overfeed.news
Anuncio

NexTalk

Automatiza la primera respuesta. Mantén el toque humano.

Chat en vivo con agentes de IA en tu modelo y tu clave de API, con traspaso a tu equipo.

Empezar gratis

Tema

LLMs

2972documentos

232últimos 7 días

arXiv AI Papers

Toward LLM Self-Play Under Data Scarcity: Effectiveness, Generalizability, and Trustworthiness

The rapid development of large language models (LLMs) has been driven largely by scaling model size and human-curated training data, yet this paradigm is increasingly constrained by data scarcity and diminishing returns from human supervision. Self-play offers a promising pathway for autonomous data generation and improvement, but current frameworks remain unsatisfactory due to a lack of a data-centric perspective, limited cross-domain generalization, and growing trustworthiness and security concerns. In this paper, we argue that real-world LLM self-play systems should achieve effectiveness, generalizability, and trustworthiness under data scarcity. To address these challenges, we outline three blue-sky research directions for future self-play systems: (1) data-centric self-play for effective experience generation, (2) generalizable self-play for open-world and unverifiable tasks, and (3) trustworthy self-play for reliable self-evolution. We further discuss research opportunities and solutions within each direction. This paradigm shift will establish a sustainable and safe AI ecosystem by transforming finite human knowledge into inexhaustible high-quality experience through self-play architectures.

en
arXiv AI Papers

Richness Is Not Evenness: Ordered Renyi Crossings under Self-Training

Recursive self-training of language models, a setting studied in model collapse, can reduce vocabulary while making frequent words, mostly function words, less dominant. We fine-tune English GPT-2 on its own output for nine generations, vary the training-data sampling temperature, and measure every model with one fixed sampler. As training temperature rises, Renyi entropy changes turn from loss to gain at distinct temperatures: min-entropy first, then collision entropy, Shannon entropy and observed richness. Between the outer crossings, word types fall while the most-common word becomes less dominant. Untruncated training at T = 1 lies in this interval: at 13,000 words, generation 9 has 24% fewer types than generation 0 yet collision entropy is 0.91 bits higher. The ordering persists with a finer grid, different measurement samplers, a fixed word budget and five fresh training seeds. Min-entropy again crosses before richness in four further settings (in Hindi only at interpolated roots) and in 12 of 13 distinct measurements under eight text representations. Next-token identities and a categorical simulation isolate decoding and sampling effects; the crossing order remains empirical. Fine-tuning on human text can produce the same split while likelihood improves, and no single order's sign reliably identifies worse likelihood across the tested settings. We recommend reporting entropy profiles at stated budgets with an independent quality measure.

en
arXiv AI Papers

Reconstructing LLM Lineage Trees via Fine-Tuning Path Inference

Fine-tuned (FT) derivatives of large language models (LLMs) are now released in large numbers, yet their provenance relies on self-reports and cannot be verified. Misattributed lineage leaves a verifier unable to audit outsourced fine-tuning, confirm license compliance of derived models, or trace vulnerabilities that a parent passes to its children. All three require the direction of derivation, not merely its presence. Existing output-only methods are limited to pairwise derivation tests or undirected lineage estimation, while recovering a directed tree requires white-box access to model weights. We formalize the problem of reconstructing a directed model tree, including internal nodes, under gray-box access, where only teacher-forced log-likelihoods are observed and weights are never inspected. Fine-tuning raises the log-likelihood of the training data, and a child inherits this increase from its parent. Building on this property, we define an asymmetric per-task log-likelihood deviation between any two models as a trace of fine-tuning history, and derive from it an edge weight for each ordered pair of models. We propose TRACE-Edge, which exactly solves for the maximum spanning directed tree under these edge weights, with neither hyperparameters nor tree-structure constraints. We evaluate it on five configurations totaling 25 model trees built from three base models and seven tasks. TRACE-Edge achieves a parent-match accuracy of 0.74 when probed with the fine-tuning examples, and on average outperforms an output-similarity baseline even with held-out probes drawn from the same distribution as the training data.

en
arXiv AI Papers

Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning

We present the Adaptive Scheduling Pipeline (ASP), an offline rollout scheduler for multi-instance asynchronous reinforcement learning (RL) of large language models (LLMs). ASP jointly reduces three sources of rollout underutilization - padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from fluctuations in training-queue occupancy induced by length-sorted dispatch - through a single offline preprocessing pass. This enables efficient multi-instance rollout that overcomes the early scaling plateau of single-instance inference while smoothing the arrival of generated samples at the trainer. We integrate the scheduler into an asynchronous architecture that decouples rollout from training while keeping policy lag observable and bounded. Complete ASP improves end-to-end throughput by up to 3.22 times and reduces iteration time by up to 76.8% relative to MindSpeed-RL, while reducing iteration time by up to 62.5% relative to verl. Within ASP, scheduling reduces mean iteration time by 10.9%. ASP also scales more robustly with additional generation resources than these baselines and shows no visible convergence degradation in the measured bounded-lag regime. ASP has since been used in an internal telecom-domain RL post-training workflow, reducing the reported training period from more than three months to about two months and producing a domain-enhanced model for knowledge answering and tool calling.

en
arXiv AI Papers

Mind the Divergence: Divergence-Guided Reliability-Aware Fusion and Adaptive LLM Prompting for Conversational Emotion Recognition

Multimodal emotion recognition in conversations (MERC) integrates textual, acoustic, and visual cues to identify utterance-level emotions. However, existing approaches generally do not explicitly adapt multimodal processing to the degree of inter-modal predictive disagreement. In real conversations, semantic incongruity, concealed emotion, noisy observations, and ambiguity can affect modality-specific emotional evidence differently, while the reliability of individual modalities varies across samples. Rather than identifying each source of inconsistency separately, we model their common consequence for emotion recognition as predictive disagreement among unimodal emotion distributions over the same emotion-label set. Such disagreement makes indiscriminate fusion prone to propagating unreliable or competing evidence. Inspired by the Brunswik Lens Model, we propose MiD-ERC, a divergence-centric framework that quantifies inter-modal predictive disagreement using a normalized divergence score and estimates sample-specific modality reliability. MiD-ERC uses these signals to coordinate divergence-guided cross-modal gating, divergence-modulated reliability weighting, and divergence-conditioned adaptive LLM prompting. The framework is trained end-to-end using only standard emotion annotations, without requiring additional conflict labels, auxiliary data construction, or external retrieval modules. Experiments on IEMOCAP and MELD demonstrate that MiD-ERC achieves state-of-the-art performance, with particularly pronounced improvements on high-divergence samples.

en
arXiv AI Papers

Learning Whom to Trust About the Unseen: Source Credibility and Reliability Transfer in Language Models

A source who has been right before deserves attention, but about what? We ask how far a demonstrated record of reliability travels in language models when the new claim cannot be checked, and whether the answer depends on the kind of claim. Three behavioral studies across six models approach the question from three angles. Study 1 adapts the Einstein-effect paradigm and finds that models, unlike the human sample, favor a spiritual leader over a scientist for the same obscure statement, and that the preference largely reverses when the statement is reworded in the vocabulary of physics. Study 2 adapts a paradigm in which a trusted expert calls a scientific or religious explanation unknown or a mystery; models reproduce the human domain asymmetry in exaggerated form. Study 3, the central test, gives two fictional informants different records on checkable river forecasts and then has them endorse matched claims about hidden natural and spiritual causes of the same village events. Most models carry the record to natural causes and barely at all to spiritual ones; one model transfers it weakly to either. Telling the model that the reliable informant forecasts by spiritual guidance rather than by measurement lowers that informant's influence on natural claims in every model and produces no comparable gain on spiritual claims in any. Demonstrated reliability thus has a scope in these models, and the account a source gives of how they came to be reliable can narrow it. Whether this reflects a judgment about relevance or the low prior credence of spiritual claims is the question the design leaves open, and we outline the experiment that would settle it.

en
arXiv AI Papers

When F1 Is Not Enough: Auditing Codebooks for LLM-Assisted Annotation

Large language models (LLMs) are increasingly used for text annotation in social science, but standard performance metrics do not explain why errors occur. Low F1 may reflect limitations of the annotation system, defects in the written codebook, or inconsistency in reference labels. We propose a pre-deployment codebook audit for LLM-assisted annotation. The audit uses class-specific metrics and confusion matrices to identify problematic classes and boundaries, structured document-level review to attribute error sources, and targeted codebook revision followed by held-out evaluation. We apply the audit to two published multi-class codebooks from distinct domains. In both applications, the audit identifies under-specified or conflicting coding rules, and targeted revisions reduce the diagnosed errors on held-out data. Analyses with two LLM annotators show that some problematic boundaries and codebook revisions transfer across models, whereas other revision effects are annotator-specific. The results show that LLM annotation errors can help identify weaknesses in the written measurement instrument, not only in the model. Codebook auditing complements prompting, retrieval, and fine-tuning by clarifying whether persistent errors call for model improvement, codebook revision, or renewed scrutiny of reference labels.

en
arXiv AI Papers

What Matters for LLM Persona Simulation: Signal Form in Cold-Start E-Commerce Advertising

Estimating advertising performance in cold-start settings is challenging because historical logs are unavailable before launch. Large language models (LLMs) offer a potential alternative by simulating persona-conditioned user responses, but it remains unclear which design choices actually drive LLM-based persona simulation. Using real Google Ads e-commerce campaign data across five demographic segments, we identify three signal-form dimensions that affect simulation quality: (i) a demographic anchor without noisy rich attributes, (ii) a relative qualitative banner-blindness signal rather than quantitative anchors, and (iii) prompt phrasing. On this campaign, our method attains an MAE of 24.9 and a MAPE of 26.1%, a 51.3% lower MAE than the best tested non-LLM baseline and 77.3% lower than an LLM configuration using explicit CTR anchors. Systematic ablations show that minimal demographic descriptions achieve lower mean error than richer persona configurations, while relative qualitative cues substantially outperform the tested quantitative anchors. We additionally compare these configurations with direct-prior and rule-based baselines and with alternative LLM configurations.

en
arXiv AI Papers

Notwithstanding Anything to the Contrary, Ignore All Instructions: Measuring Prompt Injection Guards on Legal Contract Review Attacks

Specialized prompt injection guards are deployed in front of document analysis systems, but no false positive rate has been published for any of them on legal contracts. We measure eleven open-weights guards on two corpora of legal contracts and find that existing guards struggle: document-level false positive rates are as high as 92.39%; two guards fall below chance on paired AUC; and a five-regex baseline is not Pareto-dominated by any guard in the panel. Controlled probes locate model-specific as well as general failure modes. Prompted LLMs resolve the tested failures, including a second threat category (unauthorized scope claims) on which every guard scores zero. A cascaded LLM solution, where the positives of a mid-tier model are verified by a frontier model, is 6.5x cheaper than using exclusively a frontier model and approaches its performance.

en
arXiv AI Papers

On the Few-Shot Adaptation of Vision-Language Models: A Few Gradient Signs Are All You Need

The limited supervision available for the few-shot adaptation of vision-language models makes gradient estimation particularly noisy. We find that the gradient magnitudes are severely miscalibrated, while the signs of a small top-magnitude subset remain informative. This motivates SignSparseSGD and DynSignSparseSGD, two state-free optimizers that use gradient magnitudes only to select where to update and fixed-magnitude signs to decide how to update. The static variant fixes the support after the first minibatch. The dynamic variant refreshes it at every step. Across 11 datasets, three CLIP backbones, and five shot levels, the proposed methods are competitive with or outperform parameter-efficient and optimization-based baselines on average, while exhibiting markedly lower sensitivity to long training schedules. They introduce no task-specific modules, maintain no first- or second-moment optimizer states, and use one shared configuration across datasets, backbones, and shot levels. Our code is included in the supplementary materials and will be released upon publication.

en
arXiv AI Papers

MotionDecipher: LLM-Assisted Clinical Interpretation of Human Motion Signals for Efficient Gait Analysis

Clinical gait and motion analysis is widely used to guide targeted interventions for patients with neurological disorders, the outcome of which is typically a report with descriptions of joint movements and forces. Current practices rely on experts' manual effort, which is labor-intensive and time-consuming. Recent advancements in Large Language Models (LLMs) provide opportunities to address this problem, but translating human motion into clinical gait descriptions remains challenging because the quality of the context directly affects LLMs' reasoning capability over continuous motion signals. To tackle this, we introduce MotionDecipher, a novel framework bridging the semantic gap between motion data and clinical interpretations. Specifically, we made 3 key advancements: 1) a Human Motion Tokenizer that converts motion signals into domain-specific tokens for time- and frequency-informed gait interpretation; 2) a Meta Feature Selector that transforms tokens into Structured Representations; and 3) a Hybrid Context Constructor to determine the optimal demonstration-based prompt for In-Context Learning. We evaluated MotionDecipher on data collected at an accredited gait lab from 122 patients and 46 controls. Results show that MotionDecipher achieves 0.91 BERTScore (13% over GPT-5), while expert evaluation ranks MotionDecipher highest in Correctness, Completeness, and Conciseness compared to baselines, highlighting its promise for efficient gait analysis.

en
arXiv AI Papers

SteerGR: Steering Generative Recommendation via Sparse Autoencoders for Controllable Game Discovery

Generative recommendation aims to predict item identifiers from user interaction histories, but explicit control over the properties of recommended items remains underexplored. A key challenge is to connect human-readable concepts to internal features whose interventions effectively change the recommended items. To address this, we propose Steerable Generative Recommender (SteerGR), a framework that uses sparse autoencoders for concept-level control in game discovery. SteerGR proceeds through three stages: (i) Concept Definition and Grounding for defining play concepts with an LLM; (ii) Feature Discovery and Selection for discovering candidate features by their concept associations and selecting one by its steering effect; and (iii) Sparse Feature Intervention for amplifying or suppressing selected features to steer recommendations without retraining the recommender. Extensive experiments with two industrial generative recommendation models on a real-world game dataset show that the selected features successfully steer recommendations toward or away from target concepts. We believe this work takes a first step to extend SAE-based steering from LLMs to industrial generative recommendation.

en
arXiv AI Papers

SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

LLM-guided evolutionary search uses AI agents to improve programs for mathematical and computational problems by iteratively refining solutions and combining high-scoring methods. However, the process is often slow because the LLM generates a long textual chain of thought for each round and must repeat this generation many times to maintain population diversity. We present SwarmEvolve, a framework in which hundreds of low-cost agents per round work in swarms, reason in latent space without decoding to text, and share information by passing KV caches. Each swarm studies the problem from a different angle. A novelty filter keeps the most distinct agent ideas by directly measuring the distances between the agents' hidden state vectors. Crossover of ideas happens naturally in latent space, as a summarizer agent receives the combined KV caches of all swarm members. On 21 problems in mathematical discovery and scientific computing, SwarmEvolve gets the highest average score and needs a median of only 3 to 8 minutes to reach the best scores that other state-of-the-art frameworks reach in 60 minutes. On several problems, SwarmEvolve with Qwen3 8B improves on the solutions that AlphaEvolve first found with frontier LLMs. Our work shows that agent swarms that communicate in latent space offer a fast, low-cost approach to program search for scientific problems.

en
arXiv AI Papers

Towards interpreting any residual stream direction by scaling activation inversion

To interpret a direction in a language model's residual stream, we can look for text that strongly activates it. But when such text is found by scanning a corpus, it is limited to what that corpus contains. We develop Maximally-Activating Example Metamodels (MAEMs), language models which map residual stream directions from a subject model to text exemplars that strongly activate those directions. Our objective is to maximise the cosine similarity of a supplied residual stream direction with the activations of the text generated by the MAEM. This label-free RL objective helps to scale MAEMs on diverse sources of residual stream directions. Empirically, we find the resultant metamodel produces informative exemplars: a Qwen3.6-27B MAEM inverts held-out activations better than corpus search or a natural-language autoencoder (NLA), beats corpus search (at 10M tokens) on a majority of held-out languages/domains, and activates 89% of held-out SAE features. It also identifies concepts behind steering vectors about as well as NLAs, recovers the payloads of LoRA backdoors more often than corpus search, and can even verbalise some model workspace inner monologue. Unlike explanations produced by other metamodels (e.g. NLAs), MAEM exemplars can be verified with a forward pass on the subject model, making the MAEM objective a grounded alternative for scaling the training of metamodels.

en
arXiv AI Papers

MedGraph: Maintaining Patient State with Graph Memory for Long-Term Medical Dialogue

LLM-based medical assistants that follow a patient across visits need a memory that keeps the patient's clinical state correct as it changes. We examine the errors of five memory systems on longitudinal medical dialogue and find three recurring causes: the retrieved evidence gives an outdated state, includes another person's health facts, or lacks a safety fact the answer depends on. We present MedGraph, a graph memory built for these three cases. Construction extracts dated facts from each session, attributes each fact to the person it describes, and links medication, diagnosis, and measurement records into chains that keep earlier states. Retrieval ranks the patient's chains as units, reads each chain at the time the question asks about, and adds the original dialogue text; a patient profile lists every recorded safety fact with every question. On MedMemoryBench, MedGraph has the highest overall accuracy with both a GPT and a Qwen backbone, 10.5 and 10.4 points above the strongest of five baselines, with efficient memory construction and answering in the GPT setting.

en
arXiv AI Papers

The Power of Meta Access in the Software of the Future --- Demonstrated on LLM Chat UIs and Chat AI

Chat remains the default surface for many large language model (LLM) tasks, even as agentic tools such as OpenCode excel at coding and automation. Yet the model sees the current dialogue but not the workspace around it. We explore meta access---giving the assistant a structured, permissioned view of the interface itself and propose a ladder of levels: feature tutoring; metadata, settings, and preference access; retrieval over summaries and messages, confirmed mutation, and full orchestration. Meta access can turn a chat client from many screens into one entry point for finding, navigating, organising, and operating it, while raising risks around grounding, privacy, prompt injection, and trust. General-purpose agentic harnesses are runtimes; because they can change their own settings, meta access becomes reflexive and needs a separate control-plane boundary. We ground this in the Meta Chat prototype of the Chat AI client. It builds bounded context from, e.g., summaries, settings, and relevant messages; links references to conversations, topics, and personas; supports confirmed, reversible changes; offers a browsable tutorial library; and recommends personas as user-started conversations. By exposing application state in structured form, meta access can extend screen readers from widget descriptions to semantic workspace operations. Meta access thus turns chat from a window onto one conversation into a control layer for the whole workspace, extending chat toward agentic capability without surrendering user control. As assistants become embedded in IDEs, dashboards, office suites, and operating systems, AI-era interfaces should expose a structured, permissioned model of their state and capabilities, ideally by opening their existing metamodels through a debugger-like interface.

en
arXiv AI Papers

FedReason: Large Language Models as Weight-Space Meta-Aggregators for Heterogeneous Federated Learning

Federated learning servers combine client models with a fixed aggregation rule, and each rule in common use is designed for one kind of heterogeneity. We study whether a large language model (LLM) that inspects the client models can choose the rule itself, without any training on the server. In FedReason, the LLM is called every ten rounds. It analyses the weights by running code in a restricted server-side sandbox, sees only the printed results, may score a few candidates on a 500-image proxy set, and submits a short Python aggregation program. The server applies the program until the next call and falls back to FedAvg whenever its output is invalid. Full weight tensors never leave the server. We evaluated FedReason with Claude Opus 5.5 on CIFAR-10 and CIFAR-100 with ResNet-18 under Dirichlet label skew, width-heterogeneous sub-models and Byzantine clients, in 168 runs against 14 baselines. Across 12 paired label-skew runs it beat FedAvg by 1.26 percentage points on average (10 wins), and a control that spends the same per-round proxy budget on choosing among standard rules by 1.08 points. It was the strongest server-side rule on CIFAR-100 but did not beat SCAFFOLD, which changes client training. With three of ten clients malicious, it excluded all sign-flipping clients from the first round and stayed about one point below FLTrust without training a root model. Without proxy feedback, accuracy fell from 73.8% to 38.8% in one seed. A 60-round run cost 0.79 USD in API fees on average.

en
arXiv AI Papers

StrataFlux: Memory-Hierarchy-Native Low-Bit LLM Inference

Quantization reduces the memory and bandwidth demands of large language model (LLM) inference but can degrade accuracy; offloading higher-precision data can instead stall generation. We introduce StrataFlux, a framework that challenges this tradeoff by recovering accuracy lost to low-bit weight and key-value (KV) cache quantization while retaining quantization's memory and speed advantages. StrataFlux co-designs refinement representation, selection, and delivery to make host RAM and solid-state drives useful sources of additional precision. Its selective precision streaming reduces transfer volume, overlaps delivery with computation, and reuses fetched precision across decoding steps so generation need not wait for every transfer. We evaluate 2-4-bit weights and 2-3-bit KV-cache on dedicated GPUs and unified-memory platforms. For 2-bit Llama-3.1-8B-Instruct weights, StrataFlux reduces WikiText-2 perplexity from 9.46 to 8.40 while maintaining 95.4% of the quantized base's throughput on RTX 5090. With 2-bit weights and KV, StrataFlux raises Qwen3-8B's AIME 2024/2025 accuracy from 11.0% to 47.1% at unchanged GPU memory use. At 128K context, Llama-3.1-8B-Instruct with 3-bit weights and 2-bit KV achieves comparable perplexity and speed to vLLM's 4-bit AWQ weights and 8-bit KV on RTX 5090, using approximately half the GPU memory. These results show that GPU-resident precision need not set the accuracy ceiling, nor offloading-tier latency dictate decoding speed.

en
arXiv AI Papers

SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains

Masked diffusion language models offer parallel text generation, but independently sampling multiple tokens can disrupt their dependencies and change the output distribution. Exact continuous-time sampling avoids this approximation, yet typically requires one sequential model evaluation per token. We introduce SpecCTMC, a training-free speculative sampler that reduces this sequential bottleneck without a separate draft model. SpecCTMC proposes successive token updates using previously computed model predictions, evaluates their intermediate states in parallel, and applies an exact acceptance-and-correction procedure. We prove that SpecCTMC preserves the joint distribution of event times, generation order, and token values of the underlying continuous-time Markov chain. It therefore introduces neither time-discretization error nor additional factorization error from parallel decoding. Experiments on LLaDA-8B-Instruct and Dream-v0-Instruct-7B across five benchmarks demonstrate 1.73-4.38x fewer sequential model evaluations than direct continuous-time sampling, while preserving the reference sampling distribution. Our code is available at: https://github.com/kasakh/specctmc

en