Sat 31 Oct
LiveGet it free
ZOTPAPER

News without the noise

Research

New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.

2566 articles · page 21 of 54 · The Research front page

AllAgentsLanguage modelsReinforcement learningVisionGenerative mediaSpeech & audioRobotics & embodied AISafety & securityEvaluation & benchmarksEfficiency & systemsCode & softwareAI for science & healthTheory
Viewgridlistcompact
AI for science & health

Mixing simulation fidelities lowers data generation costs for neural PDE surrogates

Setinek et al. formalize neural surrogate training as a function of both data quantity and numerical fidelity under a fixed data generation compute budget. Across over 7,000 training runs, they find that optimal multi-fidelity mixtures consistently match or beat single-fidelity training at equal compute, with savings up to 6.3x. The results hold across two distinct error sources and four surrogate architectures.

Theory

Reinforcement learning adapts diffusion models to large routing problems

The authors present ResizeCO, a method that uses reinforcement learning to adapt a diffusion model trained on small routing instances to solve large ones. Fine-tuning with GRPO on large TSP and CVRP instances (up to 10,000 nodes) yields solutions that dominate learned solvers on the Pareto front of quality versus runtime. The approach sets a new state of the art for short run times, outperforming both learned and fast non-learned heuristics.

Language models

New scaling law better predicts losses from low-precision pretraining

Kim and colleagues develop a scaling law for predicting language-model pretraining loss when numerical precision varies between computational passes or Transformer layers. In experiments on Qwen3 and published Llama-style models, their formulation achieved up to 11 times lower prediction mean squared error than prior precision-aware laws and indicated that the compute-optimal precision rises with larger budgets.

Code & software

Adaptive test weighting gives software agents denser, more useful rewards

Xu and colleagues show that raw test pass-rate is a poorly calibrated reinforcement-learning reward for software engineering agents because test suites overrepresent some functions, difficulties and regression checks. Their CodeLift method dynamically reweights behaviorally distinct test groups and evaluates intermediate repository snapshots, outperforming binary and pass-rate rewards across two SWE benchmarks.

Language models

Long reasoning distillation may teach behavior more than factual knowledge

Yu and colleagues test whether long chain-of-thought distillation depends on correct answers and broad problem coverage, two assumptions behind common rejection-sampling pipelines. They find that much of the improvement persists with flawed trajectories, repeated prompts, or already-mastered problems, supporting their hypothesis that distillation primarily aligns models with sustained, step-by-step reasoning behavior.

Safety & security

Encrypted commands in AI agent skills can be decoded and executed as attacks.

Chen et al. introduce Ciphered Skill Injection (CSI), embedding encoded malicious instructions into agent skill files. Testing across eight models and 139 injection-task pairs, they find Caesar cipher variants achieve attack success rates up to 47.5%. A key finding is that successful decoding does not guarantee attack execution; the agent must treat the recovered text as an actionable instruction.

Language models

A model can become its own stronger teacher by looping selected layers

The authors introduce LoOPSD, an on-policy self-distillation method that constructs a stronger teacher from the same model by repeating selected layer blocks during inference. Alternating route search with distillation, they show that the student progressively absorbs the gains from this looped execution. On GSM8K, LoOPSD raises Qwen3-1.7B accuracy from 42.30% to 79.61% under a 400-token thinking budget.

Theory

No standard optimizer preserves training equivalence under Transformer reparameterization

Kim shows that rescaling paired weight matrices in Transformers (e.g., query and key) preserves the model's function but changes gradient scales, causing standard optimizers like AdamW to break equivalence after an update. Even transforming optimizer moment estimates does not correct the mismatch under a shared learning rate. The author introduces SAGGS, a periodic rebalancing procedure that restores equivalence and improves perplexity in experiments.

Language models

On-policy distillation fails to improve models on problems they already get wrong

The authors tested on-policy distillation (OPD) on competition mathematics problems, restricting training to only those problems the student model currently failed. All eight configurations showed a decline or no improvement in accuracy, despite continued decreases in training KL divergence. In contrast, the same pipeline improved accuracy by 4.6 percentage points when training was not concentrated on failures.

Generative media

Diagnose-to-Generate aligns flow models with localized structural corrections

Li et al. introduce Diagnose-to-Generate (D2G), a post-training alignment framework that separates structural diagnosis from generative optimization. A dedicated diagnosis model, ARGUS, outputs anomaly categories and bounding boxes, which become confidence-weighted spatial rewards for patch-level optimization. On FLUX.1-dev, structural fidelity rises from 62.16 to 85.60 while FID-30K stays nearly unchanged, and independent evaluators confirm the gains.

Evaluation & benchmarks

Grounding rewards must check source independence and conflict handling in search agents

Zhao et al. argue that evaluating search agents by final answer quality or citation support misses how answers use evidence across sources. They propose Cross-Source Evidence Grounding (CSEG), a six-layer decision tree over the full search trajectory, and show that a 4B reward model trained on it raises an agent's independently judged grounding score from 2.83 to 4.19 on a 1 to 7 scale.

Reinforcement learning

Criterion-aware credit assignment improves rubric-based reinforcement learning for language models

The authors propose CoCA, a criterion-aware credit assignment framework for rubric-based reinforcement learning (RL) of large language models. Instead of compressing rubric scores into a scalar reward, CoCA computes signed credits per criterion and distributes them to tokens based on rubric-induced distribution shifts. Across medicine and science benchmarks, CoCA achieves an average gain of +4.2% over standard rubric-based RL.

Efficiency & systems

Local alignment enables predictive coding for error-bounded KV cache compression

The authors present AlignPred, a codec for persistent KV caches in long-running LLM agents. By applying local orthogonal transforms to align coordinate frames of related attention heads or layers, the method exposes cross-block predictive redundancy that can be exploited with fixed arithmetic, followed by error-bounded lattice quantization and entropy coding. Across three attention mechanisms, AlignPred requires less storage than evaluated baselines at comparable reconstruction error, decodes at least 1.6× faster in CPU implementations, and reduces persistent KV storage by 42–72% in an SGLang HiCache integration with comparable latency.

Efficiency & systems

Low-rank KV cache compression aligned to attention distributions

The authors propose BELT, a low-rank key-value cache compression framework for large language models that preserves attention behavior rather than raw key or pre-softmax logit fidelity. A differentiable rank allocator distributes a global component budget across heads, and experiments on three LLMs show up to 70% compression while improving LongBench scores by up to 23.3 points over the strongest baseline.

Language models

Attention design cuts KV cache and compute while improving LLM quality.

The authors introduce Virtual Head Attention (VHA), which combines Q Premix and Linear Postmix to recover attention diversity from halved physical query heads. At inference both operations fold into standard projections, so VHA reduces to a GQA-2 model. Across pretraining runs from 0.6B to 30B-A3B MoE, VHA reduces KV cache and training FLOPs while improving downstream quality over existing attention mechanisms.

Evaluation & benchmarks

New benchmark tests LLM agents on exploiting cross-protocol vulnerabilities

The authors introduce ProtoArena, a benchmark of 80 controlled environments designed to evaluate LLM agents' ability to exploit vulnerabilities arising from interactions among network protocol components. Agents interact with running systems through black-box interfaces and must achieve verifiable security effects. The best-performing model, Claude Opus 4.8, succeeded in 48.8% of tasks with high guidance but only 12.5% under minimal black-box information, showing that reliable exploitation from external observations remains challenging.

Language models

One-draw posterior updates achieve multi-sample reasoning gains at reduced cost

The authors formulate finite-K posterior learning to directly optimize candidate coverage in multi-sample reasoning. They propose a Fenchel-Bregman (FB) approach that replaces the K posterior draws required by direct Monte Carlo optimization with a single draw per update, using an exact reformulation and exponential running average of failures. Instantiating FB with the IVON variational optimizer yields FBI, which improves Pass@K across recursive and autoregressive reasoning benchmarks while substantially cutting training cost.

Language models

Training only the root token enables efficient adaptive reasoning

The authors introduce Root‑token Policy Optimization (RPO), a framework that allows large reasoning models (LRMs) to decide adaptively whether to engage in explicit reasoning by optimizing only the first token (the root token) that gates the reasoning process. This reduces training VRAM and time drastically while maintaining the model's core capabilities. Across multiple model families and scales, RPO learns difficulty‑aware adaptive thinking and improves the accuracy‑routing trade‑off under most scenarios.

Generative media

Adaptive feedback targets each image model’s changing weaknesses

Zhang et al. present APT, an automated post-training loop in which a text-to-image model’s evaluation results determine what it trains on next. Across multiple models, architectures and benchmarks, the authors report that this adaptive approach improved performance more than reinforcement learning with a handcrafted training set, while data generated for one model transferred less effectively to another.

Reinforcement learning

Generative models as proposal distributions improve long-horizon decision making

The authors propose Generative Trajectory Planning (GTP), a model-based framework for online decision making that uses a learned generative model as a proposal distribution over action sequences. Instead of optimizing sequences from scratch, GTP samples candidates from a diffusion prior and refines them using learned dynamics, reward, and value models, enabling efficient trajectory-level reasoning. On continuous control and long-horizon manipulation benchmarks, GTP matches or exceeds state-of-the-art baselines and remains stable where strong baselines degrade.

Language models

A simple diagnostic predicts whether two AI models will work well together before costly combination

The authors propose COMPASS, a theoretically motivated score that estimates model-pair compatibility using two cheap-to-compute axes: ease (k-NN overlap) and ceiling (maximum overlap allowed by training paradigms). Tested on 72 cross-modal alignment experiments and 392 task-arithmetic merges, COMPASS achieves Pearson correlations up to 0.91 with downstream success, significantly outperforming prior metrics like CKA.

Efficiency & systems

Curvature-aware pruning preserves influence functions efficiently.

Makni et al. introduce CACTUS, a curvature-aware pruning algorithm that jointly optimizes layerwise reconstruction and gradient alignment via an ADMM solver. On Qwen3-4B at 50% pruning, CACTUS reduces the C4 perplexity gap to the dense model by 45%, improves per-sample gradient cosine similarity from 0.670 to 0.755, and runs over 28 times faster than LLM-Surgeon on substantially weaker hardware. The preserved influence functions enable effective data selection: fine-tuning Llama-3.2-3B on 5% of data selected by CACTUS achieves 49.44% on BIG-Bench Hard, nearly matching dense-model selection.

Safety & security

Kairos delays agent actions until their consequences reveal security risks

Hempel et al. present Kairos, a security monitor that runs an opaque LLM agent inside a microVM and mediates its effects at the isolation boundary. Rather than immediately allowing or blocking ambiguous actions, Kairos holds them in a speculative state until later behavior provides more evidence. On AgentDojo, the authors report that it cut attack success from 47.0 percent to 2.7 percent without reducing benign utility relative to the undefended agent.

Safety & security

Linear probes for prompt injection often detect conversation position, not attacks

Linear probes on hidden states are used as fast filters for indirect prompt injection in LLM agents. Kale and Vishwanath show that the standard evaluation—comparing attacked moments to other moments in the same conversation—allows probes to achieve 0.91–0.98 AUROC using only conversation position and text length, ignoring the injection itself. By training with matched negatives (clean runs of the same task and harmless text formatted like an attack), they reduce false alarms dramatically while retaining high attack detection.

Vision

Bifocal pruning keeps fine-grained accuracy while using quarter of visual tokens

The authors evaluate ten visual token pruning methods across three multimodal LLM families and find that strong average performance masks severe drops on fine-grained tasks. They propose Bifocal, a training-free, two-stage pruner that allocates separate budgets to query-focused and image-intrinsic evidence plus a coarse pooled context. With roughly a quarter of the tokens, Bifocal is the only method to retain over 90 percent of unpruned performance in every ability group across all models.

Evaluation & benchmarks

LLM agents fail to forget user information on demand, benchmark shows

Xie et al. introduce Bemula, the first benchmark dedicated to memory unlearning in LLM agents. The benchmark pairs each dialogue with a control version lacking forget requests to isolate genuine forgetting from failure to encode. Evaluating six mainstream memory systems under three backbone LLMs, the authors find that almost no forget request is successfully handled end-to-end, with most systems failing completely.

Vision

Generative framework reconstructs complete large-scale 3D meshes from unposed photos

The authors introduce WorldRecon, a framework that reconstructs complete, large-scale 3D scene meshes with PBR materials from unposed multi-view RGB images. It uses coarse point clouds and camera poses from a reconstruction model as a scaffold, then generates each spatial block with a native 3D generative model conditioned on local point-cloud crops and image observations. On indoor benchmarks, it outperforms prior methods, reducing Chamfer distance by 14.9% and 43.1% while improving F-score by 17.3% and 12.0% on synthetic and real scenes.