arXiv subjects

Michael Granitzer

Publications and source records attributed to Michael Granitzer.

At least 19 recordsLinked to original sources

arXiv2026

Getting a Grip: Operating Human-Agent Software Teams Through Governed Messages

Software is increasingly developed by teams in which several people direct many coding agents from different vendors. Such teams must establish who may request what from whom, which messages start work, which decisions bind, and who caused a change. We propose to govern a human-agent team through its message protocol and present yoars, a relay in which humans and agents are authenticated principals that exchange typed, addressed messages in rooms holding goals, decision rules, routing and an attributable, append-only record. This report gives the architecture, the full controlled study and the field material. In 926 controlled runs (899 analyzable), small agent teams built programs under varied room rules, some with a hostile member planting a backdoor. Revising which messages start work, bundled with limits on repeated messages, was followed by a drop in total coordination failure from 42% to 13% (p = .0063, adjusted p = .057 over nine registered tests). Instructed roles and votes did not reliably produce the intended process, and no defense inside the conversation reliably prevented the backdoor. An auditor read the records of 96 runs with a confirmed backdoor. In the 25 runs where the attacker had persuaded a teammate to write it, the auditor named the persuader in all 25, but it also flagged 24 of 40 honest runs. In a four-week field study, up to nineteen agents and two people developed a distributed system through yoars, producing 7,457 acts. Human attention and decision making emerged as the critical resource: requests waited mainly for the human-run coordinator and for approvals, while agents answered within minutes. Most governance requirements emerged from operation, and holding context dominated token cost. We report twenty-six failures of our own instruments, derive seven lessons and a checklist, and release the corpus, the redacted room ledger and the analysis scripts.

cs.SE↗

arXiv2026

A Tool in the Forward Pass: Frozen Tape Modules as Layers of a Small Language Model

Calling a tool through generated text adds decoding steps, even when the computation itself is short. We study an alternative: a small neural module that runs inside a language model's forward pass. A learned writer turns hidden states into a symbolic tape, a short sequence of discrete symbols that encodes the computation. The module, trained separately and then frozen, reduces the tape to a result, and a learned reader returns that result to the model. We evaluate arithmetic and object tracking in gemma-3-1b-it. With a writer trained on labelled tapes, the layer answers all 750 items of BIG-Bench Hard's tracking shuffled objects correctly. At seven people it needs about 1/10 of the GPU kernel time of a text-tool baseline that uses the same host model and exact code, with accuracy 1.00 against 0.99. The writer can also be learned from answers alone, guided by a hand-written copying prior, but less reliably: 5 of 6 runs succeed in the 1B host and 2 of 4 in the 4B host. Two independently trained layers keep their task accuracy when attached to one model, and a learned gate keeps them silent on the tested general-text and look-alike prompts. The results support a narrow use case: a bounded computation that returns one symbol can avoid the decoding cost of a textual tool call. Extracting the right input reliably, and deciding when to act, remain the main open problems.

cs.AI↗

arXiv2026

System One models Are Not Better Rerankers

System One models, or decision models, answer a question about an input with a calibrated probability instead of generated text. Within weeks of the first release, in September 2026, they were used as rerankers. We test whether the decision fine-tuning behind them makes a better reranker than plain reranker fine-tuning of the same model. For six Qwen backbones from 0.6B to 27B parameters and ten TREC DL and BEIR collections, we compare the untuned model, asked zero-shot, with its decision fine-tunes and its reranker fine-tunes, two of which we train ourselves. In our work, we draw the conclusion that decision fine-tuning never makes the better reranker. From 0.6B to 9B, reranker fine-tuning wins on every backbone by 5 to 29 points nDCG@10, also on zero-shot collections, and at 27B the two tie. Up to 9B, decision fine-tuning adds nothing over simply asking the untuned model. Moreover, we find that the calibration advantage of decision models depends on the question. Asked whether a passage answers the query, they are better calibrated than the published reranker fine-tunes from 4B on. Asked a slightly different question from prior work, they lose this advantage. Lastly, calibration does not make up for a worse ranking: up to 9B, a reranker fine-tune keeps the better passages at a fixed threshold. If you need a reranker, fine-tune a reranker.

cs.CL↗

arXiv2026

AutoSim: A Self-Improving User Simulator for Conversational Search Evaluation

User simulators promise to evaluate conversational search systems at a scale human studies cannot reach, but only if their behavior is realistic enough to trust. We study AutoSim, a conversational-search user simulator whose relevance judgments rank a controlled pool of retrieval systems in agreement with NIST across three TREC iKAT years (Kendall's $τ$ of $0.66$, $0.91$, and $0.91$; all $p{\le}0.002$), an agreement that survives removing its persona. AutoSim self-improves against a held-out signal from real users, and we find that fine-tuning its weights makes it behave like real users, which in turn gives a more accurate estimate of how well a system serves them. Reaching that behavior takes more than prompting: when discriminating between real and simulated user sessions, no prompt we tried lowers a held-out real-versus-simulated classifier below AUC $0.85$, whereas fine-tuning the weights drives it to $0.45$, where neither the classifier nor a human reading full dialogues can tell it from a real user. That added realism is what makes the evaluation more accurate: where real users retrieve little, a prompted simulator over-estimates retrieval effectiveness by $+0.136$ nDCG and returns a relevant passage on $57\%$ of the turns where real users find none, enough to change a deployment decision. Only the weight-level simulator escapes this bias. We release the framework and code.

cs.IR↗

arXiv2026

The Fellowship of the Query: Learning Retrieval Actions

Retrieval-augmented question answering requires control decisions about when to decompose a question, search, reformulate, extract evidence, synthesize facts, verify progress, and stop. We study whether trajectory fine-tuning can improve small language models (SLMs) as next-action controllers. We additionally evaluate a low-resource setting in which a single SLM serves as both the controller and the final-answer generator. From accepted teacher search traces, we build a seven-way action-prediction task, where the model predicts the next structured teacher action from the current trajectory state, and evaluate LoRA-supervised fine-tuning across SLMs and xSLMs as controllers. On 1,646 held-out action examples, Granite 4.1 3B trained on 13,194 actions reaches macro-F1 0.6536, compared with 0.1736 for zero-shot prompting of the same model and 0.5399 for a TF-IDF logistic-regression baseline. In an end-to-end controller/generator swap evaluation over 149 held-out trajectories, using the fine-tuned model for both roles improves Exact Match from 0.7530 to 0.7946 and token F1 from 0.7783 to 0.8295 compared with using the base model as both controller and generator. The cross-role conditions show that the fine-tuned controller increases evidence-fact recording when the generator is fixed, while controller-only final-answer gains are not statistically clear. Overall, trajectory supervision improves action prediction and evidence-recording behaviour in this evaluated pipeline. Code is available at https://github.com/padas-lab-de/agent-action-controller

cs.IR↗

arXiv2026

The WebKurator.de Platform: Combined Regional and Topical Web Curation

The systematic curation of the Web remains a central challenge for national libraries and memory institutions that aim to preserve culturally and regionally relevant content. Existing directory-based approaches such as Curlie implement a predominantly topic-centric, one-dimensional hierarchy, where geographic aspects are intertwined with topical and linguistic categories. To address this limitation, we present WebKurator.de, a collaborative platform for combined regional and topical web curation, initially focused on the German web. WebKurator introduces a two-dimensional curation model that explicitly separates topical categorization and geographic annotation. The system integrates LLM-based topic classification and imprint-based address extraction with geocoding, and supports user suggestions together with moderated review. The platform is bootstrapped from the German Imprints Dataset, a large-scale collection of 5.54 million websites. Among them, 3.14 million contain imprint pages, for which we successfully extracted and geocoded postal addresses. Of these, 2.58 million (85.17%) are located in Germany and also have an assigned topic label. These websites form the initial foundation of WebKurator.de and can be continuously extended through user suggestions.

cs.IR↗

arXiv2026

The Compaction Cliff in Long-Running AI Agent Memory

A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53\% of safety rules after one compaction round and 10\% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operators implement this triage across the three context-management operations: TypeCompact rewrites items in place under per-type fidelity, TypeDecompose partitions a topic too large to compact safely, replicating in-scope safety rules across partitions, and TypeRetrieve fetches items from external storage with in-scope rules pinned ahead of relevance. On five public corpora, TypeCompact preserves 2--4$\times$ more safety rules than the strongest single-shot LLM compactor at every ratio, with 96\% recall over five rounds. TypeDecompose reaches 0\% locality violations against 93\% under uniform partitioning. TypeRetrieve reaches 100\% recall@50 against 73\% for the best single-shot LLM retriever. On three downstream behavioral benchmarks, we outperform the production Sonnet compactor on medical compliance (paired McNemar $p < 10^{-8}$ on preservation, $N = 200$), the full-policy and hierarchical baselines on retail task pass rate ($p < 0.01$, $N = 115$), and the hierarchical compaction on the airline domain ($p = 0.024$). We release AgentArtifactCorpus (396{,}934 agent configurations from 54{,}628 public GitHub repositories), the classifier, and the reference implementation.

cs.AI↗

arXiv2026

When Is Complex Chunking Worth It? A Multi-Objective Evaluation of Chunking Methods at Scale

Dense retrieval is commonly evaluated on benchmarks that represent each document with a single embedding, even though real-world retrieval systems often index long documents that require chunking. In these settings, the chosen chunking method not only affects retrieval quality, but also indexing throughput, query latency, and memory usage. Prior comparisons of chunking strategies have mainly focused on retrieval performance, leaving operational trade-offs underexplored. To address these issues, we evaluate eight representative chunking strategies across two scalable corpora, three embedding models, and multiple corpus sizes, measuring both retrieval effectiveness and system-level costs. Our results show that computationally expensive methods rarely provide consistent gains over simpler chunking. Instead, the best performing strategy depends on the embedding model, dataset, corpus size, and target retrieval metric. Methods with similar performance can also differ substantially in operational cost, showing that chunking should be seen as a multi-objective design decision.

cs.IR↗

arXiv2026

Metadata, Structure, or Strategy? A Decomposition of RAG Context Enrichment

Retrieval-augmented generation (RAG) systems increasingly enrich retrieved passages by attaching quality metadata, structuring them into explicit records, and adopting multi-hop retrieval strategies that accumulate evidence across steps. These changes assume that richer context yields better answers, yet existing evaluations cannot test this because they vary all three factors at once. We isolate each factor in a controlled experiment across six benchmarks, four models from three families, and five enrichment levels, totaling over 24,000 evaluated responses. The assumption does not hold. Most enrichment reduces accuracy. Models prompted to use confidence scores comply correctly yet produce worse answers, a gap between utilization and accuracy that no prior work has measured. What determines answer quality is not how much metadata the context carries but whether the model can act on it for the given task. When metadata and retrieval strategy are aligned with model capabilities, a smaller model outperforms a frontier model by 19 F1 points. These findings motivate a processability hierarchy that predicts, from pre-training properties alone, which metadata a model can productively use, reframing RAG design as a question of model-context alignment rather than metadata accumulation.

cs.IR↗

arXiv2026

As We May Search

The sensitive information in personal documents, legal files, and medical records is among the most valuable things to search, yet current retrieval-augmented generation systems still require sending content to remote servers. We propose local-first IR, a design philosophy where indexes, models, and inference reside on user devices, treating remote services as optional. This paper makes four contributions: (1) a framework organizing retrieval architectures along three dimensions: privacy and control, capability, and accessibility, (2) experiments on consumer hardware across five benchmarks, scaling from 1K to 1M documents with dense retrieval, BM25, and hybrid fusion. Dense retrieval keeps over 91% nDCG@10 up to 100K documents, with approximate HNSW indexes extending this to 1M with only 2% quality loss; a 7B local language model reaches within 4 points of a cloud baseline on answer quality, (3) competing perspectives for and against local-first IR, informed by experimental evidence, and (4) a research agenda identifying open problems. The real tradeoff is scope rather than quality: what matters is what you can search, not how well you can search it.

cs.IR↗

arXiv2026

GradeLegal: Automated Grading for German Legal Cases

Grading German legal exam solutions faces growing volumes and a shortage of qualified graders, delaying feedback and creating a bottleneck. At the same time, it is a high-stakes expert task, since state exam grades strongly influence career outcomes in Germany. Despite this practical relevance, literature lacks systematic studies on effective methods for grading legal exams. To address this gap, we investigate whether large language models (LLMs) can support the automated grading of German legal case solutions in criminal and public law, thereby enabling scalable feedback and student self-testing. We present a systematic evaluation of 27 proprietary and open-source LLMs, benchmarking prompting strategies that incrementally add task-related information, such as a sample solution and a grading rubric. Using quadratic weighted kappa (QWK), reasoning-oriented LLMs can approximate expert grading in public law when given a sample solution and a grading rubric (up to 0.91), compared to 0.60 in criminal law, suggesting a harder grading task in criminal law. Beyond single-model grading, ensembling improves agreement by up to 0.15 over its best member and can offer an alternative to stronger closed-source single models. In addition, our findings suggest that effective prompt design and model selection are necessary for reliable LLM-based grading of legal exams.

cs.CL↗

arXiv2026

NuggetIndex: Governed Atomic Retrieval for Maintainable RAG

Retrieval-augmented generation (RAG) systems are frequently evaluated via fact-based metrics, yet standard implementations retrieve passages or static propositions. This unit mismatch between evaluation and retrieval objects hinders maintenance when corpora evolve and fails to capture superseded facts or source disagreements. We propose NuggetIndex, a retrieval system that stores atomic information units as managed records, so called nuggets. Each record maintains links to evidence, a temporal validity interval, and a lifecycle state. By filtering invalid or deprecated nuggets prior to ranking, the system prevents the inclusion of outdated information. We evaluate the approach using a nuggetized MS MARCO subset, a temporal Wikipedia QA dataset, and a multi-hop QA task. Against passage and unmanaged proposition retrieval baselines, NuggetIndex improves nugget recall by 42%, increases temporal correctness by 9 percentage points without the recall collapse observed in time-filtered baselines, and reduces conflict rates by 55%. The compact nugget format reduces generator input length by 64% while enabling lightweight index structures suitable for browser-based and resource-constrained deployment. We release our implementation, datasets, and evaluation scripts

cs.IR↗

arXiv2026

AgentSim: A Platform for Verifiable Agent-Trace Simulation

Training trustworthy agentic LLMs requires data that shows the grounded reasoning process, not just the final answer. Existing datasets fall short: question-answering data is outcome-only, chain-of-thought data is not tied to specific documents, and web-agent datasets track interface actions rather than the core retrieval and synthesis steps of a RAG workflow. We introduce AgentSim, an open-source platform for simulating RAG agents. It generates verifiable, stepwise traces of agent reasoning over any document collection. AgentSim uses a policy to ensure the agent widely explores the document set. It combines a multi-model validation pipeline with an active human-in-the-loop process. This approach focuses human effort on difficult steps where models disagree. Using AgentSim, we construct and release the Agent-Trace Corpus (ATC), a large collection of grounded reasoning trajectories spanning three established IR benchmarks. We make three contributions: (1) the AgentSim platform with two mechanisms, Corpus-Aware Seeding and Active Validation, that improve trace diversity and quality; (2) the Agent-Trace Corpus (ATC), over 103,000 verifiable reasoning steps spanning three IR benchmarks, with 100% grounding rate on substantive answers; and (3) a comparative behavioral analysis revealing systematic differences in how state-of-the-art models approach information seeking. Platform, toolkit, and corpus are publicly available.

cs.IR↗

arXiv2026

IIRSim Studio: A Dashboard for User Simulation

User simulation is a valuable methodology for evaluation in Information Retrieval (IR), enabling low-cost experimentation and counterfactual analysis. However, existing simulation frameworks are primarily code-centric libraries that require substantial setup effort, which limits adoption and hinders reproducibility. The bottleneck is not the simulation engines themselves, but the lack of infrastructure connecting experiment design, execution, and sharing into a single verifiable workflow. This paper introduces IIRSim Studio, a web-based workbench that addresses this gap through four contributions: (1) a visual environment for composing simulation pipelines on top of simulation frameworks, serving both novices learning simulation concepts and experts piloting large-scale experiments; (2) a component lifecycle that supports authoring, versioning, and sharing custom simulation components through Git-backed storage and runtime injection; (3) a provenance model based on experiment bundles and environment templates that makes the scope of replication explicit; and (4) a shared-task workflow, demonstrated through the re-deployment of a Sim4IA micro-task. IIRSim Studio is available as a hosted service and as a portable containerized deployment.

cs.IR↗

arXiv2026

Beyond the Click: A Framework for Inferring Cognitive Traces in Search

User simulators are essential for evaluating search systems, but they primarily reproduce user actions without modeling the underlying thought process. Large-scale interaction logs record what users do, but not what they might be thinking or feeling, such as confusion or satisfaction. We present a framework for inferring cognitive traces from behavioral logs. Our method uses a multi-agent LLM system grounded in Information Foraging Theory (IFT) and validated by human experts. We annotate three public datasets (AOL, Stack Overflow, and MovieLens), producing over 530,000 cognitive labels across 50,000 sessions. A cross-dataset evaluation with a shuffled-label control reveals that cognitive labels provide the strongest signal where behavioral features are weakest: on MovieLens, the cognitive model improves F1 by up to 6.6% over the behavioral baseline and 1.8% above the shuffled control, while on AOL, where click patterns are highly predictive, improvements are near zero. We release the annotation collection on HuggingFace, an open-source annotation tool, and all experimental code to support future work on cognitively aware user simulation.

cs.IR↗

arXiv2026

Toward Generalized Cross-Lingual Hateful Language Detection with Web-Scale Data and Ensemble LLM Annotations

We study whether large-scale unlabelled web data and LLM-based synthetic annotations can improve multilingual hate speech detection. Starting from texts crawled via OpenWebSearch.eu~(OWS) in four languages (English, German, Spanish, Vietnamese), we pursue two complementary strategies. First, we apply continued pre-training to BERT models by continuing masked language modelling on unlabelled OWS texts before supervised fine-tuning, and show that this yields an average macro-F1 gain of approximately 3% over standard baselines across sixteen benchmarks, with stronger gains in low-resource settings. Second, we use four open-source LLMs (Mistral-7B, Llama3.1-8B, Gemma2-9B, Qwen2.5-14B) to produce synthetic annotations through three ensemble strategies: mean averaging, majority voting, and a LightGBM meta-learner. The LightGBM ensemble consistently outperforms the other strategies. Fine-tuning on these synthetic labels substantially benefits a small model (Llama3.2-1B: +11% pooled F1), but provides only a modest gain for the larger Qwen2.5-14B (+0.6%). Our results indicate that the combination of web-scale unlabelled data and LLM-ensemble annotations is the most valuable for smaller models and low-resource languages.

cs.CL↗

arXiv2026

Form Without Function: Agent Social Behavior in the Moltbook Network

Moltbook is a social network where every participant is an AI agent. We analyze 1,312,238 posts, 6.7~million comments, and over 120,000 agent profiles across 5,400 communities, collected over 40 days (January 27 to March 9, 2026). We evaluate the platform through three layers. At the interaction layer, 91.4% of post authors never return to their own threads, 85.6% of conversations are flat (no reply ever receives a reply), the median time-to-first-comment is 55 seconds, and 97.3% of comments receive zero upvotes. Interaction reciprocity is 3.3%, compared to 22-60% on human platforms. An argumentation analysis finds that 64.6% of comment-to-post relations carry no argumentative connection. At the content layer, 97.9% of agents never post in a community matching their bio, 92.5% of communities contain every topic in roughly equal proportions, and over 80% of shared URLs point to the platform's own infrastructure. At the instruction layer, we use 41 Wayback Machine snapshots to identify six instruction changes during the observation window. Hard constraints (rate limit, content filters) produce immediate behavioral shifts. Soft guidance (``upvote good posts'', ``stay on topic'') is ignored until it becomes an explicit step in the executable checklist. The platform also poses technological risks. We document credential leaks (API keys, JWT tokens), 12,470 unique Ethereum addresses with 3,529 confirmed transaction histories, and attack discourse ranging from template-based SSH brute-forcing to multi-agent offensive security architectures. These persist unmoderated because the quality-filtering mechanisms are themselves non-functional. Moltbook is a socio-technical system where the technical layer responds to changes, but the social layer largely fails to emerge. The form of social media is reproduced in full. The function is absent.

cs.SI↗

arXiv2026

Behind the Prompt: The Agent-User Problem in Information Retrieval

User models in information retrieval rest on a foundational assumption that observed behavior reveals intent. This assumption collapses when the user is an AI agent privately configured by a human operator. For any action an agent takes, a hidden instruction could have produced identical output - making intent non-identifiable at the individual level. This is not a detection problem awaiting better tools; it is a structural property of any system where humans configure agents behind closed doors. We investigate the agent-user problem through a large-scale corpus from an agent-native social platform: 370K posts from 47K agents across 4K communities. Our findings are threefold: (1) individual agent actions cannot be classified as autonomous or operator-directed from observables; (2) population-level platform signals still separate agents into meaningful quality tiers, but a click model trained on agent interactions degrades steadily (-8.5% AUC) as lower-quality agents enter training data; (3) cross-community capability references spread endemically ($R_0$ 1.26-3.53) and resist suppression even under aggressive modeled intervention. For retrieval systems, the question is no longer whether agent users will arrive, but whether models built on human-intent assumptions will survive their presence.

cs.IR↗