AI papers — 2026-10-11

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Multimodal large language models can be used for image classification, which is important because the way we evaluate these models often hides real performance issues. We see conflicting results when comparing them to supervised and vision-language models, and this confusion usually comes down to flawed evaluation protocols that either make the model look better or worse than it truly is.

The work by Multimodal Large Language Models as Image Classifiers shows that MLLMs benefit most when labels are corrected, showing an improvement of up to ten point eight percent, which significantly closes the gap with supervised models. This suggests that noisy ground truth is a major factor in how we perceive model performance.

Another area of focus is scaling native multimodal pre-training from scratch, as this approach aims to overcome the limitations of text-only training by allowing models to integrate visual information deeply. This research demonstrates that there is a budget compensation rule, meaning an extra budget for multimodal tokens can offset the efficiency penalty incurred by incorporating visuals into the training process.

In terms of practical application, we have work on improving software porting using Certificate-Driven Evolutionary Search, which introduces enforceable restrictions based on failed attempts to guide code translation between CPU and GPU. This method achieves substantial speedups, with one function showing a throughput exceeding an expert implementation by nearly fifteen percent.

Furthermore, the development of SkillGraph offers a way for language model agents to reuse past experiences by structuring skills as nodes in a directed graph that encodes dependencies, which is crucial for complex tasks requiring the composition of multiple skills. This framework has shown state-of-the-art performance on tasks demanding multi-step decision making.

The work that matters most right now is LearnPruner because it tackles the computational strain of vision language models by intelligently removing unnecessary visual tokens while preserving performance, which is crucial for making these large models practical. This framework proposes a two-stage pruning process: first, using a learnable module to remove redundant vision tokens after the vision encoder, and second, retaining only task-relevant tokens in the LLM's middle layer AKMPC. This approach has shown promising results, preserving about ninety-five percent of original performance while cutting the token count by over ninety percent and speeding up inference three point two times.

ShapeY provides a principled way to measure how well object recognition systems grasp shape cues, which is important because current deep networks often rely too much on texture rather than true shape. This framework uses nearest-neighbor matching across 68,200 images to probe the embedding space for 3D object similarity, revealing that even state-of-the-art models struggle with robust generalization across different viewpoints ShapeY.

IAD-Unify addresses the complex requirements of industrial anomaly inspection by unifying a multimodal language model, a visual expert, and a diffusion editor through task-specific interfaces IAD-Unify. This architecture uses dense anomaly fields to generate evidence tokens that are then consumed by Qwen3.5 for grounded answers and converted into semantic residuals for the diffusion editor, achieving a high segmentation average of seventy-three point zero two percent on one reference image IAD-Unify.

Domain-Adapted Retrieval for In-Context Annotation of Pedagogical Dialogue Acts shows that adapting retrieval by fine-tuning an embedding model on tutoring corpora significantly boosts annotation accuracy, achieving Cohen's kappa scores between zero point five two six and zero point five eight zero on TalkMoves Domain-Adapted Retrieval for In-Context Annotation of Pedagogical Dialogue Acts. This suggests that grounding retrieval is a practical way to improve annotation quality without needing to fine-tune the generative model itself.

PhysMoDPO focuses on making text-conditioned human motion generation physically plausible by integrating a whole-body controller into the training pipeline PhysMoDPO. This framework optimizes diffusion models using physics-based rewards to ensure the resulting motions are compliant with both physical laws and the original text instructions, leading to consistent improvements in physical realism on simulated robots.

Agentic Critical Training introduces a method called Agentic Critical Training that uses reinforcement learning with verifiable rewards to train language model agents to judge actions directly Agentic Critical Training. This technique rewards models only for selecting the expert action during trajectory states, showing gains over imitation learning and even over supervised reflection baselines on several benchmarks.

One Permutation Is All You Need demonstrates a way to get reliable feature importance by using a single, max-min rank-optimal deterministic permutation instead of many random shuffles One Permutation Is All You Need. This method eliminates estimation variance and provides a principled framework for auditing model reliance on features in complex scenarios.

RT-DETRv4 proposes a cost-effective distillation framework that uses vision foundation models to enhance lightweight object detectors, introducing a deep semantic injector and gradient-guided adaptive modulation RT-DETRv4. This approach allows for stable semantic transfer into resource-constrained detectors without increasing inference overhead, achieving state-of-the-art results on COCO with high frame rates.

The most critical finding relates to whether the reasoning that drives planning in vision-language driving agents is actually causal, which is explored by DriveMind AKMPC. This matters because if the reasoning is just a byproduct, it changes how we trust these autonomous systems. The results showed a clear disconnect: removing ego or navigation priors caused significant drops in planning scores, but removing the chain-of-thought reasoning only caused minor changes, suggesting that planning relies heavily on those initial priors rather than the generated text.

This causal disconnect is further illuminated by attention analysis in AKMPC, which indicated that planning primarily focused on these priors instead of the reasoning steps. This contrasts with work like Inverse-LLaVA Inverse-LLaVA, which shows how multimodal alignment can be achieved by reversing the typical projection, using text to guide visual feature updates through decoder attention.

Another area of investigation involves improving action localization accuracy; specifically, LLM-powered Query Expansion LLM-powered Query Expansion suggests expanding language queries with boundary descriptions can reduce uncertainty in training data. This idea connects to the need for robust reasoning, which is addressed by Uncertainty-Aware Budget Allocation Uncertainty-Aware Budget Allocation, a framework that intelligently samples reasoning steps based on their initial disagreement to improve accuracy without uniform compute allocation.

Furthermore, disentangling world and ego dynamics in embodied video generation is key for long-horizon tasks, as the World-Ego Model World-Ego Modeling proposes separating scene evolution from robot behavior to maintain consistency during navigation. This decomposition relates conceptually to how models learn representations, similar to the fusion techniques explored in Inverse-LLaVA Inverse-LLaVA, but applied specifically to embodied video streams rather than static image-text pairs.

The work that matters most is PixVL because it tackles the fundamental problem of creating useful pixel-level multimodal large language models by introducing a unified Mask--Text Consistency Cycle to bridge the gap between region segmentation and region understanding. This framework allows these models to learn from unlabeled data by self-verifying their descriptions against geometric reconstruction, which is a significant step toward more robust visual reasoning.

This approach builds upon prior efforts where direct cycle-based methods were found to be unreliable because re-segmentation IoU does not accurately reflect semantic quality or referring sufficiency. To fix this, PixVL introduces confuser-aware semantic verification, which rewards correct choices when a model is uncertain between similar regions and penalizes incorrect ones with zero reward. Furthermore, it employs cross-view verification using temporally separated video frames or geometrically transformed image views to prevent the learning process from collapsing into simple positional shortcuts.

This self-verification strategy is then coupled with a quality-coupled bidirectional learning strategy that uses the highest-reward description to guide text-to-mask learning, effectively turning region understanding and segmentation into mutual generators and verifiers. This entire framework demonstrates improvements in both tasks, showing that PixVL enhances both region understanding and segmentation performance.

The concept of improving model performance through structured optimization is also relevant when considering how we sample from diffusion models for generation. Score distillation methods aim to turn pretrained 2D diffusion models into 3D generators, but they suffer from high variance because the per-step gradient relies on a single random view, which is blind to global shape consistency.

To mitigate this variance, Multi-View Aggregated Score Distillation introduces MV-SDI, a training-free sampler that replaces the single-view gradient with an average over K views at a fixed UNet call budget. By averaging K views, the per-step gradient variance is reduced toward one over K of its single-view value.

Drawing antithetic antipodal pairs from these K views adds no further variance reduction, but it does stratify angular coverage across the sampling steps, removing clustering that occurs when sampling only from the same hemisphere. When tested on the 43-prompt SDI benchmark with a budget of ten thousand UNet calls, setting K to two halves the optimization steps and raises CLIP R-Precision from seventy-four point eight percent to eighty-three point eight percent.

This method is drop-in for existing gradient-based score distillation pipelines, requiring no retraining or multi-view data. This contrasts with methods that attempt to retrain the diffusion prior on multi-view data, which can conflate the sampling contribution with the quality of the prior itself.

In a different context, when we look at how models handle visual information for real-time control, Learned Image Compression for Vision-Language-Action Models introduces SPARC. This framework is tailored for VLA models by employing a lightweight temporal mask selector that adaptively allocates bitrate based on task relevance while using temporal context.

SPARC's key insight is that the importance of visual information changes across both camera views and spatial regions within an image, which informs its adaptive rate allocation over latent representations. Experiments on benchmarks like RoboCasa365 show that SPARC achieves stronger control performance than conventional codecs under the same bitrate budget.

Finally, when we consider how errors propagate through a model's long-term reasoning capabilities, MemTrace provides a framework for tracing and attributing errors in LLM memory systems. It transforms memory pipelines into executable memory evolution graphs to allow fine-grained tracing of operational information flow.

This system constructs MemTraceBench from various systems like Long-Context and RAG to study failure modes, using an automatic attribution method that traces operation subgraphs to pinpoint root causes. This allows for a closed-loop system that corrects faults and boosts end-task performance by up to seven point six two percentage points.

The most important development here concerns Constrained Dynamic Gaussian Splatting because it directly addresses the practical deployment hurdle of high-fidelity 4D reconstruction that is currently bottlenecked by memory constraints on edge devices. This work proposes a novel framework called Constrained Dynamic Gaussian Splatting which turns scene reconstruction into a budget-constrained optimization problem. The core idea involves using a differentiable budget controller guided by a multi-modal unified importance score to precisely regulate the Gaussian capacity during training. This is further supported by an adaptive static-dynamic allocation strategy that splits the representation and distributes global capacity based on motion complexity, all integrated through a three-phase training strategy to strictly adhere to the target Gaussian count.

This approach achieves over three times model compression compared to Ex4DGS while maintaining rendering quality under varying capacity limits. This is significant because it moves beyond simple heuristic pruning, which has been shown in other contexts to fail in achieving optimal rendering quality under preset budgets. Building on this success, GUI-KV introduces a plug-and-play KV cache compression method for GUI agents that requires no retraining.

This method shows that exploiting GUI-specific redundancies can reduce decoding FLOPs by nearly forty percent while increasing step accuracy, demonstrating a clear path toward making agent inference more feasible. This contrasts with other efficiency efforts; for instance, GLACIER rethinks mass spectrum prediction by framing it as an object detection problem on molecular graphs, achieving significant speedups over prior two-stage models.

Similarly, CMGL improves cancer subtype classification by producing separate reliability estimates for each patient's modality before fusing them into a consensus graph classification. This method consistently surpasses strong baselines in accuracy on four single-cancer tasks and shows that representations can transfer between different cancer types without fine-tuning.

On the alignment side, Decoding-time Realignment DeRa offers a way to explore different regularization strengths for language models without the resource drain of retraining multiple models. This allows users to smoothly transition between unaligned and aligned states by controlling the degree of alignment during decoding.

Finally, V-ECE provides a more accurate way to estimate general expected calibration errors in probabilistic classification by using prediction-dependent proper scores instead of traditional binning methods. This new estimator shows superior performance across both binary and multiclass settings for estimating calibration error, offering a robust diagnostic tool for model reliability.

The most crucial work here is the effort to expose exploitable shortcuts in multimodal large language models, which matters because if we don't know what patterns a model learns from test sets, we can't truly trust its performance on unseen data. The Test-set Stress-Test cross-validates a text only Qwen2-7B on the test set's questions and answer options, showing that it achieves a benchmark level score while also having a per question bias score s(x); this means we can measure how much the model relies on specific patterns in the questions themselves. This is supported by the work AKMPC, which shows that on template based benchmarks like VSI-Bench and CV-Bench, a fine tuned model's vision blind gap widens more than random removal when using Iterative Bias Pruning (IBP).

The DDL dataset is also significant because it addresses the severe lack of diversity in deepfake detection data, providing over one million forged samples across eighty different methods and various manipulation modes. This large scale makes it a better testbed for building next generation localization and interpretation methods. Building on this data challenge, the work DDL provides precise spatial masks and temporal segments for these forged samples, which is vital for interpretability.

Towards reasonable concept bottleneck models offers a framework that lets practitioners explicitly encode prior knowledge about concept relationships in models called CREAMs, which supports efficient interventions even when only a limited number of concepts are available. This complements the idea of probing model behavior, as it provides an internal mechanism for grounding predictions.

Policy Learning with a Language Bottleneck introduces PLLB, which allows AI agents to generate linguistic rules that capture high level strategies for tasks like maze navigation and robot grasp planning, making their learned behaviors more interpretable and shareable with humans. This contrasts with the blind success seen in MMMU where GPT-4o correctly answers a large percentage of questions without the image.

Finally, the work on Quantum Multi-Armed Bandits and Linear Bandits establishes tight lower bounds for these quantum learning problems, proving that regret cannot be independent of T or d in certain scenarios. This sets a theoretical limit on how well we can optimize action selection using quantum reward oracles.

The work by AKMPC is most significant because it addresses a fundamental flaw in how we evaluate fine-tuned reasoning models: the loss of explicit reasoning traces during adaptation. This happens when models are fine-tuned on instruction response data that lacks those initial reasoning steps, leading to a collapse where the model produces plausible answers but loses its structural validity.

This issue is compounded by how we measure performance; standard metrics can mask this failure because they focus only on the final answer correctness rather than the underlying reasoning structure. The framework introduced by AKMPC separates answer correctness from reasoning-trace validity, showing that supervised fine-tuning can rapidly suppress valid reasoning traces while answer-only metrics hide this decline.

In a related area, the GeoFuse framework aims to improve drone geo-localization in poor weather by integrating road map data as a free geometric prior. This method uses precisely aligned road map tiles with satellite imagery and employs a flexible fusion module to combine these features, consistently outperforming existing methods on benchmarks like University-1652 and DenseUAV.

Moving toward model training robustness, the work Untrained CNNs Exceed Backpropagation in V1 Alignment at High Evaluation Resolution investigates how different learning rules behave when compared against human fMRI data. This study found that an untrained random-weights baseline often outperforms backpropagation at higher evaluation resolutions, suggesting that the choice of learning rule matters significantly depending on the resolution used for testing.

Furthermore, in the realm of medical AI, Similarity-as-Evidence proposes a framework to calibrate overconfident vision-language models by treating text-image similarities as evidence rather than deterministic scores. This Similarity Evidence Head mitigates overconfidence by explicitly quantifying lack of evidence and dissonance, leading to superior calibration on medical imaging datasets.

Finally, for complex decision-making under uncertainty, Context-Aware Error Mitigation Orchestration for Hybrid Quantum Reinforcement Learning on NISQ Systems introduces Adaptive Policy-Guided Error Mitigation to dynamically select the best error mitigation strategy during quantum reinforcement learning training based on policy indicators. This approach has shown that integrating this adaptive layer substantially improves the robustness of QRL on noisy hardware.

The most significant development is how predictive representations can reveal underlying linguistic organization beyond simple word classes. This is because when a residual network predicts future states at different time horizons, it doesn't just see words; it sees structures like nouns, verbs, and adjectives emerging from the geometry of those predictions Word-Class and Construction-Like Structure Emerges in Neural Successor Representations Trained on Natural Language. This suggests that word classes are merely coarse regions within a much richer predictive space where functional groupings appear, tracing constructions like date or title names across different resolutions.

This structural insight is supported by work showing how social context influences instruction following in language models Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models. This research demonstrates that the way instructions are followed changes depending on the social register, suggesting models treat commands as social acts rather than just technical specifications. Furthermore, the concentration of answers across various language models when forced to pick a single word is quite high The One-Word Census: Answer-Choice Conformity Across 105 Language Models. This implies that even highly specialized models tend to converge on certain choices, though heavily post-trained assistants show more conformity than lightly tuned ones.

On the technical side, improving the efficiency of vision models is crucial for real-world deployment; for instance, a confidence adaptive aggregation network achieves better accuracy and speed when estimating object poses from images compared to prior methods Accurate and Efficient Object Pose Estimation via the Aggregation of Diffusion Features. This method uses confidence to select the most discriminative features, leading to substantial gains in performance on unseen objects. Relatedly, techniques for compressing vision tokens are gaining traction; ERASE proposes an adaptive two-stage framework that reduces token count while retaining high accuracy in vision-language models ERASE: Eliminating Redundant Visual Tokens via Adaptive Two-Stage Token Pruning.

Finally, the integration of multiple models is being refined through preference optimization; InfiFPO replaces the reference model in direct preference optimization with a fused source model that synthesizes multi-source probabilities InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models. This approach successfully improves performance on various benchmarks, showing that fusing knowledge from different models can lead to better reasoning capabilities.

Today's papers

The papers

Important terms

LearnPruner
This framework intelligently removes unnecessary visual tokens from vision-language models to reduce computational strain while preserving performance. It uses a two-stage pruning process to cut token count by over ninety percent and speed up inference three point two times.
AKMPC
This research investigates whether the reasoning in vision-language driving agents is truly causal or just a byproduct. It shows that planning heavily relies on initial priors rather than the generated text, highlighting a disconnect in autonomous system reasoning.
PixVL
PixVL introduces a unified Mask--Text Consistency Cycle to bridge region segmentation and understanding. It uses self-verification against geometric reconstruction to learn from unlabeled data, improving both region understanding and segmentation performance.
ShapeY
ShapeY provides a principled way to measure how well object recognition systems grasp shape cues instead of just texture. It probes embedding spaces for 3D object similarity, revealing that current models struggle with robust generalization across different viewpoints.