arXiv:2511.23064v2 Announce Type: replace Abstract: Variational phase-field models of brittle fracture pose a local constrained minimization problem of a non-convex energy functional. In the discrete setting, the problem is most often solved by alternate minimization, exploiting the separate convexity of the energy with respect to the two unknowns. This approach is theoretically guaranteed to converge, provided each of the individual subproblems is solved successfully. However, strong non-linearities of the energy functional may lead to failure of iterative convergence within one or both subproblems. We analyze and visualize the energy along Newton directions to illustrate why Newton's method without line search fails. Motivated by this, we propose to employ an exact line search algorithm based on bisection, which (under certain conditions) can guarantee global convergence of Newton's method for each subproblem and consequently the successful determination of critical points of the energy through the alternate minimization scheme. Through several benchmark tests computed with various strain energy decompositions and two strategies for the enforcement of the irreversibility constraint in two and three dimensions, we demonstrate the robustness of the approach and assess its efficiency in comparison with other commonly used line search algorithms. With the outlined approach, we are able to compute the especially demanding Brazilian test featuring contact in 3D with the star-convex model.
Science Journals
arXiv:2512.01467v2 Announce Type: replace Abstract: Controlling autonomous systems under real-world conditions often requires policies that can be evaluated with low latency and minimal energy consumption. Unfortunately, these conditions are at odds with the use of high-precision deep neural networks as controllers. In this work, we introduce Differentiable Weightless Controllers (DWCs), a symbolic-differentiable architecture that learns flexible, non-linear, yet highly efficient control policies. DWCs can be trained end-to-end via gradient-based techniques, yet compile directly into FPGA-compatible circuits with few- or even single-clock-cycle latency and nanojoule-level energy cost per action. Across five MuJoCo benchmarks, including high-dimensional Humanoid, DWCs achieve returns competitive with standard deep policies (full-precision or quantized neural networks). Furthermore, DWCs exhibit structurally sparse and interpretable connectivity patterns, enabling direct inspection of which input values influence control decisions.
arXiv:2512.04860v3 Announce Type: replace Abstract: The stochastic density functional theory (sDFT) has exhibited advantages over the standard Kohn-Sham DFT method and has become an attractive approach for large-scale electronic structure calculations. The sDFT method avoids the expensive matrix diagonalization by introducing a set of random orbitals and approximating the density matrix via Chebyshev expansion of a matrix-valued function. In this work, we study the sDFT with a plane-wave discretization, and discuss variance reduction algorithms in the framework of multilevel Monte Carlo (MLMC) methods. In particular, we show that the density matrix evaluation in sDFT can be decomposed into many levels by increasing the plane-wave cutoffs or the Chebyshev polynomial orders. This decomposition renders the computational cost independent of the discretization size or temperature. To demonstrate the efficiency of the algorithm, we provide rigorous analysis of the statistical errors and present numerical experiments on some material systems.
arXiv:2512.10656v3 Announce Type: replace Abstract: As context windows in large language models continue to expand, it is essential to characterize how attention behaves at extreme sequence lengths. We introduce token sample complexity: the rate at which attention computed on $n$ tokens converges to its infinite-token limit. We estimate finite-$n$ convergence bounds at two levels: pointwise uniform convergence of the attention map, and convergence of moments for the transformed token distribution. For compactly supported (and more generally sub-Gaussian) distributions, our first result shows that the attention map converges uniformly on a ball of radius $R$ at rate $C(R)/\sqrt{n}$, where $C(R)$ grows exponentially with $R$. For large $R$, this estimate loses practical value, and our second result addresses this issue by establishing convergence rates for the moments of the transformed distribution (the token output of the attention layer). In this case, the rate is $C'(R)/n^{\beta}$ with $\beta<\tfrac{1}{2}$, and $C'(R)$ depends polynomially on the size of the support of the distribution. The exponent $\beta$ depends on the attention geometry and the spectral properties of the token distribution. We also examine the regime in which the attention parameter tends to infinity and the softmax approaches a hardmax, and in this setting, we establish a logarithmic rate of convergence. Experiments on synthetic and real data support our predictions and show that the predicted slowdown is reflected in downstream accuracy.
arXiv:2512.16349v2 Announce Type: replace Abstract: We propose a collaborative edge-to-server inference framework for vision-language models (VLMs) that reduces communication cost while maintaining inference accuracy. In typical deployments, visual data captured at edge devices (clients) is transmitted to the server for VLM inference. However, transmitting full-resolution images incurs high communication cost. Conversely, aggressive downsizing or excessive compression to mitigate communication overhead can discard fine-grained details, leading to accuracy degradation. To overcome this limitation, we design a communication-efficient two-stage framework. In the first stage, the server performs inference on the downsized thumbnail (global image) and quantifies the min-entropy of the output tokens. If the min-entropy exceeds a predefined threshold, the server identifies a region of interest (RoI) using the VLM's internal attention and requests the edge device to send a detail-preserved local image of the RoI. The server then refines its inference by jointly leveraging the global and local images. This selective retransmission strategy ensures that only essential visual content is additionally transmitted. Experimental results consistently confirm that the proposed framework substantially reduces communication overhead while maintaining inference accuracy across diverse VQA benchmarks.
arXiv:2601.04498v2 Announce Type: replace Abstract: Infographics are composite visual artifacts that combine data visualizations with textual and illustrative elements to communicate information. While recent text-to-image (T2I) models can generate aesthetically appealing images, their reliability in generating infographics remains unclear. Generated infographics may appear correct at first glance but contain easily overlooked issues, such as distorted data encoding or incorrect textual content. We present IGENBENCH, the first benchmark for evaluating the reliability of text-to-infographic generation, comprising 600 curated test cases spanning 30 infographic types. We design an automated evaluation framework that decomposes reliability verification into atomic yes/no questions based on a taxonomy of 10 question types. We employ multimodal large language models (MLLMs) to verify each question, yielding question-level accuracy (Q-ACC) and infographic-level accuracy (I-ACC). We comprehensively evaluate 10 state-of-the-art T2I models on IGENBENCH. Our systematic analysis reveals key insights for future model development: (i) a three-tier performance hierarchy with the top model achieving Q-ACC of 0.90 but I-ACC of only 0.49; (ii) data-related dimensions emerging as universal bottlenecks (e.g., Data Completeness: 0.21); and (iii) the challenge of achieving end-to-end correctness across all models. We release IGENBENCH at https://igen-bench.vercel.app/.
arXiv:2601.05261v2 Announce Type: replace Abstract: Online consumer reviews are important decision-support resources in e-commerce, yet the increasing volume of reviews often creates information overload and makes it difficult for users to identify content that matches their individual preferences. Existing review-ranking approaches commonly rely on aggregate signals such as star ratings, helpfulness votes, or recency, which may not reflect user-specific interests. This paper proposes a personalized review ranking and summarization framework that integrates user preference modeling, hybrid sentiment estimation, aspect-level review matching, and Large Language Model (LLM)-based summarization. The framework first extracts aspect-level preferences and sentiment signals from historical reviews. It then incorporates user-selected product aspects and written review input to build a personalized user profile. Candidate reviews are ranked by comparing this profile with review-level aspect and sentiment representations. The top-ranked reviews are then summarized to provide concise, preference-aligned information. The proposed method was evaluated using an Amazon Mobile Electronics review dataset and a structured user study involving 70 participants across common consumer electronics categories. Results show that the proposed ranking method outperformed random ordering, star-rating-based ranking, helpfulness-vote ranking, recency-based ranking, and semantic-similarity-based ranking. User-study results further indicate improvements in satisfaction, perceived relevance, decision-making confidence, ease of finding information, and reading efficiency. The findings suggest that combining aspect-level personalization, sentiment-aware ranking, and LLM-based summarization can reduce review overload and support more efficient user-centered decision-making.
arXiv:2601.07994v5 Announce Type: replace Abstract: Large Language Models (LLMs) increasingly operate over long-form dialogues with frequent topic shifts. While recent LLMs support extended context windows, efficient management of dialogue history in practice is needed due to inference cost and latency constraints. We present DyCP, a lightweight context management method implemented outside the LLM that dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn, without offline memory construction. DyCP manages dialogue context while preserving the sequential nature of dialogue without predefined topic boundaries, enabling adaptive and efficient context selection. Across three long-form dialogue benchmarks-LoCoMo, MT-Bench+, and SCM4LLMs-and multiple LLM backends, DyCP achieves competitive answer quality in downstream generation, with more selective context usage and improved inference efficiency.
arXiv:2601.08020v3 Announce Type: replace Abstract: Self-imaging represents a core hallmark of paraxial wave evolution; yet, across its many realizations and generalizations over the past two centuries, the uniformity of recurrence planes along the propagation axis has been considered fundamental. Here we reformulate the general phenomenon of self-imaging within the natural framework of canonical phase-space geometry, revealing a hidden canonical coordinate in which all exact self-imaging is indeed uniform, but which need not correspond to the physical propagation axis. This leads to a general law of self-imaging, in which the spacing of the physical recurrence planes can be prescribed through the choice of initial transverse phase structure. Using a single programmable spatial light modulator, we demonstrate the construction of Talbot carpets characterized by recurrence spacings that accelerate and decelerate along the propagation axis, as well as those that follow polynomial, exponential, and sinusoidal axial trajectories. These results reveal a hidden geometric freedom in paraxial wave propagation: exact self-imaging is rigid in canonical coordinates, but freely programmable in physical space, allowing for qualitatively new forms of optical recurrence.
arXiv:2601.09285v2 Announce Type: replace Abstract: Metal-organic frameworks (MOFs) are porous crystalline materials with broad applications such as carbon capture and drug delivery, yet accurately predicting their 3D structures remains a significant challenge. While Large Language Models (LLMs) have shown promise in generating crystal structures, their application to MOFs is hindered by MOFs' high structural complexity arising from the large number of atoms in unit cell. Inspired by the success of block-wise paradigms in deep generative models for MOFs, we pioneer the application of LLMs in this domain by introducing MOF-LLM, the first LLM framework specifically adapted for block-level MOF structure prediction. To effectively harness LLMs for this 3D modular assembly task, our training paradigm integrates spatial-aware continual pre-training (CPT), structural supervised fine-tuning (SFT), and matching-driven reinforcement learning (RL). By incorporating explicit spatial priors and optimizing structural stability via Soft Adaptive Policy Optimization (SAPO), our approach substantially enhances the spatial reasoning in a Qwen-3 8B model for MOF structure prediction. Comprehensive experiments demonstrate that MOF-LLM achieves state-of-the-art performance with a match rate of 35.78% while exhibiting superior sampling efficiency of 0.04 seconds per structure.
arXiv:2601.09764v3 Announce Type: replace Abstract: A new interferometer topology for kilohertz gravitational-wave detection was recently proposed in [Zhang et al. Phys. Rev. X 13, 021019 (2023)]. The design is based on an L-shaped optical cavity pumped through a Sagnac-like vortex. We report a tabletop experiment that characterizes the interferometer's optical response near DC. When the laser frequency is locked to the resonance of the L-shaped cavity, we observe that the cavity input coupler becomes effectively transparent, yielding a simple Michelson-like response. Moreover, the Sagnac vortex separates into upper and lower paths, which behave as two independent pumping paths driving the cavity. These observations are in agreement with theoretical predictions. Our results provide an intuitive physical picture of this interferometer topology and offer insight into its lock acquisition strategy.
arXiv:2601.12735v2 Announce Type: replace Abstract: REST APIs, based on the REpresentational State Transfer (REST) architecture, are the primary type of Web API. The OpenAPI Specification (OAS) serves as the de facto standard for describing REST APIs and is crucial for multiple software engineering tasks. Automated OAS generation can help developers identify and correct issues in manually maintained OAS, but existing approaches rely on technology-specific rules and human expert intervention. LLMs' powerful code understanding capabilities offer the potential to overcome these limitations, but introduce additional challenges such as context length limitations and hallucinations. To address these challenges, we propose OOPS, the first technology-agnostic approach that leverages LLM-based static analysis of server code for OAS generation. Through an LLM agent workflow comprising two key steps, endpoint method extraction and OAS generation, OOPS eliminates the need for technology-specific rules or human expert intervention. By constructing an API dependency graph, it establishes necessary file associations to address LLMs' context length limitations. By multi-stage generation and self-refine, it mitigates both syntactic and semantic hallucinations during OAS generation. We evaluated OOPS on 12 real-world REST APIs spanning 5 programming languages and 8 development frameworks. Experimental results demonstrate that OOPS accurately generates high-quality OAS for REST APIs implemented with diverse technologies, achieving an average F1-score exceeding 98% for endpoint method inference, 97% for both request parameter and response inference, and 92% for parameter constraint inference. The input tokens average below 5.6K with a maximum of 16.13K, while the output tokens average below 0.9K with a maximum of 7.63K.
arXiv:2601.15165v4 Announce Type: replace Abstract: Diffusion Large Language Models (dLLMs) break the rigid left-to-right constraint of traditional LLMs, enabling token generation in arbitrary orders. Intuitively, this flexibility implies a solution space that strictly supersets the fixed autoregressive trajectory, theoretically unlocking superior reasoning potential. However, in this paper, we find that for general reasoning tasks (e.g., mathematics and coding), arbitrary order generation may in fact limit the reasoning potential of dLLMs. We observe that dLLMs tend to exploit this order flexibility to bypass high-uncertainty tokens that are crucial for exploration, which can lead to a premature collapse of solution coverage. This observation motivates a rethink of RL approaches for dLLMs, where considerable complexities, such as handling combinatorial trajectories and intractable likelihoods, are often devoted to preserving this flexibility. We show that effective reasoning can be elicited by simply forgoing arbitrary order and applying standard Group Relative Policy Optimization (GRPO) instead. Our approach, JustGRPO, is minimalist yet surprisingly effective (e.g., 89.1% accuracy on GSM8K) while fully retaining the parallel decoding ability of dLLMs. Project page: https://nzl-thu.github.io/the-flexibility-trap
arXiv:2601.18957v2 Announce Type: replace Abstract: We show how several important classical problems, with positive definite potential energy, can be solved by starting from the factorization of the total mechanical energy using complex numbers. In particular, we derive in a new way exact analytical solutions for: simple harmonic oscillator, vertical projectile motion, motion under a repulsive inverse cube force, and damped harmonic oscillator (with linear damping). We also show how this approach easily yields an excellent approximation of the energy decay and a new approximate analytical solution in the case of a weakly damped harmonic oscillator. Our derivations are suitable for undergraduate physics teaching as an alternative to solving Newton's equations of motion. In addition, we comment on the limitations of our approach, but also on the insights it provides and opportunities for further research.
arXiv:2601.20600v2 Announce Type: replace Abstract: In the recent years, there has been active research on self-orthogonal embeddings of linear codes since they yielded some optimal self-orthogonal codes. LCD codes have a trivial hull so they are counterparts of self-orthogonal codes. So it is a natural question whether one can embed linear codes into optimal LCD codes. To answer it, we first determine the number of columns to be added to a generator matrix of a linear code in order to embed the given code into an LCD code. Then we characterize all possible forms of shortest LCD embeddings of a linear code. As examples, we start from binary and ternary Hamming codes of small lengths and obtain optimal LCD codes with minimum distance 4. Furthermore, we find new ternary LCD codes with parameters including $[23, 4, 14]$, $[23, 5, 12]$, $[24, 6, 12]$, and $[25, 5, 14]$ and a new quaternary LCD $[21, 10, 8]$ code, each of which has minimum distance one greater than those of known codes. This shows that our shortest LCD embedding method is useful in finding optimal LCD codes over various fields.
arXiv:2601.21149v3 Announce Type: replace Abstract: Recent progress in geospatial foundation models highlights the importance of learning general-purpose representations for real-world locations, particularly points-of-interest (POIs) where human activity concentrates. Existing approaches, however, focus primarily on place identity derived from static textual metadata, or learn representations tied to trajectory context, which capture movement regularities rather than how places are actually used (i.e., POI's function). We argue that POI function is a missing but essential signal for general POI representations. We introduce Mobility-Embedded POIs (ME-POIs), a framework that augments POI embeddings derived, from language models with large-scale human mobility data to learn POI-centric, context-independent representations grounded in real-world usage. ME-POIs encodes individual visits as temporally contextualized embeddings and aligns them with learnable POI representations via contrastive learning to capture usage patterns across users and time. To address long-tail sparsity, we propose a novel mechanism that propagates temporal visit patterns from nearby, frequently visited POIs across multiple spatial scales. We evaluate ME-POIs on five newly proposed map enrichment tasks, testing its ability to capture both the identity and function of POIs. Across all tasks, augmenting text-based embeddings with ME-POIs consistently outperforms both text-only and mobility-only baselines. Notably, ME-POIs trained on mobility data alone can surpass text-only models on certain tasks, highlighting that POI function is a critical component of accurate and generalizable POI representations.
arXiv:2601.21754v3 Announce Type: replace Abstract: While Large Language Models (LLMs) excel in language-based agentic tasks, their applicability to unseen, nonlinguistic environments (e.g., symbolic or spatial tasks) remains limited. Previous work attributes this performance gap to the mismatch between the pretraining distribution and the testing distribution. In this work, we demonstrate the primary bottleneck is the prohibitive cost of exploration: mastering these tasks requires extensive trial-and-error, which is computationally unsustainable for parameter-heavy LLMs operating in a high dimensional semantic space. To address this, we propose SCOUT (Sub-Scale Collaboration On Unseen Tasks), a novel framework that decouples exploration from exploitation. We employ lightweight "scouts" (e.g., small MLPs) to probe environmental dynamics at a speed and scale far exceeding LLMs. The collected trajectories are utilized to bootstrap the LLM via Supervised Fine-Tuning (SFT), followed by multi-turn Reinforcement Learning (RL) to activate its latent world knowledge. Empirically, SCOUT enables a Qwen2.5-3B-Instruct model to achieve an average score of 0.86, significantly outperforming proprietary models, including Gemini-2.5-Pro (0.60), while saving about 60% GPU hours consumption.
arXiv:2601.21881v2 Announce Type: replace Abstract: Humans can infer mechanical outcomes by learning from a few observations. This capacity for mechanics intuition is acquired in a data-efficient manner. Here, we propose a reinforcement learning framework to mimic this process, in which an agent encodes continuous physical observation parameters into its state and is trained via episodic switching across closely related observations. With merely two or three similar observations, the agent acquires robust mechanics intuition that generalizes over wide parameter ranges beyond the training data. Our method is demonstrated on the brachistochrone, a large-deformation elastic plate, and the quantum harmonic oscillator. We explain this generalization through a unified theoretical view: it is associated with cross-parameter Bellman consistency encouraged by episodic switching across neighboring task parameters, promoting approximate stationarity of the Bellman residual with respect to physical variations. This is consistent with a smooth policy that tracks a low-dimensional solution manifold underlying the continuum of tasks. Our work identifies episodic switching as a practically effective and theoretically motivated route to artificial mechanics intuition and suggests a computational analogy to data-efficient generalization in biological learners.
arXiv:2601.21996v2 Announce Type: replace Abstract: While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.
arXiv:2602.00238v2 Announce Type: replace Abstract: Existing retrieval-augmented generation (RAG) systems often assume that each query has a single correct answer. This assumption overlooks open-ended information-seeking scenarios where multiple plausible answers are valuable, and where diversity is important for creativity, fairness, and inclusive access to information. We show that standard RAG systems fail to fully use diverse retrieved contexts: simply increasing retrieval diversity does not necessarily lead to diverse generations. To address this limitation, we propose Diverge, a plug-and-play agentic RAG framework that improves the diversity--quality trade-off through iterative, reflection-guided exploration of diverse viewpoints and diversity-aware retrieval support. We further introduce evaluation metrics for characterizing the diversity-quality trade-off in open-ended question answering. Experiments across multiple real-world datasets and backbone LLMs show that Diverge achieves the best trade-off among competitive baselines, increasing diversity by $\sim2\times$ without noticeable quality degradation. These results reveal a systematic limitation of current RAGs and show the value of explicit diversity modeling.
arXiv:2602.05845v2 Announce Type: replace Abstract: We propose a multi-task formulation of self-predictive Siamese SSL in which each spatial transformation defines a distinct latent-space alignment task, solved by a dedicated predictor over a shared encoder. This perspective directly explains a long-standing failure of multi-crop training in self-predictive methods such as BYOL, SimSiam, and MoCo v3: a shared predictor is forced to solve heterogeneous alignment tasks simultaneously, leading to unstable optimization. Assigning one predictor per view type resolves this interference, unlocking linear evaluation gains of 3.8-4\% across frameworks. This perspective also suggests a principled way to enrich pre-training by introducing additional spatial transformations as complementary tasks. We demonstrate this by introducing asymmetric cutout views, in which a masked online view is aligned with a complete target, forming a semantic inpainting objective. The resulting framework is stable, backbone-agnostic, and consistently improves the performance of ResNet and ViT models on ImageNet and COCO.
arXiv:2602.06307v2 Announce Type: replace Abstract: Spoken bilingual conversations pose substantial challenges for syntactic parsing because they often include disfluencies and discourse-driven structures that complicate dependency parsing under standard Universal Dependencies (UD) assumptions and evaluation practices. To systematically study these challenges, in this work, we first introduce a linguistically grounded taxonomy of conversational bilingual phenomena, together with SpokeBench, an expert-annotated English-Spanish benchmark for structurally complex speech. To address the limitations of existing evaluation practices, we propose Flex-UD, an ambiguity-aware evaluation metric that distinguishes catastrophic structural failures from linguistically acceptable variations. Finally, we introduce DECAP, a decoupled agentic parsing framework that separates spoken-phenomena handling from core syntactic analysis, enabling robust and interpretable dependency parsing without retraining. Experiments across both proprietary and open-weight LLMs show that DECAP substantially improves performance on complex conversational phenomena and achieves over 60% improvements in UPOS-F1 Score over baselines, while Flex-UD evaluations reveal gains that otherwise remain partially hidden under standard attachment-based metrics.
arXiv:2602.06677v3 Announce Type: replace Abstract: We analyze the Double Fourier Sphere (DFS) method on the rotation group $\mathcal{SO}(3)$ in the frequency domain and demonstrate its central role in fast algorithms. Fast Fourier algorithms on $\mathcal{SO}(3)$ are commonly formulated as a Wigner transform - mapping harmonic to Fourier coefficients - followed by a Fourier transform. We revisit this formulation and interpret the Wigner transform as an explicit realization of the DFS method, lifting functions from $\mathcal{SO}(3)$ to $\mathbb{T}^3$. In this context, we analyze the Sobolev regularity loss induced by this lifting. Furthermore, we compare different Wigner transform implementations, examine additional symmetry enhancements, and observe that the direct method is often faster and more stable than the fast polynomial transform approaches.
arXiv:2602.08733v2 Announce Type: replace Abstract: Ordinary differential equations (ODEs) are central to scientific modelling, but inferring their vector fields from noisy trajectories remains challenging. Current approaches such as symbolic regression, Gaussian process (GP) regression, and Neural ODEs often require complex training pipelines and substantial machine learning expertise, or they depend strongly on system-specific prior knowledge. We propose FIM-ODE, a pretrained Foundation Inference Model that amortises low-dimensional ODE inference by predicting the vector field directly from noisy trajectory data in a single forward pass. We pretrain FIM-ODE on a prior distribution over ODEs with low-degree polynomial vector fields and represent the target field with neural operators. FIM-ODE achieves strong zero-shot performance, matching and often improving upon ODEFormer, a recent pretrained symbolic baseline, across a range of regimes despite using a simpler pretraining prior distribution. Pretraining also provides a strong initialisation for finetuning, enabling fast and stable adaptation that outperforms modern neural and GP baselines without requiring machine learning expertise.
arXiv:2602.11934v2 Announce Type: replace Abstract: Robot manipulation often fails in the final millimeters: a policy may recognize the right object yet miss the pose offsets, boundaries, or pre-contact alignments needed for action. We argue that such failures arise when semantic invariance suppresses correspondence cues for closed-loop control, or when these cues are not exposed to the policy in a usable form. Modern visual encoders provide strong semantic abstractions, but contact-rich manipulation requires correspondence sensitivity: discriminative feature responses to action-relevant changes in pose, boundary, and contact geometry. Diffusion features provide a strong prior for dense correspondence, but direct use is impractical due to stochasticity, latency, and representation drift. We introduce Robot-DIFT, a deterministic diffusion-derived backbone for real-time control. Through Manifold Distillation, Robot-DIFT converts a noise-conditioned diffusion Teacher into a clean-input, single-pass Student while preserving the teacher's feature manifold. A Spatial--Semantic Feature Pyramid Network (S2-FPN) fuses coarse-to-fine Student decoder features into visual tokens that expose semantic context and fine contact detail to the policy. Across RoboCasa, LIBERO-10, and real robots, Robot-DIFT outperforms vision--language, self-supervised, geometry-oriented, and diffusion baselines on contact-sensitive tasks. Controlled backbone/readout swaps show that S2-FPN unlocks, rather than replaces, the diffusion correspondence prior.