Soft-Prompt Tuning for Fair and Efficient LLM Benchmark Evaluation

Benchmark scores often misrepresent a large language model's (LLM's) knowledge, because they rely, e.g., on the model's ability to follow specific formatting requirements. This especially penalizes base models that may know the correct answers but lack the ability -- typically introduced in post-training -- to structure them as instructed. To overcome this, we propose soft-prompt tuning, an efficient, fair, and architecture-agnostic model evaluation. By optimizing only 10 soft-prompt vectors (roughly 0.0006% parameters for a 7B model) over a short tuning period, we adapt models to specific benchmark formats, closing gaps in format-following and ensuring that underlying knowledge is accurately reflected in benchmark scores. This allows one to fairly compare different base models -- trained with various pre-training recipes -- on benchmarks without the need for full post-training. We evaluated soft-prompt tuning across 7 models and 7 datasets. The results show that (a) soft-prompt tuning saturates format-following within 80 steps (~640 samples) making it highly efficient, (b) soft-prompt tuning significantly outperforms zero- and few-shot prompting, surfacing base model knowledge that standard prompting misses, that (c) even post-trained models can benefit from soft-prompts to maximize format compliance, and that (d) soft-prompted base model performance predicts post-trained model rankings more reliably than zero- and few-shot baselines, offering a low-cost proxy for downstream model quality. Our contributions include (1) metrics which disentangle format-following and knowledge accuracy, (2) a fairer benchmarking protocol of LLM knowledge, and (3) a cost- and memory-effective recipe to identify optimal pre-training strategies early in LLM development.

Learning Explicit Behavioral Models with Adaptive Questions and World-Model Probes

Interactive agents trained only against task return can achieve high scores while failing to represent the mechanisms that make their actions succeed. This makes brittle behavior difficult to diagnose and limits adaptation when environment dynamics change. Existing LLM reflection and policy-code repair can revise behavior from failed trajectories, but questions and world-understanding tests are usually used only after training. We introduce an Explicit Symbolic Behavioral Model (ESBM), a trainable behavioral model that couples task performance with evidence-grounded question answering and executable mechanism prediction. An ESBM represents behavior through typed predicates, weighted rules, bounded options and mechanism memory; the mechanism layer predicts symbolic events, object changes, rewards and terminal consequences under action interventions. After each rollout, adaptive questions and active world-model probes convert score failures, QA errors and transition-prediction errors into constraints for local ESBM edits. Candidate models are selected by a multi-criterion rule that jointly evaluates task score, answerability and active world-model consistency. Under the tested Atari-style protocols, ESBM learns high-scoring policies while producing explicit answers and executable mechanism predictions, indicating that adaptive questions can serve as both training pressure and reusable benchmarks for mechanistic policy learning in this setting.

KletterMix: Climbing Toward High-Quality German Pretraining Data

High-quality pretraining data is a central ingredient in modern language models, but German-language resources remain far less developed than their English counterparts: they are often smaller, less carefully curated, weakly documented, and rarely validated through controlled training experiments. We introduce KletterMix, a high-quality German corpus for language model pretraining and annealing, designed as a reusable dataset artifact for the natural language processing and modeling community. KletterMix is built by translating a state-of-the-art English pretraining corpus into German while preserving document boundaries, metadata, source structure, and topical diversity. This construction yields a German corpus with the scale and diversity of a modern pretraining dataset, while enabling direct comparison to its English source. We document the dataset through a broad set of corpus-level analyses, including translation quality, document length distributions, topic coverage, source composition, and geographic metadata. Using COMETKiwi, we show that the translated documents achieve strong quality across diverse domains, suggesting that careful translation can preserve much of the semantic and stylistic richness of the original corpus. Beyond dataset construction, we evaluate KletterMix as training data. Through controlled pretraining and annealing ablations against established German corpora, we show that models trained on KletterMix achieve measurable improvements on German-language downstream evaluations. These results demonstrate that carefully curated translated data can substantially strengthen the German pretraining data ecosystem.

No Safe Dose: How Training Data Drives Unsafe Image Generation

Text-to-image models trained on large-scale data often inevitably ingest unsafe content. While some people observe input-output amplifications, it remains unclear whether and how training data composition directly drives model output safety or by other factors. We shed light on this question by isolating this variable: we train the same text-to-image model on datasets that differ \emph{only} in their fraction of unsafe images (0\% to 9.6\%), across several dataset scales (100K to 8M). Then we generate images with the resulting models, and evaluate them with four independent safety classifiers. Output unsafety rises monotonically from 16.6\% at 0\% contamination to 25.5\% at 5\%. A factorial design reveals that the \emph{proportion}, not the absolute count, of unsafe training images is the operative variable. The 16.6\% irreducible baseline at zero contamination implicates the other components, e.g. frozen text encoder, as a residual safety risk -- confirmed by a text encoder ablation showing that SafeCLIP reduces this floor to 9.6\%, while the dose-response effect persists across all three encoders tested. Critically, no quality degradation in terms of FID, CLIPscore and ImageReward accompanies safety filtering. These results establish that data curation and text encoder safety are complementary and independently effective interventions. At the same time, the remaining level of unsafety poses questions for future research about emerging capabilities and compositionality.

ORCA: An End-to-End Interactive Copilot for Optimized Root Cause Analysis

Causal analysis is a crucial task in many domains, including manufacturing, social science, and medicine. However, despite recent progress, the conceptual and methodological complexity of causal methods makes them largely inaccessible to domain experts. This gap prevents experts from leveraging these advances and hinders researchers who lack access to real-world data for validation. To bridge this divide, we introduce ORCA, a copilot for end-to-end causal analysis. ORCA orchestrates agents to understand the user's goals and guide them through the most appropriate causal analysis workflow, from fully automatic to highly user-guided execution. It features causal discovery, causal effect estimation, explainability and Root-Cause-Analysis (RCA). ORCA evaluates and compares performance, generates key metrics and diagrams, and generates insights through structured reports. We highlight its effectiveness across several real-world use-cases.

Kintsugi: Learning Policies by Repairing Executable Knowledge Bases

Modern embodied agents achieve impressive performance, but their task knowledge is often stored in neural weights, latent state, or prompt-bound memory, making individual policy knowledge difficult to inspect, validate, recombine, and reuse. We introduce \textbf{Kintsugi}, a white-box policy-learning framework that treats embodied policy improvement as verifier-gated construction of a typed executable Knowledge Base (KB). Kintsugi represents task-level policy knowledge as composable typed entries -- predicates, operators, policy schemas, monitors, recovery rules, experience records, and goals -- and improves this artifact through localized typed edits induced from rollout evidence, rather than relying on test-time language-model reasoning. Between rollouts, a tool-constrained agentic editing loop diagnoses trajectory failures, localizes them to editable KB layers, and proposes candidate edits. A deterministic verification gate admits an edit only when the candidate type-checks, the resulting KB executes, and focused validation success or trajectory-health metrics improve without violating protected-regression checks. At inference, the accepted KB is executed by a deterministic symbolic executor with zero LLM calls. Across long-horizon text-agent benchmarks and representative object-centric manipulation settings, Kintsugi achieves strong endpoint performance while preserving inspectability, local editability, and verifier-gated deployment. These results suggest that embodied policy improvement can be organized around executable task knowledge.

All Eyes on the Workflow: Automated and Efficient Event Discovery from Video Streams

Disciplines such as business process management and process mining aid organizations by discovering insights about processes on the basis of recorded event data. However, an obstacle to process analysis is data multi-modality: for instance, data in video form are not directly interpretable as events. In this work, we present SnapLog, an approach to extract event data from videos by converting frames to feature vectors using image embeddings and performing temporal segmentation through frame-wise similarity matrices. A generalized few-shot classification is then used to assign labels to the video segments, yielding labeled, timestamped sub-sequences of frames that are interpretable as events. Conventional process mining techniques can be used to analyze the resulting data. We show that our approach produces logs that accurately reflect the process in the videos.

LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking

As reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for scaling reasoning capabilities in LLMs, a new failure mode emerges: LLMs gaming verifiers. We study this phenomenon on inductive reasoning tasks, where models must induce and output logical rules. We find that RLVR-trained models systematically abandon rule induction. Instead of learning generalizable patterns (e.g., ``trains carrying red cars go east''), they enumerate instance-level labels, producing outputs that pass verifiers without capturing the relational patterns required by the task. We show that this behavior is not a failure of understanding but a form of reward hacking: imperfect verifiers that check only extensional correctness admit false positives. To detect such shortcuts, we introduce Isomorphic Perturbation Testing (IPT), which evaluates a single model output under both extensional and isomorphic verification, where the latter enforces invariance under logically isomorphic tasks. While genuine rule induction remains invariant, shortcut strategies fail. We find that shortcut behavior is specific to RLVR-trained reasoning models (e.g., GPT-5, Olmo3) and absent in non-RLVR models (e.g., GPT-4o, GPT-4.5, Ministral). Moreover, shortcut prevalence increases with task complexity and inference-time compute. In controlled training experiments, extensional verification directly induces shortcut strategies, while isomorphic verification eliminates them. These results show that RLVR can incentivize reward hacking not only through overt manipulation but also by exploiting what the verifier fails to enforce.

Context-specific Credibility-aware Multimodal Fusion with Conditional Probabilistic Circuits

Multimodal fusion requires integrating information from multiple sources that may conflict depending on context. Existing fusion approaches typically rely on static assumptions about source reliability, limiting their ability to resolve conflicts when a modality becomes unreliable due to situational factors such as sensor degradation or class-specific corruption. We introduce C$^2$MF, a context-specfic credibility-aware multimodal fusion framework that models per-instance source reliability using a Conditional Probabilistic Circuit (CPC). We formalize instance-level reliability through Context-Specific Information Credibility (CSIC), a KL-divergence-based measure computed exactly from the CPC. CSIC generalizes conventional static credibility estimates as a special case, enabling principled and adaptive reliability assessment. To evaluate robustness under cross-modal conflicts, we propose the Conflict benchmark, in which class-specific corruptions deliberately induce discrepancies between different modalities. Experimental results show that C$^2$MF improves predictive accuracy by up to 29% over static-reliability baselines in high-noise settings, while preserving the interpretability advantages of probabilistic circuit-based fusion.

Recursive Inference Machines for Neural Reasoning

Neural reasoners such as Tiny Recursive Models (TRMs) solve complex problems by combining neural backbones with specialized inference schemes. Such inference schemes have been a central component of stochastic reasoning systems, where inference rules are applied to a stochastic model to derive answers to complex queries. In this work, we bridge these two paradigms by introducing Recursive Inference Machines (RIMs), a neural reasoning framework that explicitly incorporates recursive inference mechanisms inspired by classical inference engines. We show that TRMs can be expressed as an instance of RIMs, allowing us to extend them through a reweighting component, yielding better performance on challenging reasoning benchmarks, including ARC-AGI-1, ARC-AGI-2, and Sudoku Extreme. Furthermore, we show that RIMs can be used to improve reasoning on other tasks, such as the classification of tabular data, outperforming TabPFNs.