Searcharxiv⌕ Search

arXiv subjects

Ravid Shwartz-Ziv

Publications and source records attributed to Ravid Shwartz-Ziv.

At least 19 recordsLinked to original sources

Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)

Open-weight language models are widely observed to get worse the longer they generate. Often, attempts to fix this involve more training. At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives. We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens, covering SWE-bench Verified coding agents, long chain-of-thought reasoning, long-form writing, long structured output and long-context reading. Samplers that keep their pool of candidate tokens small and steady as temperature rises don't lose accuracy. Samplers that let in low-probability tokens fall apart. At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62 and win every issue the two samplers disagree on, on three Qwen3.5 models. Over all benchmarks top-nsigma keeps a median 100% of its temperature-1 score where top-p keeps 26%. Counting how many tokens each sampler lets through explains why. In one Llama-3.1-8B story, going from temperature 1 to 2 grows top-p's pool from 4 tokens to 51,845 while top-nsigma stays at 3. Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77), and when we pick the pool at temperature 1 and only apply temperature afterwards, Top-H, min-p and p-less stop collapsing. At the default temperature of 1 most samplers perform about the same, We release all code and data to the public.

cs.AI↗

Do Multi-Agents Dream of Electric Screens? Achieving 100% on AndroidWorld Through Task Decomposition

We present Minitap's mobile-use, a multi-agent system for Android UI control, and evaluate it on the 116 tasks of AndroidWorld, on Android emulators in Minitap's cloud device infrastructure, real Pixel 7 phones from our device farm and the stock emulator. On fresh, held-out task instances (three pre-registered seeds never used during development), the final system reaches a pass@3 of 100.0% (95% CI 96.8-100.0); pass^3 is 94.0% (95% CI 89.7-97.4) and the mean success rate 98.0% (95% CI 96.0-99.7). Minitap decomposes each task across five LLM agents (planning, subgoal tracking, state capture, decision, execution) and deterministic nodes, with mechanisms that read typed text back from the device, verify the goal on device evidence and detect repeated failures. A study of the initial system measures what each part contributes: removing only the decomposition changes success by -26.1 [-35.6, -16.7] points, the mechanisms together add +11.7 [+2.2, +21.1] points over a single agent, and the decision agent is the one role that needs a frontier model. We then improve reliability with general changes, a few of them triggered by the goal's wording. On MobileMiniWoB++, a second benchmark never used during development, the final system improves on the initial one by 7.1 points (McNemar p = 0.035; 95% CI -0.5 to +14.7). Minitap is open source: https://github.com/minitap-ai/mobile-use

cs.AI↗

Delayed Effects of Temporary Weight Decay in Muon Pretraining

A temporary optimization intervention can increase validation loss when it ends yet reduce it at the final training budget. We demonstrate this with temporary weight decay in Muon-trained decoder-only Transformers with 162M, 405M, 1.4B, and 3.4B parameters. For models up to 1.4B parameters, moving a fixed-duration decay window from early to late training reverses its final effect. At 162M, changing window duration alters final loss even at approximately matched total shrinkage. Most strikingly, a schedule selected at 162M and transferred without retuning to 1.4B is 0.26 nats worse than no decay when removed, overtakes it only after 73% of training, and ultimately finishes 0.028 nats better while also outperforming the best sampled constant-decay baseline. Early decay leaves a persistent weight-norm deficit, which keeps relative steps larger after the intervention ends. An approximate norm-dynamics recurrence captures the slow convergence of these relative steps, and fixed-norm replays of the measured trajectories reproduce the performance ordering and a similar loss difference. At 162M and 405M, norm-reset experiments show that smaller weight norms at the end of the decay window explain most of the final loss improvement. A separate Gram intervention exhibits related post-removal norm and relative-step dynamics. These results show that temporary optimization schedules can have delayed effects long after removal and should therefore be evaluated at the intended final training budget, not only at removal.

cs.LG↗

SCOPE: Observation-Conditioned Full-Target Prediction for Sparse PDE Inference

Recovering complete physical fields from sparse observations is challenging because the measurements may not uniquely determine the underlying state. Diffusion-based PDE solvers address this problem through iterative sampling whereas neural operators provide deterministic one-pass predictions. We propose SCOPE (Sparse-Context Observability-aware Predictive Embeddings) to recover complete PDE fields from sparse observations by coupling full-field latent prediction with physical reconstruction. A shared decoder reconstructs fields from both predicted and complete-view representations so that representation learning is guided by both physical recovery and latent matching. We derive a quadratic risk decomposition at fixed teacher-decoder pairs showing why optimal latent prediction need not yield optimal field reconstruction. We also establish sufficient conditions for decoder improvements on complete inputs to transfer to recovery from partial observations. Experiments across five PDE settings show that SCOPE outperforms mask-aware neural operators on all ten forward and inverse tasks and achieves lower errors than those reported for diffusion-based solvers including DiffusionPDE and FunDPS. Decoder-only adaptation further improves recovery without retraining the backbone while retaining deterministic single-pass inference.

cs.LG↗

HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning

Graph self-supervised learning aims to learn transferable representations from large-scale unlabeled graph data. Joint-embedding predictive architectures (JEPAs) avoid explicit negative-pair construction and raw-input reconstruction by predicting masked targets directly in latent space. However, existing graph JEPAs typically rely on a single predefined graph partition, biasing the learned representations toward one structural granularity and limiting their ability to capture complementary patterns at different graph scales. To address this limitation, we propose HP-JEPA, a hierarchical partitioning framework for multi-resolution graph joint-embedding prediction. HP-JEPA organizes each graph into an ordered bank of coarse-to-fine partition resolutions and performs context-target latent prediction separately at each resolution using an online encoder, an exponential-moving-average target encoder, and a latent predictor. The resulting resolution-specific graph representations are subsequently integrated through concatenation or task-specific resolution weighting, allowing downstream models to combine complementary local, regional, and global structural information. Experiments on seven graph classification benchmarks and one graph regression benchmark show that HP-JEPA outperforms the fixed-resolution Graph-JEPA baseline on 6 of 8 tasks, improving upon Graph-JEPA on most evaluated benchmarks. Size-stratified analyses further show that HP-JEPA achieves higher accuracy than Graph-JEPA in most evaluated graph-size quartiles on three representative datasets. These results highlight the effectiveness of hierarchical multi-resolution partitioning for transferable graph representation learning.

cs.LG↗

PDE-OBS: Controlled Evaluation Across Observation Patterns

Physical-field reconstruction and forecasting depend on both measurement density and spatial layout, yet evaluation under a single observation pattern does not characterize performance when that pattern changes. We introduce PDE-OBS, an integrated benchmarking platform spanning numerical data generation, model training, and inference and evaluation under varying observation conditions. It combines 560,000 fields and trajectories from seven partial differential equation families with configurable observation operators and seven adapted baseline methods for stationary reconstruction and short-horizon forecasting. Separating observation construction from physical records allows users to specify parameterized patterns and deterministic mixtures for training and testing while preserving prediction targets and data splits. The evaluation protocol uses references trained for each test pattern to compare models on identical test observations and targets, alongside equal-count groups for spatial-layout comparisons. On a 14,000-record subset, we evaluate 441 trained models under nine test patterns, yielding 3,969 evaluations. Mean cross-pattern error exceeds mean matched-pattern error in all 49 PDE-method pairs, and this finding persists in a configuration-matched subset of 117 models. Denser test observations do not consistently reduce error for a fixed model. Mixed-pattern training on five completed pairs reduces large single-pattern transfer errors, although destination-trained references usually remain more accurate. Together, the benchmark and findings support systematic evaluation of observation-pattern sensitivity and provide a reusable workflow for developing methods under changing measurement conditions. Code: https://github.com/ru1ch3n/PDE-OBS.

cs.LG↗

Learning Functional Subspaces for Neural Network Compression

Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keeping the matrices dense, and thus efficient on standard hardware. Existing methods, however, choose the subspace to remove from each weight matrix with local closed-form criteria: activation energy, layer-wise reconstruction error, or a quadratic approximation of the loss. These criteria ignore how errors propagate through the network, so at high compression the errors compound with depth and performance collapses. We introduce Learnable Subspace Projections (LSP), which instead learns the subspaces to discard end-to-end. Each linear layer, or tied group of layers that read the same activations, is assigned an orthogonal projector. All projectors are optimized jointly against a global objective--the KL divergence to the dense model's output distribution or the model's original training loss--while the pretrained weights remain frozen. Projectors are initialized from a whitened SVD truncation, and ranks are allocated by the output KL each projector induces per parameter saved. After training, the projectors merge into standard low-rank factors, with each tied group sharing one factor. In attention, this also lets the model cache one narrow latent in place of full keys and values. Across LLMs (OPT-125M/1.3B, Qwen3-4B, Llama-2-7B) and ViT-B/16, LSP outperforms baselines, and its advantage widens as compression increases. At -70% compression, LSP brings Llama-2-7B to 10.9 WikiText-2 perplexity and 42.2% mean zero-shot accuracy, versus 13.3 and 36.0% for the strongest baseline. The factorized model decodes up to 1.6x faster than the dense model at small batch sizes, and aching the shared latent shrinks the combined memory of weights and KV cache by 13.5x at a 128k-token context, versus at most 6.5x for untied baseline factorizations.

cs.LG↗

LLMs Trust Their Own: Identity-Dependent Conformity in Multi-Agent Systems

Large language models (LLMs) are increasingly deployed in multi-agent settings, where agents observe and influence one another, making social influence a key dimension of AI behavior and safety. We investigate whether LLMs' responses depend on the social identity of other agents, beyond the effect of their consensus. We construct judgment tasks with a single correct answer, and place models in a multi-agent setting where they receive incorrect answers from other agents whose social identities (AI or human, model family, or an arbitrary minimal group) are either shared with or distinct from their own. Across 12 open-weights models and nine tasks, we find a bidirectional effect of group identity on conformity to incorrect answers: in-group consensus increases conformity (in-group favoritism), whereas out-group consensus decreases it (out-group divergence). Unlike humans, for whom one ally breaking the consensus sharply reduces conformity, models are unmoved by an ally from the majority's group. Worse, a correct ally from the opposing group intensifies this bidirectional effect. Chain-of-Thought reasoning suppresses most of these effects, yet an in-group ally still reduces conformity to an incorrect out-group majority. Labeling peers as safety-aligned shifts overall conformity but leaves in-group favoritism and out-group divergence intact. These results show that group identity shapes how LLMs aggregate information across agents, independently of its correctness, and identify a manipulation surface for multi-agent AI systems.

cs.CL↗

XTC: Head-Aware Sampling by Excluding Top Choices

Standard decoding rules for autoregressive language models promote diversity by rescaling the full next-token distribution or truncating its low-probability tail. These strategies overlook a common regime of open-ended generation in which several continuations are plausible but too much probability mass remains concentrated on the most generic choice. We introduce XTC (Exclude Top Choices), a lightweight head-aware decoding operator that targets this regime directly. XTC identifies tokens whose probabilities exceed an absolute plausibility threshold $τ$: when at least two qualify, it removes the dominant eligible choices with probability $ρ$ and retains only the weakest plausible alternative before renormalization. Across 60 experiments on Gemma 3 27B Q4, Gemma 3 12B Q6, and DeepSeek R1 14B Q6, with scaling validation on Llama 3.3 70B Q4, XTC improves the diversity-repetition Pareto frontier. On creative generation, Distinct-2 increases by 11--15% and repeat trigrams decrease by 27--47% across the four models. Combined with temperature scaling, gains reach 38% in Distinct-2 and 71% in repeat-trigram reduction over baseline. A blinded Amazon Mechanical Turk study with 150 Master raters yields a 62.3% creativity preference for XTC ($p<10^{-4}$) without reduced fluency, while a GPT-4o control judge reproduces the Anthropic-judge direction on every measure. On IFEval with Llama 3.3 70B Q4, XTC preserves prompt-level strict accuracy within 1.7 percentage points of baseline while recovering most of the diversity gain; a temperature setting matched on Distinct-2 reduces IFEval by 8.8 points. The effect is additive with temperature and repetition penalties, robust across quantization levels and model families, and consistent across twelve prompt genres. XTC has been adopted by llama.cpp, ExLlamaV2, and text-generation-webui.

cs.CL↗

Don't Repeat Yourself: Stopping Verbatim Loops at Sampling Time

Large Language Models generate text autoregressively, but open-ended generation is prone to verbatim looping, in which models repeat spans already present in context. Standard defenses such as repetition, presence, and frequency penalties and n-gram blocking act on token recurrence rather than the sequential structure of a loop, and often suppress looping only at strengths that also degrade formatting or fluency. We propose Don't Repeat Yourself (DRY), a sampling-time logit adjustment that penalizes a candidate token only when generating it would extend the current suffix into an exact continuation of a span seen earlier in the context. Sequence breakers protect chat templates and formatting tokens. Across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study, DRY reduces suffix-extension rate by 47% while improving lexical diversity. An intervention-matched placebo produces no comparable reduction, identifying suffix matching as the operative mechanism. On AWQ-quantized 70B and 120B models, DRY reduces loop rate by roughly half while preserving MT-Bench, MMLU, and GSM8K performance, whereas standard alternatives lose measurable ground. DRY has been adopted by popular open-source LLM inference frameworks including llama.cpp, ExLlamaV2, and text-generation-webui, highlighting its practical impact on text generation.

cs.CL↗

From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning

Humans organize knowledge into compact conceptual categories that balance compression with semantic richness. Large Language Models (LLMs) exhibit impressive linguistic abilities, but whether they navigate this same compression-meaning trade-off remains unclear. We apply an Information Bottleneck framework to compare human conceptual structure with embeddings from 40+ LLMs using classic categorization benchmarks. We find that LLMs broadly align with human category boundaries, yet fall short on fine-grained semantic distinctions. Unlike humans, who maintain ``inefficient'' representations that preserve contextual nuance, LLMs aggressively compress, achieving more optimal information-theoretic compression at the cost of semantic richness. Surprisingly, encoder models outperform much larger decoder models in human alignment, suggesting that understanding and generation rely on distinct representational mechanisms. Training-dynamics analysis reveals a two-phase trajectory: rapid initial concept formation followed by architectural reorganization, during which semantic processing migrates from deep to mid-network layers as the model discovers increasingly efficient, sparser encodings. These divergent strategies, where LLMs optimize for compression and humans for adaptive utility, reveal fundamental differences between artificial and natural intelligence. This highlights the need for models that preserve the conceptual ``inefficiencies'' essential for human-like understanding.

cs.CL↗

OpenDebateEvidence: A Massive-Scale Argument Mining and Summarization Dataset

We introduce OpenDebateEvidence, a comprehensive dataset for argument mining and summarization sourced from the American Competitive Debate community. This dataset includes over 3.5 million documents with rich metadata, making it one of the most extensive collections of debate evidence. OpenDebateEvidence captures the complexity of arguments in high school and college debates, providing valuable resources for training and evaluation. Our extensive experiments demonstrate the efficacy of fine-tuning state-of-the-art large language models for argumentative abstractive summarization across various methods, models, and datasets. By providing this comprehensive resource, we aim to advance computational argumentation and support practical applications for debaters, educators, and researchers. OpenDebateEvidence is publicly available to support further research and innovation in computational argumentation. Access it here: https://huggingface.co/datasets/Hellisotherpeople/OpenDebateEvidence-Anonymized

cs.CL↗

Measure what Matters: Psychometric Evaluation of AI with Situational Judgment Tests

Persona conditioning is widely used to steer large language model (LLM) behavior, but it is unclear whether it induces stable behavioral structure or superficial variation. We propose a framework to measure consistent behavioral tendencies using situational judgment tests (SJTs), multidimensional item response theory (MIRT), and structured synthetic personas, treating responses as observations of latent behavioral variables. Across large-scale SJT and persona datasets, we find that persona-conditioned behaviors are stable across runs, latent trait scores predict external benchmarks (e.g., TruthfulQA, EmoBench), and MIRT reveals consistent latent structure. We validate these results through human annotation, benchmark evaluation, and internal consistency analyses. We interpret these traits not as human personality, but as stable behavioral tendencies expressed across contexts. Our results show that scenario-based psychometric evaluation provides a more reliable alternative to classical self-report approaches for assessing LLM behavior, and we release datasets to support further study.

cs.AI↗

S-JEPA : Soft Clustering Anchors for Self-Supervised Speech Representation Learning

Self-supervised speech encoders are predominantly trained by predicting discrete hard cluster IDs at masked positions, a recipe that collapses acoustic ambiguity at category boundaries and requires interrupting training to re-cluster the entire corpus between iterations. We introduce S-JEPA, a JEPA-style encoder-predictor pair trained to match the soft posteriors of a Gaussian Mixture Model at masked positions via KL divergence. Training runs as one continuous optimization trajectory in two phases: a fixed GMM over MFCC features, then an online GMM over encoder features, with the input layer selected adaptively from a label-free signal, removing both the offline re-cluster step and the hand-tuned choice of which transformer layer to cluster on. Under the SUPERB protocol, S-JEPA achieves the lowest WER among evaluated SSL methods below 90M parameters and matches HuBERT-Base on emotion recognition at roughly half its parameter count, establishing a new Pareto frontier without offline re-clustering or teacher distillation. An analysis of the predictor's per-frame entropy on held-out speech reveals a bimodal distribution with a substantial minority of frames near the entropy of a perfect two-cluster tie, providing direct empirical evidence that the soft-target objective preserves the acoustic ambiguity that hard targets would collapse. Code is available at https://github.com/gioannides/s-jepa.

cs.SD↗

Mirage Probes: How Vision Models Fake Visual Understanding

Vision-language models (VLMs) can answer image-based questions confidently, and often correctly, even when no image is provided. This mirage behavior inflates benchmark scores without reflecting visual grounding. Prior work treats this as a single failure mode. We argue it is two. Using Mirage Probes, a contrastive probing framework that pairs paraphrased question variants with matched mirage and non-mirage labels on the same image, we show that mirage behavior is linearly decodable from internal activations across residual stream, MLP, post-attention, and attention-head sites in two open-source VLMs. We demonstrate that a Naive Bayes text baseline cannot recover this signal, ruling out surface lexical confounds. Cross-benchmark separability patterns, together with a novel Prior Harnessing Index (PHI) measuring how much a model can answer from text alone, expose two distinct regimes: textual biases, where the model answers from language priors without engaging visual representations, and spurious images, where it constructs false visual content in latent space and answers as if grounded. The distinction has direct mitigation consequences: text-distribution cleaning can address the first regime but cannot reach the second, since spurious-image mirages live in the model's visual representations rather than its text. Faithful visual grounding will require interventions at the representational level.

cs.CV↗

On Training in Imagination

State-of-the-art model-based reinforcement learning methods train policies on imagined rollouts. These rollouts are trajectories generated by a learned dynamics model and are scored by a learned reward model, but without querying the true environment during policy updates. We study this training paradigm by quantifying how errors in learned dynamics and reward models affect returns and policy optimization. First, we extend the analysis of Asadi et al. (2018) to MDPs with learned reward models, and derive the optimal sample allocation--the ratio of dynamics samples to reward samples that minimizes a bound on return error under power-law scaling assumptions. We identify lower Lipschitz constants of the learned dynamics, reward, and policy as a representation desideratum that tightens this bound, and we connect this perspective to the temporal-straightening objective of Wang et al. (2026). Second, we examine how policy optimization with REINFORCE tolerates noisy rewards, which are often cheaper to obtain. We show that zero-mean reward noise leaves the gradient estimator unbiased and adds at most a variance term that decreases with the number of rollouts. This introduces a practical tradeoff: given a fixed budget, should one buy more rollouts with cheaper but noisier rewards, or fewer rollouts with more expensive but less noisy rewards? We reduce this choice to a one-dimensional optimization problem and characterize the optimum.

cs.LG↗

UAT-LITE: Inference-Time Uncertainty-Aware Attention for Pretrained Transformers

Neural NLP models are often miscalibrated and overconfident, assigning high confidence to incorrect predictions and failing to express uncertainty during internal evidence aggregation. This undermines selective prediction and high-stakes deployment. Post-hoc calibration methods adjust output probabilities but leave internal computation unchanged, while ensemble and Bayesian approaches improve uncertainty at substantial training or storage cost. We propose UAT-LITE, an inference-time framework that makes self-attention uncertainty-aware via Monte Carlo dropout in pretrained transformer classifiers. Unlike output-level calibration (e.g., TS), UAT-LITE injects epistemic uncertainty directly into attention, enabling uncertainty-aware routing during contextualization and token-level diagnostic signals beyond global logit rescaling. Token-level epistemic uncertainty is estimated from stochastic forward passes and used to modulate self-attention during contextualization, without modifying pretrained weights or training objectives. We additionally introduce a layer-wise variance decomposition to diagnose how predictive uncertainty accumulates across transformer depth. Across SQuAD 2.0 answerability, MNLI, and SST-2, UAT-LITE achieves an average relative ECE reduction of approximately 20% compared with a fine-tuned BERT-base baseline while preserving accuracy, and yields more informative uncertainty behavior for selective prediction under distribution shift.

cs.AI↗

Latent Transfer Attack: Adversarial Examples via Generative Latent Spaces

Adversarial attacks are a central tool for probing the robustness of modern vision models, yet most methods optimize perturbations directly in pixel space under $\ell_\infty$ or $\ell_2$ constraints. While effective in white-box settings, pixel-space optimization often produces high-frequency, texture-like noise that is brittle to common preprocessing (e.g., resizing and cropping) and transfers poorly across architectures. We propose $\textbf{LTA}$ ($\textbf{L}$atent $\textbf{T}$ransfer $\textbf{A}$ttack), a transfer-based attack that instead optimizes perturbations in the latent space of a pretrained Stable Diffusion VAE. Given a clean image, we encode it into a latent code and optimize the latent representation to maximize a surrogate classifier loss, while softly enforcing a pixel-space $\ell_\infty$ budget after decoding. To improve robustness to resolution mismatch and standard input pipelines, we incorporate Expectation Over Transformations (EOT) via randomized resizing, interpolation, and cropping, and apply periodic latent Gaussian smoothing to suppress emerging artifacts and stabilize optimization. Across a suite of CNN and vision-transformer targets, LTA achieves strong transfer attack success while producing spatially coherent, predominantly low-frequency perturbations that differ qualitatively from pixel-space baselines and occupy a distinct point in the transfer-quality trade-off. Our results highlight pretrained generative latent spaces as an effective and structured domain for adversarial optimization, bridging robustness evaluation with modern generative priors.

cs.CV↗