, producing consistent improvements over fixed-routing and fixed-weight baselines in modality quality, semantic alignment and audio-video synchronization.
Chen and colleagues present a pipeline for converting egocentric human demonstrations into training data for coordinated humanoid loco-manipulation, where locomotion, balance and manipulation must work together. With 200 human demonstrations per task, their policies achieved a mean score of 51.67% across four real-world tasks, without using robot demonstrations from those target tasks.
Shi et al. propose Interpolated Policy Distillation (IPD), which defines the rollout policy at each decoding step as a linear interpolation between student and teacher token distributions, governed by a single coefficient. They accelerate exact sampling with a speculative decoding rule that preserves the interpolated distribution. IPD consistently outperforms off-policy SFT, on-policy distillation, their two-stage combination, and recent segment-interleaving baselines on text-only and multimodal reasoning benchmarks.
The authors propose Trajectory Soup, a strategy that forks multiple independent branches from a shared pretrained checkpoint, averages the strongest checkpoints within each branch, and then averages the resulting anchors across branches. Under matched compute budgets, the merged model outperforms the strongest single-trajectory average and keeps improving as more trajectories are added, even after a standard post-training pipeline.
The authors identify trace-response divergence (TRD) in a natural distillation setting where a think-mode teacher provides privileged supervision for a no-think student. They propose ThinkOPD, which combines reward gain and TRD to weight student responses within each group, improving average performance by 1.7–3.0 percentage points over uniform weighting across same-model experiments.
He and colleagues introduce LongSpark, a speculative-decoding drafter whose per-round computation and context state remain constant with respect to prefix length. Across reasoning, coding, dialogue and long-context evaluations, the authors report 1.88 to 2.13 times end-to-end speedups and lower mean time per output token than competing drafters.
The authors propose moesq, an end-to-end framework that jointly optimizes semi-structured sparsity and low-precision quantization for Mixture-of-Experts models, leveraging hardware Sparse Tensor Cores. Across models from 30 billion to one trillion parameters, moesq preserves 96.09% of the original model's accuracy while enabling up to 1.65x faster matrix multiplication and 1.18x higher serving throughput on NVIDIA B200 GPUs.
Wu and colleagues present a video world model that generates future views from a single sunny or adverse-weather image while separately controlling camera motion and weather. Their central result is that explicitly representing the change from observed to requested weather produces more accurate weather control than asking a generation backbone to infer that change implicitly, while maintaining competitive scene consistency and camera adherence.
The authors extend the Moshi paradigm to long-horizon, multi-party, bilingual full-duplex speech. They release MultiTalkPT and MultiTalkFT, a 57.6k-hour synthetic training corpus, and MultiTalkBench, a benchmark of real recordings averaging 32.6 minutes. Their model substantially outperforms open-source baselines including Moshi, MiniCPM-o-4.5, and Qwen3-Omni-30B-A3B-Instruct on this benchmark.
Sanz-Guerrero et al. identify a hidden source of non-determinism in LLM evaluation: the automatic injection of the current date into system prompts. Across 9 recent LLMs and 6 datasets covering multiple-choice QA, math reasoning, code generation, and machine translation, the authors find that performance fluctuates solely with the date, with deltas of up to 6% accuracy on MCQA, 14% on math, 7% pass@1 on code, and 2.84 BLEU on translation. This date effect exceeds variability from batch size or numerical precision, and standard techniques like chain-of-thought or few-shot prompting do not reduce it.
Fan et al. present TIPS, a reinforcement-learning method that trains a model to examine a reasoning trace, label its steps and judge its final outcome, while rewarding only the correctness of the outcome judgment. Using 3,200 outcome-labeled trajectories, their best 4-billion-parameter model reached 85.2 F1 on ProcessBench, although it remained behind o1-mini.