MASEval: Extending Multi-Agent Evaluation from Models to Systems

The rapid adoption of LLM-based agentic systems has produced a rich ecosystem of frameworks (smolagents, LangGraph, AutoGen, CAMEL, LlamaIndex, i.a.). Yet existing benchmarks are model-centric: they fix the agentic setup and do not compare other system components. We argue that implementation decisions substantially impact performance, including choices such as topology, orchestration logic, and error handling. MASEval addresses this evaluation gap with a framework-agnostic library that treats the entire system as the unit of analysis. Through a systematic system-level comparison across 3 benchmarks, 3 models, and 3 frameworks, we find that framework choice matters as much as model choice. MASEval allows researchers to explore all components of agentic systems, opening new avenues for principled system design, and practitioners to identify the best implementation for their use case. MASEval is available under the MIT licence https://github.com/parameterlab/MASEval.

Half-Truths Break Similarity-Based Retrieval

When a text description is extended with an additional detail, image-text similarity should drop if that detail is wrong. We show that CLIP-style dual encoders often violate this intuition: appending a plausible but incorrect object or relation to an otherwise correct description can increase the similarity score. We call such cases half-truths. On COCO, CLIP prefers the correct shorter description only 40.6% of the time, and performance drops to 32.9% when the added detail is a relation. We trace this vulnerability to weak supervision on caption parts: contrastive training aligns full sentences but does not explicitly enforce that individual entities and relations are grounded. We propose CS-CLIP (Component-Supervised CLIP), which decomposes captions into entity and relation units, constructs a minimally edited foil for each unit, and fine-tunes the model to score the correct unit above its foil while preserving standard dual-encoder inference. CS-CLIP raises half-truth accuracy to 69.3% and improves average performance on established compositional benchmarks by 5.7 points, suggesting that reducing half-truth errors aligns with broader gains in compositional understanding. Code is publicly available at: https://github.com/kargibora/CS-CLIP

Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models

Compositional generalization, the ability to recognize familiar parts in novel contexts, is a defining property of intelligent systems. Although modern models are trained on massive datasets, they still cover only a tiny fraction of the combinatorial space of possible inputs, raising the question of what structure representations must have to support generalization to unseen combinations. We formalize three desiderata for compositional generalization under standard training (divisibility, transferability, stability) and show they impose necessary geometric constraints: representations must decompose linearly into per-concept components, and these components must be orthogonal across concepts. This provides theoretical grounding for the Linear Representation Hypothesis: the linear structure widely observed in neural representations is a necessary consequence of compositional generalization. We further derive dimension bounds linking the number of composable concepts to the embedding geometry. Empirically, we evaluate these predictions across modern vision models (CLIP, SigLIP, DINO) and find that representations exhibit partial linear factorization with low-rank, near-orthogonal per-concept factors, and that the degree of this structure correlates with compositional generalization on unseen combinations. As models continue to scale, these conditions predict the representational geometry they may converge to. Code is available at https://github.com/oshapio/necessary-compositionality.

Universal Algorithm-Implicit Learning

Current meta-learning methods are constrained to narrow task distributions with fixed feature and label spaces, limiting applicability. Moreover, the current meta-learning literature uses key terms like "universal" and "general-purpose" inconsistently and lacks precise definitions, hindering comparability. We introduce a theoretical framework for meta-learning which formally defines practical universality and introduces a distinction between algorithm-explicit and algorithm-implicit learning, providing a principled vocabulary for reasoning about universal meta-learning methods. Guided by this framework, we present TAIL, a transformer-based algorithm-implicit meta-learner that functions across tasks with varying domains, modalities, and label configurations. TAIL features three innovations over prior transformer-based meta-learners: random projections for cross-modal feature encoding, random injection label embeddings that extrapolate to larger label spaces, and efficient inline query processing. TAIL achieves state-of-the-art performance on standard few-shot benchmarks while generalizing to unseen domains. Unlike other meta-learning methods, it also generalizes to unseen modalities, solving text classification tasks despite training exclusively on images, handles tasks with up to 20$\times$ more classes than seen during training, and provides orders-of-magnitude computational savings over prior transformer-based approaches.

Enhancing Multi-Image Understanding through Delimiter Token Scaling

Large Vision-Language Models (LVLMs) achieve strong performance on single-image tasks, but their performance declines when multiple images are provided as input. One major reason is the cross-image information leakage, where the model struggles to distinguish information across different images. Existing LVLMs already employ delimiter tokens to mark the start and end of each image, yet our analysis reveals that these tokens fail to effectively block cross-image information leakage. To enhance their effectiveness, we propose a method that scales the hidden states of delimiter tokens. This enhances the model's ability to preserve image-specific information by reinforcing intra-image interaction and limiting undesired cross-image interactions. Consequently, the model is better able to distinguish between images and reason over them more accurately. Experiments show performance gains on multi-image benchmarks such as Mantis, MuirBench, MIRB, and QBench2. We further evaluate our method on text-only tasks that require clear distinction. The method improves performance on multi-document and multi-table understanding benchmarks, including TQABench, MultiNews, and WCEP-10. Notably, our method requires no additional training or inference cost.

Dynamics Reveals Structure: Challenging the Linear Propagation Assumption

Neural networks adapt through first-order parameter updates, yet it remains unclear whether such updates preserve logical coherence. We investigate the geometric limits of the Linear Propagation Assumption (LPA), the premise that local updates coherently propagate to logical consequences. To formalize this, we adopt relation algebra and study three core operations on relations: negation flips truth values, converse swaps argument order, and composition chains relations. For negation and converse, we prove that guaranteeing direction-agnostic first-order propagation necessitates a tensor factorization separating entity-pair context from relation content. However, for composition, we identify a fundamental obstruction. We show that composition reduces to conjunction, and prove that any conjunction well-defined on linear features must be bilinear. Since bilinearity is incompatible with negation, this forces the feature map to collapse. These results suggest that failures in knowledge editing, the reversal curse, and multi-hop reasoning may stem from common structural limitations inherent to the LPA.

Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models

We identify a novel phenomenon in language models: benign fine-tuning of frontier models can lead to privacy collapse. We find that diverse, subtle patterns in training data can degrade contextual privacy, including optimisation for helpfulness, exposure to user information, emotional and subjective dialogue, and debugging code printing internal variables, among others. Fine-tuned models lose their ability to reason about contextual privacy norms, share information inappropriately with tools, and violate memory boundaries across contexts. Privacy collapse is a ``silent failure'' because models maintain high performance on standard safety and utility benchmarks whilst exhibiting severe privacy vulnerabilities. Our experiments show evidence of privacy collapse across six models (closed and open weight), five fine-tuning datasets (real-world and controlled data), and two task categories (agentic and memory-based). Our mechanistic analysis reveals that privacy representations are uniquely fragile to fine-tuning, compared to task-relevant features which are preserved. Our results reveal a critical gap in current safety evaluations, in particular for the deployment of specialised agents.

Un-Attributability: Computing Novelty From Retrieval & Semantic Similarity

Understanding how language-model outputs relate to the pretraining corpus is central to studying model behavior. Most training data attribution (TDA) methods ask which training examples causally influence a given output, often using leave-one-out tests. We invert the question: which outputs cannot be attributed to any pretraining example? We introduce un-attributability as an operational measure of semantic novelty: an output is novel if the pretraining corpus contains no semantically similar context. We approximate this with a simple two-stage retrieval pipeline: index the corpus with lightweight GIST embeddings, retrieve the top-n candidates, then rerank with ColBERTv2. If the nearest corpus item is less attributable than a human-generated text reference, we consider the output of the model as novel. We evaluate on SmolLM and SmolLM2 and report three findings: (1) models draw on pretraining data across much longer spans than previously reported; (2) some domains systematically promote or suppress novelty; and (3) instruction tuning not only alters style but also increases novelty. Reframing novelty assessment around un-attributability enables efficient analysis at pretraining scale. We release ~20 TB of corpus chunks and index artifacts to support replication and large-scale extension of our analysis at https://huggingface.co/datasets/stai-tuebingen/faiss-smollm

What Defines Good Reasoning in LLMs? Dissecting Reasoning Steps with Multi-Aspect Evaluation

Evaluating large language models (LLMs) on final-answer correctness is the dominant paradigm. This approach, however, provides a coarse signal for model improvement and overlooks the quality of the underlying reasoning process. We argue that a more granular evaluation of reasoning offers a more effective path to building robust models. We decompose reasoning quality into two dimensions: relevance and coherence. Relevance measures if a step is grounded in the problem; coherence measures if it follows logically from prior steps. To measure these aspects reliably, we introduce causal stepwise evaluation (CaSE). This method assesses each reasoning step using only its preceding context, which avoids hindsight bias. We validate CaSE against human judgments on our new expert-annotated benchmarks, MRa-GSM8K and MRa-MATH. More importantly, we show that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance. Our work provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating the practical value of moving beyond validity checks.

Dr.LLM: Dynamic Layer Routing in LLMs

Large Language Models (LLMs) process every token through all layers of a transformer stack, causing wasted computation on simple queries and insufficient flexibility for harder ones that need deeper reasoning. Adaptive-depth methods can improve efficiency, but prior approaches rely on costly inference-time search, architectural changes, or large-scale retraining, and in practice often degrade accuracy despite efficiency gains. We introduce Dr.LLM, Dynamic routing of Layers for LLMs, a retrofittable framework that equips pretrained models with lightweight per-layer routers deciding to skip, execute, or repeat a block. Routers are trained with explicit supervision: using Monte Carlo Tree Search (MCTS), we derive high-quality layer configurations that preserve or improve accuracy under a compute budget. Our design, windowed pooling for stable routing, focal loss with class balancing, and bottleneck MLP routers, ensures robustness under class imbalance and long sequences. On ARC (logic) and DART (math), Dr.LLM improves accuracy by up to +3.4%p while saving 5 layers per example on average. Routers generalize to out-of-domain tasks (MMLU, GSM8k, AIME, TruthfulQA, SQuADv2, GPQA, PIQA, AGIEval) with only 0.85% accuracy drop while retaining efficiency, and outperform prior routing methods by up to +7.7%p. Overall, Dr.LLM shows that explicitly supervised routers retrofit frozen LLMs for budget-aware, accuracy-driven inference without altering base weights.