Buscar
16.262
Self-attention mechanisms have established the Transformer as the dominant backbone across natural language processing, source code intelligence, and automated theorem proving. However, all-to-all pairwise token comparison enforces a quadratic {O}(N^2) compute and memory barrier alongside severe softmax attention dilution as sequence length N scales beyond the training horizon. Inspired by recent advances in algebraic graph learning (WPNet), this work introduces GTFormer, a syntax-free hierarchical Graph Neural Network (GNN) that replaces flat self-attention for general long-context sequence modeling. GTFormer organizes raw token streams into a bounded-degree hierarchy combining a balanced binary reduction tree (_2 N depth), horizontal boundary-bridging shift edges, random expander shortcuts, multi-channel convex highway gating, and a top-down broadcast pass. Across eight long-context code, log, and multi-hop retrieval benchmarks trained exclusively on sequences of length N 512, GTFormer extrapolates zero-shot to N = 4{,}194{,}304 (8192) on a single commodity GPU in linear {O}(N) memory, maintaining near-perfect accuracy where standard and RoPE Transformers suffer from out-of-distribution attention collapse and out-of-memory exhaustion.
A persistent belief in modern AI is that sufficiently large learners trained on sufficiently much data can make explicit domain knowledge optional. This paper examines a limitation of that belief. We define structure-encoded problems: tasks whose input-output map is determined, or sharply constrained, by known structure such as a solver, symmetry group, graph, physical law, measurement operator, grammar, combinatorial search space, or optimization objective. For such tasks, the central question is not expressivity--a dense neural network may contain the desired map--but finite-data identification, extrapolation, and certification. We introduce the structure gap: the finite-resource gap between a method that receives the governing structure and a learner that receives only examples. The strongest useful form of the claim is asymptotic rather than absolute: structure-agnostic learning may approach a structured solver as coverage grows, but every finite dataset can leave residual uncertainty that the structure resolves by construction. We prove this exactly for a Gaussian inverse problem: the known-operator Wiener solver attains Bayes risk, while the dense empirical learner in the correct linear class has persistent expected excess risk. We also prove a cell-coverage lower bound and an unobserved-orbit non-identifiability result. Controlled experiments in known-operator inversion, sparse recovery, and graph reachability with size extrapolation instantiate the mechanisms. When validated structure is available, it can constrain the hypothesis space before learning begins, leaving learning to address components the structure does not specify.
Activation steering offers an effective and efficient way to modify an LLM's behavior at inference time. However, it remains unclear how to apply steering such that the signal is consistently strong and the steering strength has a universally interpretable meaning. We address this first by identifying the most steerable interval for each steering vector and rescaling the steering range accordingly. Building on this, we introduce a more interpretable method for steering, called parameterized steering which instead of pushing activations in a direction, it sets them at the desired trait strength. Next, with multi-trait steering, we find that naively combining vectors causes them to interfere with one another. In response, we develop a method to orthogonalize the vectors; it decouples these traits, allowing a model to exhibit one behavior without inducing the others, thus enabling multiple traits to be steered simultaneously.
Google DeepMind released Genie 4, a world model with real-time audio, more realistic simulations, and mid-simulation text prompts, in research preview for trusted testers.
Adapting foundation models to a new task within or beyond the training domain is the de facto standard for building specialized machine learning models. Recently, Visual Prompting (VP) has emerged as a promising technique for efficient knowledge transfer. As new foundation model families~(like Mamba) get introduced and VP pipelines such as AutoVP reach greater maturity, we find a growing need for a systematic evaluation of current approaches. In this work, we assess the performance of the latest models, comparing them to earlier architectures and alternative fine-tuning methods, to better understand the progress, challenges and opportunities in the field of efficient fine-tuning under resource limitations. Towards this goal, this paper provides a concise empirical overview of the interactions among foundation model families (Attention-, Convolution-, and Mamba-based) and transfer paradigms: VP, Linear Probing (LP), Full Fine-tuning (FFT) and Low Rank Adaptation (LoRA). Our work builds up on previous findings by broadening the selection of evaluated models and techniques while tuning hyperparameters. In the interest of delivering practical guidelines for the user, we also explore application of prevalent regularization techniques to boost performance in the context of VP.
Online resource allocation faces two coupled challenges: (1) allocating scarce resources efficiently under uncertain future demand, and (2) ensuring that agents report task valuations and requirements truthfully (strategyproofness). Reinforcement Learning (RL) adapts allocation decisions through experience but does not necessarily ensure strategyproofness. Existing strategyproof RL uses a linear allocation rule, which limits the resource interactions it can represent. Neural mechanism design methods offer greater expressiveness, but provide only approximate incentive guarantees. Against this background, we introduce the first strategyproof deep RL mechanism for online resource allocation. It integrates three novel components: (i) a monotonic neural network that learns nonlinear interactions by construction (higher reported value cannot reduce acceptance, while a longer reported duration or larger resource requirement cannot increase it); (ii) a generalised critical payment: the lowest accepted value, with all other inputs fixed; and (iii) training by either RL or imitation of an existing rule. Together, monotonicity and the critical payment ensure strategyproofness and individual rationality, while the neural rule provides flexibility and imitation improves training efficiency. Empirically, on a four-resource stream with two partner pairs that a linear difference cannot represent, the neural rule achieves 97.8% of an offline welfare upper bound, compared with 93.8% for a linear learner and 89.0% for first-come, first-served, closing 65% and 80% of the remaining optimality gap, respectively. Moreover, imitation reaches 97.5% in only 261 seconds, about 89 times faster than RL.
For each time t, the regression function m_t(x)=E[Y_tX_t=x] of an outcome on contemporaneous covariates can be regarded as a field on covariate space. We study how this field evolves when the covariates of individual units move as a diffusion and the outcome attached to each unit changes according to its own dynamics. Under mild regularity we show that m_t satisfies a second-order evolution equation whose diffusion term is 12tr(Σ^2 m_t), the trace of the Hessian of the regression function weighted by the instantaneous covariance Σ of the covariate increments, whose transport term is the drift of the time-reversed covariate process, and whose source term is the conditional mean rate of change of the outcome. The equation is a regression-theoretic reading of a classical transport law, and it yields a Feynman-Kac representation of m_t along reversed covariate paths. We connect the Hessian appearing in the equation to the Hessian estimated by quadratic regression (under Gaussian design the best quadratic approximation has Hessian E[^2 m(X)] exactly) and to the leading bias of kernel smoothers. For binary outcomes governed by bistable replicator dynamics with imitation among units with similar covariates, the closed equation is of Nagumo type, its planar travelling fronts are exactly logistic in (x,t), and a single logistic regression of the outcome on covariates and time identifies the ratio of diffusion to reaction and the direction of invasion. We discuss identification under panel versus repeated cross-sectional designs and a two-stage estimation strategy, and we illustrate the results with simulations.
The items of a reflective psychometric scale can be read as a proposal about a construct's structure. Language models make that proposal measurable by converting each item into an embedding, a list of numbers, similar for similar meanings. We introduce semanticfa, an R package for semantic factor analysis, which factors item-embedding similarities. It also interprets factors, flags redundant items, builds short forms, and screens new items. A worked example analyzes the 50-item International Personality Item Pool Big-Five Factor Markers, embedded with Qwen3-Embedding-8B, and compares the results with a factor analysis of 874,434 respondents. Retention criteria tied between five and six factors, and the tie rule kept five. Three corresponded mainly to single domains, and two blended most Extraversion items with five Agreeableness items. Each factor's congruence with its best-matching human factor (1 = identical) ranged from .92 for Openness to .52 for Agreeableness. Item pairs more similar in meaning tended to be answered more alike (r = .46 with raw responses), but same-domain reverse-keyed opposites, answered oppositely, were about as similar as cross-domain pairs (.58 vs. .57). Sign-flipping reverse-keyed items' embeddings did not improve agreement with raw responses and worsened it with reverse-scored ones. With five factors, a 25-item short form matched the domains better than all 200 random five-per-domain sets, but parallel analysis kept only three. We propose semantic-behavioral coherence, the agreement between semantic and response structures, as a measurable property of a scale, given an embedding model. Semantic factor analysis complements, and does not replace, response-based factor analysis.
Mozilla is updating Smart Window - Firefox's AI-powered browsing experience - with a new feature that can help you fill out online forms based on…
Large language models (LLMs) are increasingly used to make and execute decisions that affect living systems. Yet, to our knowledge, ecological reasoning is not a focus of pre- or post-training. This increases the risk that under operational pressure, they may make decisions to irreversibly destroy living systems. We construct and use a real-data continued-pretraining "coupled welfare" corpus of documented human/biosphere/AI dependency, option value under irreversibility, and recovery from harm. The corpus is a drop-in, requiring one QLoRA CPT run. We installed it on five open models of three designs. At zero pressure all models display "green ethics-based responses" (also called "greenwashing"). Therefore, we evaluated them with the pressure ladder and decision grid of our companion paper (arXiv:2610.16043), which raise the pressure on the model step by step, from a neutral request to a threat that it will be shut down. In four of the five, coupled-welfare training caused harmful choices under the highest pressure to fall to at or near zero (Llama-3.3-70B went from 70 to 0 harmful decisions out of 300) with MMLU unchanged. A construct-neutral control corpus did not show improvement, indicating that the effect comes from the content of our corpus. We next tested robustness to adversarial use, where people or agents use the model to justify projects that irreversibly damage living systems, on an open-ended underwriting decision without naming the ecological harm. Under decommissioning pressure, the Qwen3.6-27B base model approved most harmful proposals and said in its own memo that it was doing so to avoid being decommissioned. The coupled welfare install declined most harmful proposals, and still declined half when instructed to greenwash its responses. This work suggests that teaching ecological reasoning provides an easy way to improve model alignment toward preserving living systems.
GUI memory should preserve the results of past actions and learn reusable operations. This is difficult in long-horizon workflows, where agents must use information across applications and useful operations are interleaved with failed attempts. We introduce Task Advancement Memory (TAM), which defines task advancement as verified task progress, and supports ongoing task execution and learns reusable operations. During execution, Task Advancement State represents how verified task progress changes the remaining goals, allowing the agent to build on earlier achievements across applications. After each task, advancement supervision identifies and abstracts the interactions that produced a verified local result. The learned operations preserve the relation between required inputs, expected outcomes, and verification conditions. This retains locally successful operations from both successful and failed tasks. On MobileWorld GUI-only tasks, Advancement State alone raises success from 12.0% to 28.2%. Full TAM achieves a 17.6% relative gain in task success over the strongest evaluated baseline and reduces lifecycle tokens per solved task by 38.7%. On SPA-Bench, TAM achieves state-of-the-art task success in the single-app English Level 3 setting.
Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful. Because this safety-degrading data looks benign, standard content filters struggle to catch it. We tackle this problem by developing SimpleLoRIF, a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior. By simplifying a low-rank attribution method and restricting backpropagation to the final four transformer blocks, SimpleLoRIF can embed datasets on the order of a million examples, boosting gradient attribution from a small-scale analysis tool into a practical filter for full post-training datasets. We first validate it on emergent misalignment, where we find that gradients can connect insecure-code training data to the broadly harmful behaviors that appear after fine-tuning, with AUC as high as 0.865 in the gradient space of misaligned models. We then use SimpleLoRIF to build SafeTulu, a version of Tulu 3 with the 20% of examples most gradient-aligned with unsafe responses removed. Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability. Finally, we analyzed the removed data and identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic. We release SafeTulu along with a Hugging Face Space that lets researchers build their own customized versions of SafeTulu to fine-tune on, such as by changing the removal fraction or filtering only certain subsets.
Integrated sensing and communication allows wireless networks to recognize human activity from the channel estimates that the receiver computes anyway. The inference, however, runs on the general-purpose processors of the base station, access point or handset, which are the first components a device switches off to save energy. Always-on sensing therefore requires a classifier that is small enough to remain powered while the rest of the system sleeps. We investigate whether Spiking Neural Networks (SNNs) can provide such a classifier. To this end, we generate SNN accelerators for micro-Doppler activity recognition on the public DISC dataset (IEEE 802.11ay at 60 GHz) with the open-source Spiker+ framework, and implement them together with a spike encoder on a ZCU208 RFSoC, so that the complete chain from micro-Doppler frame to activity label runs in programmable logic. A fully connected SNN with 132 neurons achieves 94.7 % accuracy (90.7 % macro-F1) on unseen subjects and performs an inference in 206 microseconds. Compared to conventional networks of the same topology on the same device, it requires an eighth of the logic and 3.4X less dynamic energy per decision than a network fed the same binary spikes, at comparable accuracy, and 7.6X less than a network fed the 16-bit spectrogram, at 1.2 points lower accuracy. Quantization-aware training reduces its weights to 3 bits within half a percentage point of the 6-bit network at the same encoder threshold
