This position paper argues that the evaluation of modern visual processing systems should no longer be driven primarily by single-metric image quality assessment benchmarks, particularly in the era of generative and perception-oriented methods. Image restoration exemplifies this divergence: while objective IQA metrics enable reproducible, scalable evaluation, they have increasingly drifted apart from human perception and user preferences. We contend that this mismatch risks constraining innovation and misguiding research progress across visual processing tasks. Rather than rejecting metrics altogether, this paper calls for a rebalancing of evaluation paradigms, advocating a more human-centered, context-aware, and fine-grained approach to assessing the visual models' outcomes.
Sparse oblique (SPO), part of the top-ranked configuration of Google's Yggdrasil Decision Forests (YDF), improve the accuracy while maintaining interpretability of Random Forests (RF) and Gradient Boosted Trees (GBT) by scanning a sparse linear combination of features instead of a single feature. Because projections are sampled at runtime, training is significantly slower, as pre-run optimizations such as presorting cannot be used. We overhaul SPO training in YDF, first by fixing inefficiencies in the released version, achieving 2-5x speedup, then by devising two novel methods, one aimed at GBTs and the other at RFs: (1) Hierarchical AVX2 and AVX-512-vectorized histogram filling speeds up typical GBT-depth trees by up to 1.6x in SPO-GBT and 2x for SPO-RF, and (2) runtime-dynamic histograms between vectorized histograms and exact splits, speeds up deeper, RF depth-typical trees (depths 16 to purity) by up to 1.6x with no detectable effect on accuracy. Our optimizations bring SPO-RF training time on par with YDF's axis-aligned RF. We extensively evaluate on 8 natural and 11 synthetic datasets of up to 10.5 million rows and 1.6 million features, from depth 6 to purity, and open-source the implementation.
Image-text retrieval has become a fundamental component in intelligent multimedia systems; however, most existing vision-language models are optimized for highresource languages and remain suboptimal for low-resource settings such as Vietnamese. This work introduces ViCLIP-OT, a foundation vision-language model specifically designed for Vietnamese image-text retrieval. The proposed framework integrates CLIP-style contrastive learning with a Similarity-Graph Regularized Optimal Transport (SIGROT) loss to enhance global cross-modal consistency and mitigate modality gap issues. Extensive experiments on three Vietnamese benchmarks (UITOpenViIC, KTVIC, and Crossmodal-3600) demonstrate that ViCLIP-OT consistently outperforms CLIP and SigLIP baselines in both in-domain and zero-shot settings. On UIT-OpenViIC, the model achieves an average Recall@K of 67.34%, improving upon CLIP by 5.75 percentage points. In zero-shot evaluation on Crossmodal-3600, ViCLIPOT surpasses CLIP by 11.72 percentage points. Embedding-space analysis further confirms improved alignment and reduced modality gap. The results indicate that integrating SIGROT provides an effective and scalable strategy for cross-modal retrieval in low-resource languages, offering practical implications for intelligent multimedia retrieval systems in Vietnamese and other underrepresented linguistic contexts.
Training selects for behavior, not circuitry: many weight configurations can implement the same function. Studying any single trained neural network thus risks describing accidents of one training run rather than the computation itself. This work shifts focus from what transformers happen to do to what they must do by extracting algorithmic cores, compact subspaces that are necessary and sufficient for a task and that recur across independently trained models. Here, Algorithmic Core Extraction (ACE) is introduced to isolate these subspaces, causally validate them, and recover the algorithms they implement, from synthetic tasks to large pretrained models. Markov-chain transformers embed 3D cores in nearly orthogonal subspaces yet recover identical transition spectra. Modular-addition transformers form compact cyclic cores at grokking that later inflate under continued regularization, redundantly distributing the same computation across many functionally equivalent modes. This functional redundancy accelerates the transition from memorization to generalization, yielding an inverse scaling law for grokking time. In six language models spanning two orders of magnitude in scale (GPT-2 Small/Medium/Large, LLaMA-3.1, Gemma-2, and Qwen2.5), subject-verb agreement is governed by a single, steerable axis that aligns across architectures. Flipping this axis inverts grammatical number throughout open-ended generation. For indirect object identification, a multi-step task, the same six models route computation through compact multi-dimensional cores at two sites, linked by a linear operator that transfers across model families. Together these results suggest that beneath the apparent complexity of trained transformers lies a simpler, shared computational structure, and that targeting invariants rather than parameterizations may offer a more tractable path to mechanistic understanding and control.
Synthetic tabular data generation has attracted growing attention due to its importance for data augmentation, foundation models, and privacy. However, real-world tabular datasets increasingly contain free-form text fields (e.g., reviews or clinical notes) alongside structured numerical and categorical attributes. Generating such heterogeneous tables with joint modeling of different modalities remains challenging. Existing approaches broadly fall into two categories: diffusion-based methods and LLM-based methods. Diffusion models can capture complex dependencies over numerical and categorical features in continuous or discrete spaces, but extending them to open-ended text is nontrivial, since they model each column as a continuous value or one of a closed set of categories, whereas text lives in an exponentially large sequence space. In contrast, LLM-based generators naturally produce fluent text, yet their discrete tokenization can distort precise or wide-range numerical values, hindering accurate modeling of both numbers and language. In this work, we propose TabDLM, a unified framework for free-form tabular data generation via a joint numerical-language diffusion model built on masked diffusion language models (MDLMs). TabDLM models textual and categorical features through masked diffusion, while modeling numerical features with a continuous diffusion process through learned specialized numeric tokens embedding; bidirectional attention then captures cross-modality interactions within a single model. Extensive experiments on diverse benchmarks demonstrate the effectiveness of TabDLM compared to strong diffusion- and LLM-based baselines. Our code is available at https://github.com/ilikevegetable/TabDLM.
Token-level steering has emerged as a pivotal approach for inference-time alignment, enabling fine-grained control over large language models (LLMs) by modulating their output distributions without parameter updates. While effective, existing methods rely on dense intervention at every decoding step. This persistent manipulation not only incurs substantial computational overhead but also risks compromising generation quality by excessively drifting from the model's intrinsic distribution. In this work, we show that dense intervention is unnecessary and propose Sparse Inference-time Alignment (SIA), which performs sparse junction steering, intervening only at critical decision points along the generation trajectory. Our key insight is that high-entropy junctions are empirically consistent with pivotal decision points in the generation trajectory and are particularly susceptible to misalignment, suggesting that these points benefit from alignment-related reward signal. Extensive experiments across different model families and alignment objectives show that steering only 20%-80% of tokens achieves the superior alignment-efficiency trade-offs. For strong base models such as Qwen3, intervening on as few as 20% of tokens matches or even surpasses heavily post-trained instruct models. This sparsity enables stronger guidance while better preserving the model's native distribution, integrates seamlessly with search-based methods (e.g., Best-of-N), and improves efficiency by reducing the required sampling budget while yielding practical wall-clock speedups under matched-quality comparisons.
We uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was injected. While the model denies injection in sampled outputs, logit lens analysis reveals clear detection signals in the residual stream, which are attenuated in the final layers. Furthermore, prompting the model with accurate information about AI introspection mechanisms can dramatically strengthen this effect: the sensitivity to injection increases massively (0.3% -> 39.9%) with only a 0.6% increase in false positives. Also, mutual information between nine injected and recovered concepts rises from 0.61 bits to 1.05 bits, ruling out generic noise explanations. Our results demonstrate models can have a surprising capacity for introspection and steering awareness that is easy to overlook, with consequences for latent reasoning and safety.
Multi-agent decision pipelines can outperform single agent workflows when complementarity holds, i.e., different agents bring unique information to inform a final decision. We propose ComplLLM, a post-training framework based on decision theory that fine-tunes a decision-assistant LLM to output signals that complement existing agent decisions, using complementary information as reward. We validate ComplLLM on synthetic and real-world tasks involving domain experts, demonstrating how the approach recovers known complementary information, produces plausible explanations of complementary signals to support downstream decision-makers, and improves human and LLM agent decision performance in controlled studies.
We study conditional average treatment effect (CATE) estimation from observational data in regimes with potentially weak overlap, when the nuisance response functions lie in a reproducing kernel Hilbert space (RKHS), but the contrast function is structurally simpler. We consider three forms of such simplicity: (i) membership in a lower-complexity RKHS with faster eigenvalue decay, (ii) a source condition relative to the nuisance kernel, or (iii) dependence on a known low-dimensional covariate representation. We develop a unified two-stage kernel ridge regression procedure that remains valid under weak overlap and attains minimax rates determined by the complexity of the contrast function. The procedure avoids propensity-score estimation and separate nuisance convergence-rate assumptions. We also show that a simple model-selection step over candidate contrast spaces and regularization levels satisfies an oracle inequality, enabling adaptation to unknown CATE regularity.
Attention scores in transformers are bilinear forms $S_{ij} = x_i^\top M x_j / \sqrt{d_h}$ whose maximum magnitude governs overflow risk in low-precision training. We derive a rank-aware concentration inequality: when the interaction matrix $M = W^Q W^{K\top}$ has rank $r \ll d$, tail probabilities for $\max_{i,j}|S_{ij}|$ decay as $\exp(-d^{2}\alpha^{2}/(\gamma r))$ rather than $\exp(-d\alpha^{2})$, where $\gamma > 1$. For transformer attention where $r = d_h$, this yields $8$ to $28\times$ tighter concentration than rank-agnostic bounds in modern architectures. We apply this result to FP8 training, deriving geometry-aware scale factors that provide principled overflow guarantees without observing activations. The method computes per-layer scales from the spectral norm $\|W^Q W^{K\top}\|_2$ via implicit power iteration, includes a grouped query attention formulation that avoids key expansion, and remains compatible with fused attention kernels. Across GPT-2 XL to Llama-2-70B, geometry-aware scaling eliminates overflows in transient scenarios where delayed scaling fails, while achieving comparable downstream MMLU accuracy.
Object detectors achieve strong performance under nominal imaging conditions but can fail silently when exposed to blur, noise, compression, adverse weather, or resolution changes. In safety-critical settings, it is therefore insufficient to produce predictions without assessing whether the input remains within the detector's nominal operating regime, which we refer to as self-aware object detection. We introduce a degradation-aware framework based on degradation manifolds, which structure a detector's feature space according to image degradation rather than semantic content. Our method augments a standard detection backbone with a lightweight embedding head trained via multi-layer contrastive learning, where images sharing the same degradation composition are pulled together while differing configurations are separated. Unlike prior contrastive IQA approaches, our method performs detector-aligned input-fidelity monitoring: it uses task-backbone features and a nominal prototype to detect degradation-induced shifts, rather than estimating perceptual image quality in a standalone representation. To anchor the representation, we estimate a pristine prototype from clean training embeddings, defining a nominal operating point in feature space. Self-awareness is then derived as geometric deviation from this reference, providing an intrinsic image-level signal of degradation-induced shift that is independent of detection confidence. Experiments on synthetic corruptions, cross-dataset zero-shot transfer, and natural weather-induced shifts show strong pristine-degraded separability, consistent behavior across detector architectures, and robust generalization across datasets with substantial content variation. These results suggest that degradation-aware detector representations provide a reliable signal for monitoring input-fidelity shifts.
The hyper-connection paradigm separates the skip-connections of transformers into $n$ parallel streams on which a linear mixer acts. Constraining the mixer to bistochastic matrices has proven effective in preventing gradient explosion but risk collapsing onto low-rank configurations, limiting expressivity. We therefore propose JPmHC (Jacobian-spectrum Preserving manifold-constrained Hyper-Connections). Our framework triages manifold constraints by computing the Jacobian singular spectrum of the end-to-end network in the mean-field limit and their spectral stability. It consists of: (i) an operator-valued free-probability library that predicts the Jacobian singular spectrum; (ii) an orthogonal mixer computed by a fixed-point Cayley iteration, which enables dynamical isometry; (iii) implicit differentiation of the Sinkhorn projection of the bistochastic baseline, with $O(1)$ backward operations; and (iv) a sensitivity theory giving the training budget before a bistochastic mixer collapses. The library is written in Lean, which certifies the mathematics and the algorithms. We validate JPmHC with TRM on ARC-AGI and nano-GPT on FineWeb-Edu: the theory predicts spectral collapse for TRM and stability for nano-GPT as observed in our experiments.
Long-context DNA modeling is computationally expensive, motivating existing approaches to reduce sequence length through uniform downsampling. However, functional information and sequence redundancy are unevenly distributed across the genome, calling for finer resolution in functionally dense regions and stronger compression of redundant sequence. We introduce GeneZip, a region-aware DNA compression framework that learns where to allocate a limited token budget. Building on H-Net-style dynamic routing, GeneZip combines a Region-Aware Ratio objective, which uses regional labels derived from GENCODE gene annotations to specify compression targets during training, with bounded routing for stable token counts. At inference, it compresses raw DNA without annotations. GeneZip provides three main benefits. First, it preserves predictive utility under high compression: GeneZip-70M achieves 137.6-fold compression with a validation perplexity increase of 0.31 relative to uncompressed Mamba2-84M, and the best average rank across four DNALongBench tasks among compression models with leakage-controlled pretraining. Second, it learns region-aware allocation with emergent repeat-aware compression. Alternative regional priors yield task-dependent performance trade-offs. Without repeat-specific supervision, GeneZip compresses transposable-element-derived and tandem repeats more strongly than matched non-repeat sequence; its tandem-repeat compression preference transfers to a nine-species evaluation panel without further training. Third, it enables efficient long-context modeling: on 450-kb inputs, GeneZip-70M achieves 36.1-fold higher inference throughput than JanusDNA-7.7M on a single H100 GPU.
Single-cell RNA sequencing (scRNA-seq) data exhibit strong and reproducible statistical structure. This has motivated the development of large-scale foundation models, such as TranscriptFormer, that use transformer-based architectures to learn a generative model for gene expression by embedding genes into a latent vector space. These embeddings have been used to obtain state-of-the-art (SOTA) performance on downstream tasks such as cell-type classification, disease-state prediction, and cross-species learning. Here, we ask whether similar performance can be achieved without utilizing computationally intensive deep learning-based representations. Using simple, interpretable pipelines that rely on careful normalization and linear methods, we obtain SOTA or near SOTA performance across multiple benchmarks commonly used to evaluate single-cell foundation models, including outperforming foundation models on out-of-distribution tasks involving novel cell types and organisms absent from the training data. Our findings highlight the need for rigorous benchmarking and suggest that the biology of cell identity can be captured by simple linear representations of single cell gene expression data.
Time-series anomaly detection (TSAD) requires identifying both immediate Point Anomalies and long-range Context Anomalies. However, existing foundation models face a fundamental trade-off: 1D temporal models provide fine-grained pointwise localization but lack a global contextual perspective, while 2D vision-based models capture global patterns but suffer from information bottlenecks due to a lack of temporal alignment and coarse-grained pointwise detection. To resolve this dilemma, we propose VETime, the first TSAD framework that unifies temporal and visual modalities through fine-grained visual-temporal alignment and dynamic fusion. VETime introduces a Reversible Image Conversion and a Patch-Level Temporal Alignment module to establish a shared visual-temporal timeline, preserving discriminative details while maintaining temporal sensitivity. Furthermore, we design an Anomaly Window Contrastive Learning mechanism and a Task-Adaptive Multi-Modal Fusion to adaptively integrate the complementary perceptual strengths of both modalities. Extensive experiments demonstrate that VETime significantly outperforms state-of-the-art models in zero-shot scenarios, achieving superior localization precision with lower computational overhead than current vision-based approaches. Code available at: https://github.com/yyyangcoder/VETime .
Identifying molecular structures from LC-MS/MS spectra is a central problem in computational metabolomics, often approached by predicting molecular fingerprints and retrieving candidates via similarity search. Despite widespread use, the relationship between training objectives for fingerprint prediction and downstream retrieval performance remains poorly understood. In this work, we investigate the alignment between these objectives and show that they can be misaligned under uncertainty. Adopting a decision-theoretic perspective, we derive novel regret bounds that characterize when Bayes-optimal predictors for fingerprint similarity diverge from those for molecular retrieval. Our analysis reveals that optimizing standard similarity-based losses can degrade retrieval performance, and that the extent of this mismatch depends on the similarity structure of candidate sets. Empirically, we validate our theory on the MassSpecGym and Enveda-180 datasets, observing a Pareto frontier between fingerprint accuracy and retrieval metrics across commonly used loss functions. These findings highlight potential trade-offs in fingerprint-based molecular identification and provide principled guidance for the design of learning objectives in small molecule mass spectrometry.
Large Reasoning Models (LRMs) have shown excellent performance in reasoning tasks using long Chains of Thought (CoTs). However, this has also led to a significant increase of computational costs and the generation of verbose output, a phenomenon known as overthinking. The tendency to overthink is often exacerbated by Reinforcement Learning (RL) algorithms such as GRPO/DAPO. In this paper, we propose BFS-PO, an RL algorithm which alleviates this problem using a Best-First Search exploration strategy. Specifically, the reasoning chains generated at training time by BFS-PO are organized into a search tree, within which the shortest correct sequence is selected as the best and expanded using a backtracking mechanism based on maximum entropy nodes. In this way, we bias the exploration of the solution space towards the search for increasingly shorter solutions, training the LRM to generate progressively more concise answers. Using different benchmarks and base LRMs, we show that BFS-PO can simultaneously increase the LRM accuracy and shorten its reasoning chains.
Building agentic systems that can autonomously self-improve from experience is a longstanding goal of AI. Large language models (LLMs) today primarily self-improve via two mechanisms: self-reflection for context updates, and reinforcement learning (RL) for weight updates. In this work, we propose Evolutionary System Prompt Learning (E-SPL), a method for jointly improving a model's experiential context and its weights. In each RL iteration, E-SPL samples trajectories under multiple system prompts in parallel, then uses them to jointly update the weights with RL and evolve the system prompts with self-reflection. E-SPL encourages a natural division between procedural knowledge in weights and interpretable declarative knowledge in system prompts. Most notably, in an easy-to-hard generalization setting (AIME $\rightarrow$ BeyondAIME), E-SPL improves RL success rate from 38.8% $\rightarrow$ 45.1%. E-SPL also improves RL on AIME 2025 (56.3% $\rightarrow$ 60.6%), HMMT 2025 (50.0% $\rightarrow$ 52.7%), and agentic search (44.2% $\rightarrow$ 48.6% on gpt-oss-120b). Across all settings, E-SPL outperforms both RL-only and evolution-only baselines, demonstrating that weight updates and context updates are deeply synergistic and can together yield gains in generalization that neither achieves alone. Code is available at https://github.com/LunjunZhang/E-SPL.
Recent years have witnessed meteoric progress in reasoning models: neural networks that generate intermediate reasoning traces (RTs) before producing a final output. Despite the rapid advancement, our understanding of how RTs support reasoning, and where this paradigm fails, remain incomplete. To promote greater clarity, we introduce PITA: a novel large-scale dataset of over 23 million statements in propositional logic and their corresponding proofs. As a benchmark for robust reasoning, we focus on length generalization: if a model is trained to determine truth or falsity on statements with proofs up to fixed length, how well does it generalize to statements requiring longer proofs? We propose notions of (1) task depth and (2) task breadth, which measure respectively (1) the number of proof steps required to solve a proposition and (2) the number of unique propositions within a task family. We vary these quantities across subsets of PITA, and find that RT models generalize well on relatively broad and shallow subsets, while deteriorating on relatively narrow and deep subsets compared to non-RT baselines. As a controlled point of comparison, we study a separate, tractable transitive inference task that exhibits qualitatively similar behavior. Our accompanying theory explains the scalings observed in this simpler setting, suggesting one mechanism by which breadth can favor RT models while depth can expose long-context weaknesses. Our findings suggest salient benefits and limitations of proof-like reasoning traces in controlled length-generalization settings.
We study instruction following in multi-task reinforcement learning, where an agent must zero-shot execute novel tasks not seen during training. In this setting, linear temporal logic (LTL) has recently been adopted as a powerful framework for specifying structured, temporally extended tasks. While existing approaches successfully train generalist policies, they often struggle to effectively capture the rich logical and temporal structure inherent in LTL specifications. In this work, we address these concerns with a novel approach to learn structured task representations that facilitate training and generalisation. Our method conditions the policy on sequences of Boolean formulae constructed from a finite automaton of the task. We propose a hierarchical neural architecture to encode the logical structure of these formulae, and introduce an attention mechanism that enables the policy to reason about future subgoals. Experiments in a variety of complex environments demonstrate the strong generalisation capabilities and superior performance of our approach.