arXiv subjects

Kristian Kersting

Publications and source records attributed to Kristian Kersting.

At least 55 records · Page 3Linked to original sources

arXiv2026

ActivationReasoning: Logical Reasoning in Latent Activation Spaces

Large language models (LLMs) excel at generating fluent text, but their internal reasoning remains opaque and difficult to control. Sparse autoencoders (SAEs) make hidden activations more interpretable by exposing latent features that often align with human concepts. Yet, these features are fragile and passive, offering no mechanism for systematic reasoning or model control. To address this, we introduce ActivationReasoning (AR), a framework that embeds explicit logical reasoning into the latent space of LLMs. It proceeds in three stages: (1) Finding latent representations, first latent concept representations are identified (e.g., via SAEs) and organized into a dictionary; (2) Activating propositions, at inference time AR detects activating concepts and maps them to logical propositions; and (3)Logical reasoning, applying logical rules over these propositions to infer higher-order structures, compose new concepts, and steer model behavior. We evaluate AR on multi-hop reasoning (PrOntoQA), abstraction and robustness to indirect concept cues (Rail2Country), reasoning over natural and diverse language (ProverQA), and context-sensitive safety (BeaverTails). Across all tasks, AR scales robustly with reasoning complexity, generalizes to abstract and context-sensitive tasks, and transfers across model backbones. These results demonstrate that grounding logical structure in latent activations not only improves transparency but also enables structured reasoning, reliable control, and alignment with desired behaviors, providing a path toward more reliable and auditable AI.

cs.LG↗

arXiv2026

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.

cs.LG↗

arXiv2026

GRAIL: Autonomous Concept Grounding for Neuro-Symbolic Reinforcement Learning

Neuro-symbolic Reinforcement Learning (NeSy-RL) combines symbolic reasoning with gradient-based optimization to achieve interpretable and generalizable policies. Relational concepts, such as "left of" or "close by", serve as foundational building blocks that structure how agents perceive and act. However, conventional approaches require human experts to manually define these concepts, limiting adaptability since concept semantics vary across environments. We propose GRAIL (Grounding Relational Agents through Interactive Learning), a framework that autonomously grounds relational concepts through environmental interaction. GRAIL leverages large language models (LLMs) to provide generic concept representations as weak supervision, then refines them to capture environment-specific semantics. This approach addresses both sparse reward signals and concept misalignment prevalent in underdetermined environments. Experiments on the Atari games Kangaroo, Seaquest, and Skiing demonstrate that GRAIL matches or outperforms agents with manually crafted concepts in simplified settings, and reveals informative trade-offs between reward maximization and high-level goal completion in the full environment.

cs.AI↗

arXiv2026

SocialGrid: A Benchmark for Planning and Social Reasoning in Embodied Multi-Agent Systems

As Large Language Models (LLMs) transition from text processors to autonomous agents, evaluating their social reasoning in embodied multi-agent settings becomes critical. We introduce SocialGrid, an embodied multi-agent environment inspired by Among Us that evaluates LLM agents on planning, task execution, and social reasoning. Our evaluations reveal that even the strongest open model (GPT-OSS-120B) achieves below 60% accuracy in task completion and planning, with agents getting stuck in repetitive behaviors or failing to navigate basic obstacles. Since poor navigation confounds evaluation of social intelligence, SocialGrid offers an optional Planning Oracle to isolate social reasoning from planning deficits. While planning assistance improves task completion, social reasoning remains a bottleneck: agents fail to detect deception at near-random chance regardless of scale, relying on shallow heuristics rather than accumulating behavioral evidence. SocialGrid provides automatic failure analysis and fine-grained metrics, enabling developers to diagnose and improve their agents. We also establish a competitive leaderboard using Elo ratings from adversarial league play.

cs.AI↗

arXiv2026

AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency

Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks. Yet ensuring that the reasoning trace both contributes to and faithfully reflects the processes underlying the model's final answer, rather than merely accompanying it, remains challenging. We introduce AtManRL, a method that leverages differentiable attention manipulation to learn more faithful reasoning through reinforcement learning. By training an additive attention mask that identifies tokens in the CoT crucial for producing correct answers, we derive a saliency reward signal that encourages the model to generate reasoning traces that genuinely influence its final predictions. We integrate this saliency reward with outcome-based rewards within the GRPO framework to jointly optimize for correctness and interpretability. Experiments on GSM8K and MMLU with Llama-3.2-3B-Instruct demonstrate that our approach can identify influential reasoning tokens and enable training more transparent reasoning models.

cs.CL↗

arXiv2026

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.

cs.LG↗

arXiv2026

Boosting deep Reinforcement Learning using pretraining with Logical Options

Deep reinforcement learning agents are often misaligned, as they over-exploit early reward signals. Recently, several symbolic approaches have addressed these challenges by encoding sparse objectives along with aligned plans. However, purely symbolic architectures are complex to scale and difficult to apply to continuous settings. Hence, we propose a hybrid approach, inspired by humans' ability to acquire new skills. We use a two-stage framework that injects symbolic structure into neural-based reinforcement learning agents without sacrificing the expressivity of deep policies. Our method, called Hybrid Hierarchical RL (H^2RL), introduces a logical option-based pretraining strategy to steer the learning policy away from short-term reward loops and toward goal-directed behavior while allowing the final policy to be refined via standard environment interaction. Empirically, we show that this approach consistently improves long-horizon decision-making and yields agents that outperform strong neural, symbolic, and neuro-symbolic baselines.

cs.AI↗

arXiv2026

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.

cs.LG↗

arXiv2026

Right in Time: Reactive Reasoning in Regulated Traffic Spaces

Exact inference in probabilistic First-Order Logic offers a promising yet computationally costly approach for regulating the behavior of autonomous agents in shared traffic spaces. While prior methods have combined logical and probabilistic data into decision-making frameworks, their application is often limited to pre-flight checks due to the complexity of reasoning across vast numbers of possible universes. In this work, we propose a reactive mission design framework that jointly considers uncertain environmental data and declarative, logical traffic regulations. By synthesizing Probabilistic Mission Design (ProMis) with reactive reasoning facilitated by Reactive Circuits (RC), we enable online, exact probabilistic inference over hybrid domains. Our approach leverages the Frequency of Change inherent in heterogeneous data streams to subdivide inference formulas into memoized, isolated tasks, ensuring that only the specific components affected by new sensor data are re-evaluated. In experiments involving both real-world vessel data and simulated drone traffic in dense urban scenarios, we demonstrate that our approach provides orders of magnitude in speedup over ProMis without reactive paradigms. This allows intelligent transportation systems, such as Unmanned Aircraft Systems (UAS), to actively assert safety and legal compliance during operations rather than relying solely on preparation procedures.

cs.RO↗

arXiv2026

Reactive Knowledge Representation and Asynchronous Reasoning

Exact inference in complex probabilistic models often incurs prohibitive computational costs. This challenge is particularly acute for autonomous agents in dynamic environments that require frequent, real-time belief updates. Existing methods are often inefficient for ongoing reasoning, as they re-evaluate the entire model upon any change, failing to exploit that real-world information streams have heterogeneous update rates. To address this, we approach the problem from a reactive, asynchronous, probabilistic reasoning perspective. We first introduce Resin (Reactive Signal Inference), a probabilistic programming language that merges probabilistic logic with reactive programming. Furthermore, to provide efficient and exact semantics for Resin, we propose Reactive Circuits (RCs). Formulated as a meta-structure over Algebraic Circuits and asynchronous data streams, RCs are time-dynamic Directed Acyclic Graphs that autonomously adapt themselves based on the volatility of input signals. In high-fidelity drone swarm simulations, our approach achieves several orders of magnitude of speedup over frequency-agnostic inference. We demonstrate that RCs' structural adaptations successfully capture environmental dynamics, significantly reducing latency and facilitating reactive real-time reasoning. By partitioning computations based on the estimated Frequency of Change in the asynchronous inputs, large inference tasks can be decomposed into individually memoized sub-problems. This ensures that only the specific components of a model affected by new information are re-evaluated, drastically reducing redundant computation in streaming contexts.

cs.AI↗

arXiv2026

Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves

Depth-recurrence facilitates latent reasoning by sharing parameters across depths. However, prior work lacks combined FLOP-, parameter-, and memory-matched baselines, underutilizes depth-recurrence due to partially fixed layer stacks, and ignores the bottleneck of constant hidden-sizes that restricts many-step latent reasoning. To address this, we introduce a modular framework of depth-recurrent attention mixtures (Dreamer), combining sequence attention, depth attention, and sparse expert attention. It alleviates the hidden-size bottleneck through attention along depth, decouples scaling dimensions, and allows depth-recurrent models to scale efficiently and effectively. Across language reasoning benchmarks, our models require 2 to 8x fewer training tokens for the same accuracy as FLOP-, parameter-, and memory-matched SOTA, and outperform ca. 2x larger SOTA models with the same training tokens. We further present insights into knowledge usage across depths, e.g., showing 2 to 11x larger expert selection diversity than SOTA MoEs.

cs.AI↗

arXiv2026

Fodor and Pylyshyn's Legacy: Still No Human-like Systematic Compositionality in Neural Networks

Strong meta-learning capabilities for systematic compositionality are emerging as an important skill for navigating the complex and changing tasks of today's world. However, in presenting models for robust adaptation to novel environments, it is important to refrain from making unsupported claims about the performance of meta-learning systems that ultimately do not stand up to scrutiny. While Fodor and Pylyshyn famously posited that neural networks inherently lack this capacity as they are unable to model compositional representations or structure-sensitive operations, and thus are not a viable model of the human mind, Lake and Baroni recently presented meta-learning as a pathway to compositionality. In this position paper, we critically revisit this claim and highlight limitations in the proposed meta-learning framework for compositionality. Our analysis shows that modern neural meta-learning systems can only perform such tasks, if at all, under a very narrow and restricted definition of a meta-learning setup. We therefore claim that `Fodor and Pylyshyn's legacy' persists, and to date, there is no human-like systematic compositionality learned in neural networks.

cs.AI↗

arXiv2026

Probabilistic Mission Design for Neuro-Symbolic Unmanned Aircraft Systems

Advanced Air Mobility (AAM) is a growing field that demands accurate and trustworthy models of legal concepts and restrictions for navigating Unmanned Aircraft Systems (UAS). In addition, any implementation of AAM needs to face the challenges posed by inherently dynamic and uncertain human-inhabited spaces robustly. Nevertheless, the employment of UAS beyond visual line of sight (BVLOS) is an endearing task that promises to significantly enhance today's logistics and emergency response capabilities. Hence, we propose Probabilistic Mission Design (ProMis), a novel neuro-symbolic approach to navigating UAS within legal frameworks. ProMis is an interpretable and adaptable system architecture that links uncertain geospatial data and noisy perception with declarative, Hybrid Probabilistic Logic Programs (HPLP) to reason over the agent's state space and its legality. To inform planning with legal restrictions and uncertainty in mind, ProMis yields Probabilistic Mission Landscapes (PML). These scalar fields quantify the belief that the HPLP is satisfied across the agent's state space. Extending prior work on ProMis' reasoning capabilities and computational characteristics, we show its integration with potent machine learning models such as Large Language Models (LLM) and Transformer-based vision models. Hence, our experiments underpin the application of ProMis with multi-modal input data and how our method applies to many AAM scenarios.

cs.AI↗

arXiv2026

Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning

Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of NLP tasks, but they remain fundamentally stateless, constrained by limited context windows that hinder long-horizon reasoning. Recent efforts to address this limitation often augment LLMs with an external memory bank, yet most existing pipelines are static and heuristic-driven, lacking a learned mechanism for deciding what to store, update, or retrieve. We present Memory-R1, a reinforcement learning (RL) framework that equips LLMs with the ability to actively manage and utilize external memory through two specialized agents: a Memory Manager that learns structured operations, including ADD, UPDATE, DELETE, and NOOP; and an Answer Agent that pre-selects and reasons over relevant entries. Both agents are fine-tuned with outcome-driven RL (PPO and GRPO), enabling adaptive memory management with minimal supervision. With only 152 training QA pairs, Memory-R1 outperforms strong baselines and generalizes across diverse question types, three benchmarks (LoCoMo, MSC, LongMemEval), and multiple model scales (3B-14B).

cs.CL↗

arXiv2026

EmoNet-Voice: A Fine-Grained, Expert-Verified Benchmark for Speech Emotion Detection

Speech emotion recognition (SER) systems are constrained by existing datasets that typically cover only 6-10 basic emotions, lack scale and diversity, and face ethical challenges when collecting sensitive emotional states. We introduce EMONET-VOICE, a comprehensive resource addressing these limitations through two components: (1) EmoNet-Voice Big, a 5,000-hour multilingual pre-training dataset spanning 40 fine-grained emotion categories across 11 voices and 4 languages, and (2) EmoNet-Voice Bench, a rigorously validated benchmark of 4,7k samples with unanimous expert consensus on emotion presence and intensity levels. Using state-of-the-art synthetic voice generation, our privacy-preserving approach enables ethical inclusion of sensitive emotions (e.g., pain, shame) while maintaining controlled experimental conditions. Each sample underwent validation by three psychology experts. We demonstrate that our Empathic Insight models trained on our synthetic data achieve strong real-world dataset generalization, as tested on EmoDB and RAVDESS. Furthermore, our comprehensive evaluation reveals that while high-arousal emotions (e.g., anger: 95% accuracy) are readily detected, the benchmark successfully exposes the difficulty of distinguishing perceptually similar emotions (e.g., sadness vs. distress: 63% discrimination), providing quantifiable metrics for advancing nuanced emotion AI. EMONET-VOICE establishes a new paradigm for large-scale, ethically-sourced, fine-grained SER research.

cs.CL↗

arXiv2025

LIME: Making LLM Data More Efficient with Linguistic Metadata Embeddings

Pre-training decoder-only language models relies on vast amounts of high-quality data, yet the availability of such data is increasingly reaching its limits. While metadata is commonly used to create and curate these datasets, its potential as a direct training signal remains under-explored. We challenge this status quo and propose LIME (Linguistic Metadata Embeddings), a method that enriches token embeddings with metadata capturing syntax, semantics, and contextual properties. LIME substantially improves pre-training efficiency. Specifically, it adapts up to 56% faster to the training data distribution, while introducing only 0.01% additional parameters at negligible compute overhead. Beyond efficiency, LIME improves tokenization, leading to remarkably stronger language modeling capabilities and generative task performance. These benefits persist across model scales (500M to 2B). In addition, we develop a variant with shifted metadata, LIME+1, that can guide token generation. Given prior metadata for the next token, LIME+1 improves reasoning performance by up to 38% and arithmetic accuracy by up to 35%.

cs.CL↗

arXiv2025

Measuring and Guiding Monosemanticity

There is growing interest in leveraging mechanistic interpretability and controllability to better understand and influence the internal dynamics of large language models (LLMs). However, current methods face fundamental challenges in reliably localizing and manipulating feature representations. Sparse Autoencoders (SAEs) have recently emerged as a promising direction for feature extraction at scale, yet they, too, are limited by incomplete feature isolation and unreliable monosemanticity. To systematically quantify these limitations, we introduce Feature Monosemanticity Score (FMS), a novel metric to quantify feature monosemanticity in latent representation. Building on these insights, we propose Guided Sparse Autoencoders (G-SAE), a method that conditions latent representations on labeled concepts during training. We demonstrate that reliable localization and disentanglement of target concepts within the latent space improve interpretability, detection of behavior, and control. Specifically, our evaluations on toxicity detection, writing style identification, and privacy attribute recognition show that G-SAE not only enhances monosemanticity but also enables more effective and fine-grained steering with less quality degradation. Our findings provide actionable guidelines for measuring and advancing mechanistic interpretability and control of LLMs.

cs.CL↗

arXiv2025

Focused Chain-of-Thought: Efficient LLM Reasoning via Structured Input Information

Recent large language models achieve strong reasoning performance by generating detailed chain-of-thought traces, but this often leads to excessive token use and high inference latency. Existing efficiency approaches typically focus on model-centric interventions, such as reinforcement learning or supervised fine-tuning, to reduce verbosity. In contrast, we propose a training-free, input-centric approach. Inspired by cognitive psychology, we introduce Focused Chain-of-Thought (F-CoT), which separates information extraction from the reasoning process. F-CoT first organizes the essential information from a query into a concise, structured context and then guides the model to reason exclusively over this context. By preventing attention to irrelevant details, F-CoT naturally produces shorter reasoning paths. On arithmetic word problems, F-CoT reduces generated tokens by 2-3x while maintaining accuracy comparable to standard zero-shot CoT. These results highlight structured input as a simple yet effective lever for more efficient LLM reasoning.

cs.CL↗