User embeddings are central components of modern recommendation, personalization, and behavioral modeling systems. Though these embeddings can compactly encode rich behavioral information, they are inherently dense and opaque, limiting interpretability. To address this gap, we argue that user embeddings can be understood as sparse superpositions of interpretable behavioral modes. This structure is directly predicted by generative models of user behavior for linear embeddings and extends approximately to deep-learning representations by the superposition hypothesis. Superimposed modes can naturally be recovered by sparse dictionary learning methods, such as sparse autoencoders (SAEs), and recovered modes can be automatically named and queried in natural language. We validate mode recovery against known ground truth on synthetic data, where SAEs recover modes better than classical sparse baselines. We then apply our methods to classical and transformer-based user embeddings from two real-world datasets. We also show how the corresponding natural-language labels enable zero-shot audience retrieval and characterization over embeddings never trained on text. The framework is post-hoc and model-agnostic, requiring no modification to the upstream embedding system.
Modern video models possess strong visual priors, yet do not natively represent human action in metric 3D space. We present EgoGeneralist, which adapts pretrained video generators to unify action-conditioned video generation, pose estimation, and joint action-video generation for egocentric manipulation. Our role-adaptive action renderer provides video-native representations for multidirectional action-video learning, making metric motion usable as both a spatial control signal and a recoverable generation target. Cross-stream action-video coupling establishes close correspondence between metric motion and visual dynamics through lightweight adaptation while preserving the pretrained visual prior. Together, these designs establish a generalist framework spanning six action-video tasks. With only narrow-domain data, EgoGeneralist generalizes to diverse unseen scenes and interactions with effective action control and visual quality, bridging pretrained visual knowledge and metric 3D human motion.
Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.
Molecular representations such as fingerprints and learned embeddings play a central role in cheminformatics and small-molecule machine learning. They serve as the primary input for a diverse array of tasks, including clustering, dataset splitting, and virtual screening, yet their evaluation is often limited to downstream model performance on narrow benchmark tasks. We carried out a systematic investigation of the quality of molecular representations, finding that none of the state-of-the-art molecular embedding models capture graph structure, 3D shape, or ADMET endpoint similarity better than fingerprints or physicochemical descriptors. However, classical fingerprints also revealed their own material limitations. For example, we show that depending on the representation, between 225M and $\sim$4B molecules have non-unique representations in the 10.8B Enamine REAL database, with implications for virtual screening. Our findings provide an insight into the current limitations of molecular representations, outlining future research directions for the development of more robust encodings of small molecules.
We identify a key limitation of existing feed-forward 3DGS methods: Gaussian attributes are typically predicted in a radial-distance-normalized space that canonicalizes a global radial statistic of the reconstructed geometry, while the spacing between neighboring Gaussian centers can still vary substantially across scene scales and camera configurations. This mismatch induces systematic shifts in Gaussian-scale targets, leading to ambiguous supervision. To address this, we propose SplatAnything with Gaussian Spacing Normalization (GSNorm), which canonicalizes a robust scene-level statistic of Gaussian spacing to provide a more consistent spatial reference for Gaussian prediction. To rigorously assess our method, we further establish a systematic benchmark spanning diverse datasets, input sparsity levels, and challenging extrapolated viewpoints. Extensive experiments demonstrate that SplatAnything achieves state-of-the-art performance, with particularly strong gains under substantial variations in scene scale and viewpoint. Project page: https://zju3dv.github.io/SplatAnything
Automating printed circuit board (PCB) routing requires coordinating board-level connectivity decisions with geometric execution under strict design rules. Although recent agents can interact with native routing tools, reliable routing on complex boards remains challenging. We present PCBRouteAgent, a multimodal PCB routing agent inspired by how human engineers coordinate global strategy with local execution. A Global Routing Planner combines board images and structured electrical and geometric information to specify connection topology, routing order, corridors, and layer transitions. A Detailed Routing Executor realizes these guides through KiCad's native routing engine and performs bounded local recovery. Unresolved failures are returned as structured feedback, enabling targeted guide revisions while preserving accepted routes. We evaluate PCBRouteAgent on 247 synthetic and real-world boards from four PCBWorld-Bench test splits. It achieves clean-pass rates of 100%, 94.9%, 100%, and 40% on D2, D3-A, D3-B, and D3-C, respectively. On the three real-world subsets, PCBRouteAgent outperforms the best evaluated baseline in clean-pass rate by 13.9, 36, and 10 percentage points, respectively.
We study unsupervised industrial anomaly detection in the single-class setting, where a detector learns only from normal images of one product category. Reconstruction-based detectors model normal features from a frozen encoder, but their target layers are usually chosen by architectural convention. We instead ask what information one block can pass to another. Treating depth as a sequence of discrete transitions, we train eight directed, token-wise predictors from block 1 and blocks 21-23 to blocks 7 and 14. Adjacent-layer cosine similarity drops at four repeatable fault lines in DINOv2 ViT-L, separating stable depth segments. The deepest blocks retain semantic context but lose local detail, so the residual of a source-to-sensitive prediction measures information that cannot be transported. The fault structure constrains the source and target roles, and the same rule transfers to ViT-B and RADIO-B. A bounded foreground prior, calibrated only on normal training images, suppresses background residuals while retaining a positive floor for evidence outside the estimated foreground. With the same eight channels and a common training-only calibration procedure in every single-category model, category-averaged image AUROC reaches 99.75 on MVTec-AD, 99.42 on VisA and 94.27 on Real-IAD; object-level AUROC on Real-IAD is 96.26. For the 448x448 runs, category-averaged P-AUROC/P-AUPRO on a common 256x256 evaluation grid is 98.46/95.97 on MVTec-AD, 99.21/96.12 on VisA and 99.26/96.59 on Real-IAD.
Anomaly detection (AD) is often framed as a representation-learning problem, where strong performance is expected to require learned features or AD-specific optimization. However, this view can overlook the intrinsic separability of anomalies from normal references, even in the absence of learned representations. We therefore investigate how far untrained neural networks can go in anomaly detection. To this end, we propose PRISM, a framework combining fixed random nonlinear features with normal-reference Mahalanobis scoring for tabular detection and dense visual localization. Across 47 tabular datasets in ADBench, PRISM-T achieves the highest mean AUROC (86.4%) and the second-highest mean AUPRC (65.9%) among the compared methods. On MVTec AD and VisA, covering 27 visual categories, PRISM-V achieves mean image-level and pixel-level AUROCs of 90.4% and 96.4%, respectively. Notably, both models achieve these results without neural weight optimization or pretrained representations. A theoretical analysis further characterizes how nonlinear transformations affect the mean and covariance contributions to the Mahalanobis score gap, while empirical analyses examine the resulting score separability. These findings establish PRISM as both a competitive anomaly detector and a prior-to-learning baseline for assessing the added value of learned representations in anomaly detection.
Frontier large language models (LLMs) have been reported to read, edit and write the weights of small neural networks, raising the prospect of auditing clinical prediction models from their files alone. Public demonstrations, however, mixed reading raw weights with running code. We separated these conditions in internally pre-specified simulation experiments. Tanh networks of 105 to 38,401 weights were trained on simulated cohorts of patients with subarachnoid haemorrhage, each with a planted relationship or none, and one network was trained on labels that under-recorded deaths in a subgroup. GPT-6 Astra and four other GPT models answered single prompts under a default or an explicit system instruction, either with the weights in the prompt and no tools, or with a calculation-only tool that exposed the weights (white box) or only the network's predictions (black box). Without tools, Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights, and under the default instruction it declined 17 of 21 audit and repair tasks as unidentifiable; the explicit instruction removed these refusals. With the tool, it recovered the planted relationship in all 39 runs on networks of up to 38,401 weights, raised no false alarm in 21 runs on null networks, and found the biased subgroup in 6 of 6 runs. Black-box access performed as well as white-box access, and other GPT-6 models read as accurately. A well-designed script matched the comparator models on the audit and the best LLM repair. For auditing these networks, what mattered was computation and elicitation, not access to the weights.
Online investment services must translate asset opportunities into investment outcomes for users' accounts, yet the same market opportunity can warrant different account actions. We study how supervision from future asset opportunities can strengthen account-level trading decisions and propose Forecast-Augmented Relative Policy Optimization (FRPO). FRPO separately supervises asset opportunities and account decisions through a jointly learned representation: a Financial State Head (FSH) learns future relative-return targets to provide asset-level opportunity supervision, while Same-Path Group-Relative Optimization compares trading trajectories under the same market path and initial account to provide account-level relative-outcome supervision. An Adaptive Gate coordinates auxiliary supervision using predictive variance and gradient alignment. FSH is used only during training; inference uses the encoder and policy for trading. Across DJIA30, TOPIX Core30, and three investment-style portfolios, ten-seed evaluations show that FRPO achieves the highest mean cumulative and annualized returns among the compared methods in all five settings. These results support strengthening investment agents' trading decisions through joint supervision from asset opportunities and trading outcomes.
Articulated 3D assets are essential for robotic simulation and training, yet their generation from casual images poses a key challenge: recovering part geometry and kinematics that remain mutually consistent. We present ArtCrafter, a novel feed-forward generative model for controllable generation of articulated 3D objects from a pair of images, one depicting the object in the rest state and the other in the open state. At the core of ArtCrafter is a DiT-based generative model that jointly produces part-level meshes for both articulation states in a single forward pass. Inspired by Slot Attention, its denoising process operates on a $1+2N$ slot layout comprising one shared base slot and $N$ paired rest/open slots, encouraging each pair to bind to the same kinematic part. To ensure geometrically consistent and disjoint part generation, we introduce Triplet Attention to enforce cross-state geometric consistency between paired rest/open slots anchored through the shared base slot, and a repulsion loss $\mathcal{L}_{\mathrm{rep}}$ on decoded part point clouds to prevent inter-part penetration throughout denoising. Given the resulting paired meshes, URDF joint parameters---type, axis, origin, and range---are derived analytically from the rigid transformation between each rest/open pair, providing an interpretable alternative to learned joint-parameter regression. Extensive experiments on both in-distribution and out-of-distribution datasets demonstrate that ArtCrafter outperforms existing method under the same dual-state setting in geometry quality and articulation precision, while remaining competitive with methods operating under different input settings and paradigms.
Expert pruning compresses Mixture-of-Experts (MoE) language models by removing the experts that a calibration set marks as unimportant. Studies of high-resource languages largely find that English calibration transfers well to other languages. We test this for Lithuanian: we pruned four MoE models with REAP, removing 25% or 50% of the experts with English, Lithuanian or mixed calibration, and evaluated both languages on reading comprehension, translation and language modelling. English calibration does not transfer to Lithuanian: at 25% pruning, English comprehension fell by a median of 1.7 points but Lithuanian by 21.4. The transfer is asymmetric. Lithuanian calibration removed most of the Lithuanian loss and never significantly hurt English comprehension, which was on average 2.5 points better than under English calibration; only English language modeling got slightly worse. Routing suggests why: in the last layers, Lithuanian sends 55-72% of its expert selections to just eight experts, against 19-39% for English, and English calibration removes some of these experts. Compressed models should be calibrated, and evaluated, in every language they must serve.
Existing context models for learned image com- pression rely primarily on internal priors from the current image, overlooking common patterns and typical textures shared across natural images. To address this limitation, we propose an Internal-External Prior Fusion-Alignment (IE-PFA) context model built upon quadtree spatial-channel joint modeling, where a learnable dictionary shared by the encoder and decoder serves as an external prior. Specifically, an RWKV-based Spatial- Channel Prior Fusion (RSCPF) module efficiently fuses inter- nal and external priors along spatial and channel dimensions with linear complexity, while a training-only Self-Distilled Prior Alignment (SDPA) mechanism aligns them without inference overhead. Moreover, our model supports variable-rate coding over 64 quality levels within a single network and is accelerated by custom CUDA operators. Experiments on seven benchmarks demonstrate state-of-the-art rate-distortion performance, achiev- ing an average BD-Rate saving of 21.65% over VTM-22.0-Intra, with only 29.9M parameters and the fastest coding speed among advanced neural codecs (61 ms encoding and 52 ms decoding). Code is available at https://github.com/stcnjtech/DCVC-X-Intra.
Resource closures, arrival delays, and processing-time extensions jointly alter station operations during disruptions. Uncertain reopening times further complicate decisions about holding trains and selecting alternative routes. We use observations of disruption-resolution stages to update conditional predictions of the remaining closure duration. A distributionally robust model accounts for errors in scenario probability estimates and jointly optimizes train routes, sequences, and timings. Logic-based Benders decomposition coordinates routing and resource scheduling. Conditional cuts exploit resource-use patterns, and a graph neural network guides the solution order of resource relaxations. Together, these components form the GNN-LBBD algorithm. Across 28 test instances for a five-track station, GNN-LBBD achieved a mean total completion delay of 61.13 train-min. This was 16.63% lower than tabu search and 20.06% lower than variable neighborhood search. Its mean total computation time was close to that of tabu search. On six representative instances, using stage observations reduced mean total completion delay by approximately 21.7%. Continued updating of reopening-time predictions allows route and entry-time adjustments during closures, reducing delays caused by waiting for reopening.
AI agents are expanding from personal assistance into professional services delivered to external clients on an individual's behalf. We define a personal external service agent as one authorized and managed on an ongoing basis by an individual. It uses that person's knowledge, experience, and professional methods to accept client work, deliver results, and provide follow-up support. Client feedback, professional reputation, and potential income may give individuals more direct incentives to keep developing and improving agents for external services than for personal assistance. Yet building a safe, reliable external service system independently remains difficult, even with AI-assisted programming. People without AI expertise need platforms that are easy to use and safe to operate, with support for creating, developing, and running their agents. We propose the DelegateEval evaluation framework and experimental design to assess these platforms and guide their development. The framework covers eight dimensions: ease of creation and management, service discovery and matching, delivery quality, expertise inheritance, service continuity, learning and improvement, reliability and governance, and efficiency and value. A complete dataset, executable environment, and unified scoring protocol remain to be developed, and empirical validation across foundation models has yet to be conducted.
Vision-language classifiers can learn new labels from text alone, so the label lists they serve keep growing. We show that this growth silently changes conformal decisions about old labels. Common softmax-based conformal scores compare each label with an average over all labels; adding labels changes that average, and with it the score of every old label, even when the model, the image, and the old calibration data stay the same. Recalibrating after the update keeps the coverage guarantee, yet across eleven settings it changes at least one old decision on 53-90% of new images. Coverage cannot reveal this, because it checks only the true label: in five controlled settings that keep the old calibration, so that only the normalizer changes, old-class coverage moves by at most 0.01 while 60-67% of images change. We propose a lifecycle audit that measures these changes alongside coverage and set size, and a simple fix, PrefixCP: each class keeps the label registry that existed when it was added. Its old decisions then never change, each class keeps exact class-conditional coverage under exchangeability, and one prefix scan computes all historical normalizers at once. PrefixCP recovers 40.9-58.5% of the set-size gap between never updating the reference and recalibrating, which gives smaller sets but changes old decisions.
Large Reasoning Models (LRMs) have achieved impressive performance via explicit Chain-of-Thought (CoT) reasoning, yet this process introduces critical safety risks. We formalize LRM safety objectives as requiring non-malicious reasoning and answers, safety consistency, and explicit intent identification. We argue that existing alignment methods suffer from poor adversarial generalization, a significant ``safety tax'' on reasoning, and unfaithful coarse-grained rewards, primarily due to their neglect of the LRM's intrinsic reasoning capabilities for safety. To address these, we propose \textbf{SAVeR$^2$}, a novel \textbf{S}afety \textbf{A}lignment method via \textbf{R}ewards with \textbf{R}easoning capability. SAVeR$^2$ decomposes the safety objective into four fine-grained verifiable reward signals and utilizes a difficulty-aware data selection strategy to stabilize training. Crucially, we introduce a reasoning-preserving gradient projection mechanism that analytically resolves conflicts between safety and reasoning gradients, effectively eliminating the safety tax. Extensive experiments on models ranging from 1.5B to 671B demonstrate that SAVeR$^2$ achieves superior safety performance and significantly lower over-refusal rates across multiple benchmarks without compromising general reasoning capabilities.
In embedding visualizations, nearby instances are interpreted as semantically related, although neighborhoods derived from input embeddings may not reflect content relationships. Neighborhood-based dimensionality reduction (DR) methods can propagate this limitation by optimizing layouts from embedding-based neighborhood graphs. We propose Knowledge-Guided Neighborhood Graph Construction (KGN), which represents relational knowledge derived from each instance as a set of knowledge graph (KG) triples. It uses the degree of triple sharing as pairwise semantic evidence to adjust embedding distances and reconstruct the top-k neighborhood graph; neither KG extraction nor neighborhood construction uses class labels. We also introduce IDProfile (IDP), which interprets residual mismatches by independently assessing whether the source content supports the assigned class label and the neighborhood-majority class. Across eight text and image dataset-encoder conditions, we characterize KGN over the full guidance-strength sweep and four KG similarity measures. Across this sweep, the maximum NMI and LT_2D each exceed the corresponding kNN baseline in all eight conditions, with the largest gains reaching 0.68 and 1.35, respectively. The gains and the extent of neighborhood reconfiguration vary across conditions and are larger under encoder conditions with lower kNN baseline scores. The IDP analysis shows that the largest reductions are concentrated in Placement Issue candidates, whereas Label Issue and Unresolved Case instances account for a larger share of the remaining review set. These results indicate that KGN can narrow the evaluated gap between embedding-based neighborhoods and content-supported semantic relations, while IDP shows that a residual mismatch should not automatically be interpreted as a projection failure.
Reinforcement learning has achieved strong performance in many tasks by letting agents learn from nothing more than a reward signal. However, in more complex tasks, agents still struggle to learn good, generalizable policies, even with carefully designed rewards and abundant samples. We argue that this difficulty stems from the reward decoding burden: designers encode their knowledge of the desired behavior into rewards and penalties, and agents must infer this knowledge from long sequences of noisy scalars. To address this burden, we propose prescriptive policy learning (PPL), which directly teaches agents higher-level knowledge about the tasks, termed principles of behavior, such as how to maintain balance, how to accumulate energy, and how to land smoothly. Through appropriately selected key quantities and references, PPL can directly indicate at every step whether the agent follows the principles and to what extent. Our experiments show that agents can learn good, generalizable policies without any reward. Taught only simple heuristic principles, PPL agents outperform reward-based methods with far fewer samples. Their policies also transfer directly to harder settings and keep performing well, demonstrating their strong generalization ability. In summary, PPL provides a new approach in which rough knowledge of the desired behavior is enough for agents to learn good, generalizable policies, even when the system dynamics cannot be accurately modeled.
Deep search and long-horizon reasoning unfold as branching trajectories that pursue parallel sub-queries, abandon dead-end detours, and cross-check sibling findings. Yet modern serving engines, harnesses, and RL trainers execute episodes as flat, append-only sequences: stretching a context window leaves failed detours that dilute attention, in-place edits break KV caching and recurrent states, and all-layer cache splicing overwrites hybrid recurrent working memory and triggers hijacking. We introduce Octopus, a co-designed serving runtime, memory manager, and RL framework executing episodes as a GPU-resident DAG over radix cursors (89.6% KV hit rate). Agents coordinate via five primitives: fork shares parent KV prefixes in $O(1)$ with isolated 32K budgets, rewind restores recurrent checkpoints, read_thread mounts sibling KV pages in zero copy, and wait and done synchronize results. Octopus splices sibling KV pages only into full-attention layers under monotonic position offsets while recurrent layers skip mounted spans, retaining 92.2% of full-text answer information and cutting hijacking $21\times$ ($57.9\% \to 2.7\%$) with zero new KV pages. During RL, our packed DAG kernel and DAG-GRPO objective deduplicate shared prefixes via loss masks while keeping mounted KV edges differentiable, cutting rollout-training KL divergence $32\times$ and routing reader gradients into producer KV projections. Trained on Qwen3.5-4B with binary rewards, Octopus self-organizes a commander-worker hierarchy and latent KV mesh, lifting BrowseComp-Plus accuracy under BM25 from 26.0% (32K ReAct) and 32.1% (143K GRPO) to 49.6% (+17.5 pts; 51.2% on OpenResearcher), surpassing 9B context-editing RL (42.5%), matching o3 (49.3%), and outperforming dense-retrieval agents up to $250\times$ larger: Tongyi-DeepResearch (44.5%), DeepSeek-R1 (16.4%), Kimi-K2 (35.4%), GPT-4.1 (36.4%), and Claude 4 Opus (36.8%).