Papers

Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.

Data-intensive applications, ranging from large-scale retrieval systems to advanced data pipelines, are increasingly bottlenecked by the processing of highly redundant text corpora. We present Merlin, a local-first, agnostic, high-throughput deduplication and context optimization engine designed to mitigate these inefficiencies. Utilizing a highly optimized, SIMD-friendly open-addressing flat hash set combined with xxHash3-64, Merlin performs rapid, byte-exact deduplication of text passages and data chunks. While broadly applicable to any text-processing workflow, its impact is particularly pronounced in Large Language Model (LLM) ecosystems, such as Retrieval-Augmented Generation (RAG). Our empirical evaluations demonstrate an input reduction ranging from 13.9% in low-redundancy datasets to over 71% in high-redundancy pipelines, maintaining absolute data fidelity. Furthermore, we detail the system's integration architecture via the Model Context Protocol (MCP), enabling secure, zero-network-interception deployment across major IDEs and autonomous agents. This paper outlines the core algorithmic design, performance benchmarks, and the architectural principles required to process data at sustained speeds of up to 8.7 GB/s.

Hospitals, companies, and research consortia want language models that learn from, and answer questions over, data that must stay where it is held, over links with limited bandwidth, such as the satellite links of remote clinics and ships. Averaging the sites' model parameters makes each message grow with the model and requires one shared architecture; methods that exchange predictions avoid both limits, but, to our knowledge, none bounds what compressing those predictions costs. We introduce two methods that compress every message at a small and predictable cost. Federated Probe-Logit Distillation (FPLD) trains a shared language model: each site sends compressed logits on shared unlabeled probes, and a hub distills one model from their average. Federated Conformal Retrieval-Augmented Generation (FC-RAG) answers questions: each site scores candidate answers using its own documents and sends compressed scores, and the hub returns a conformal prediction set that contains the correct answer at a guaranteed rate. Both use one codec, which centers each message and rounds it after adding a random offset known to the receiver. For logits within a fixed clipping range, we bound the expected log-loss added by compression independently of vocabulary size; the bound depends only on the bits, the range, and the number of sites, falls as more sites take part, and lets the bit budget be set before deployment. We also prove that recalibrating after compression keeps the coverage guarantee, and that FPLD's training error bounds how far coverage can slip when a threshold is reused after a model update. In experiments with GPT-2 and Qwen2.5, 8 bits per number cause no detectable loss in the trained model or in coverage. FPLD also trains a shared model across different architectures, where parameter averaging does not apply; with one shared architecture, parameter averaging remains more accurate.

Reliable detection of personally identifiable information (PII) is increasingly important across modern data-processing systems, yet the task remains difficult: PII spans are heterogeneous, locale-dependent, context-sensitive, and often embedded in noisy or semi-structured documents. We present GLiNER2-PII, a small 0.3B-parameter model adapted from GLiNER2 and designed to recognize a broad taxonomy of 42 PII entity types at character-span resolution. Training such systems, however, is constrained by the scarcity of shareable annotated data and the privacy risks associated with collecting real PII at scale. To address this challenge, we construct a multilingual synthetic corpus of 4,910 annotated texts using a constraint-driven generation pipeline that produces diverse, realistic examples across languages, domains, formats, and entity distributions. On the challenging SPY benchmark, GLiNER2-PII achieves the highest span-level F1 among five compared systems, including OpenAI Privacy Filter and three GLiNER-based detectors. We publicly release the model on Hugging Face to support further research and practical deployment of open PII detection systems.

How should hidden states generated autoregressively be collapsed into a representation that reflects a language model's internal state? Despite tokens being generated under causal masking, we find that mean pooling across their hidden states yields more semantic representations than any individual token alone. We quantify this through kernel alignment to reference spaces in language, vision, and protein domains. The improvement through mean pooling is consistent with information being distributed across generated tokens rather than localized to a single position. Furthermore, representations derived from generated tokens outperform those from prompt tokens, and alignment across generation reveals interpretable dynamics in model behavior.

Self-evolving LLMs excel in verifiable domains but struggle in open-ended tasks, where reliance on proxy LLM judges introduces capability bottlenecks and reward hacking. To overcome this, we introduce G-Zero, a verifier-free, co-evolutionary framework for autonomous self-improvement. Our core innovation is Hint-$\delta$, an intrinsic reward that quantifies the predictive shift between a Generator model's unassisted response and its response conditioned on a self-generated hint. Using this signal, a Proposer model is trained via GRPO to continuously target the Generator's blind spots by synthesizing challenging queries and informative hints. The Generator is concurrently optimized via DPO to internalize these hint-guided improvements. Theoretically, we prove a best-iterate suboptimality guarantee for an idealized standard-DPO version of G-Zero, provided that the Proposer induces sufficient exploration coverage and the data filteration keeps pseudo-label score noise low. By deriving supervision entirely from internal distributional dynamics, G-Zero bypasses the capability ceilings of external judges, providing a scalable, robust pathway for continuous LLM self-evolution across unverifiable domains.

Disagreement in annotation is a common phenomenon in the development of NLP datasets and serves as a valuable source of insight. While majority voting remains the dominant strategy for aggregating labels, recent work has explored modeling individual annotators to preserve their perspectives. However, modeling each annotator is resource-intensive and remains underexplored across various NLP tasks. We propose an agreement-based clustering technique to model the disagreement between the annotators. We conduct comprehensive experiments in 40 datasets in 18 typologically diverse languages, covering three subjective NLP tasks: sentiment analysis, emotion classification, and hate speech detection. We evaluate four aggregation approaches: majority vote, ensemble, multi-label, and multitask. The results demonstrate that agreement-based clustering can leverage the full spectrum of annotator perspectives and significantly enhance classification performance in subjective NLP tasks compared to majority voting and individual annotator modeling. Regarding the aggregation approach, the multi-label and multitask approaches are better for modeling clustered annotators than an ensemble and model majority vote.

Multimodal large language models increasingly solve vision-centric tasks by calling external tools for visual inspection, OCR, retrieval, calculation, and multi-step reasoning. Current tool-using agents usually expose the executed tool trajectory and the final answer, but they rarely specify which tool observation supports each generated claim. We call this missing claim-level dependency structure the provenance gap. The gap makes tool use hard to verify and hard to optimize, because useful evidence, redundant exploration, and unsupported reasoning are mixed in the same trajectory. We introduce TRACER, a framework for verifiable generative provenance in multimodal tool-using agents. Instead of adding citations after generation, TRACER generates each answer sentence together with a structured provenance record that identifies the supporting tool turn, evidence unit, and semantic support relation. Its relation space contains Quotation, Compression, and Inference, covering direct reuse, faithful condensation, and grounded derivation. TRACER verifies each record through schema checking, tool-turn alignment, source authenticity, and relation rationality, and then converts verified provenance into traceability constraints and provenance-derived local credit for reinforcement learning. We further construct TRACE-Bench, a benchmark for sentence-level provenance reconstruction from coarse multimodal tool trajectories. On TRACE-Bench, simply adding tools often introduces noise. With Qwen3-VL-8B, TRACER reaches 78.23% answer accuracy and 95.72% summary accuracy, outperforming the strongest closed-source tool-augmented baseline by 23.80 percentage points. Compared with tool-only supervised fine-tuning, it also reduces total test-set tool calls from 4949 to 3486. These results show that reliable multimodal tool reasoning depends on provenance-aware use of observations, not on more tool calls alone.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K–32K context lengths; on RULER, it raises CWE aggregation from 72.9% to 81.1% at 16K; and on GPQA with agentic tool use, it yields a 24% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529$\times$ and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT

Tool-integrated reasoning (TIR) enables large language models (LLMs) to enhance their capabilities by interacting with external tools, such as code interpreters (CI). Most recent studies focus on exploring various methods to equip LLMs with the ability to use tools. However, how to further boost the reasoning ability of already tool-capable LLMs at inference time remains underexplored. Improving reasoning at inference time requires no additional training and can help LLMs better leverage tools to solve problems. We observe that, during tool-capable LLM inference, both the number and the proportion of erroneous tool calls are negatively correlated with answer correctness. Moreover, erroneous tool calls are typically resolved successfully within a few subsequent turns. If not, LLMs often struggle to resolve such errors even with many additional turns. Building on the above observations, we propose PruneTIR, a rather effective yet efficient framework that enhances the tool-integrated reasoning at inference time. During LLM inference, PruneTIR prunes trajectories, resamples tool calls, and suspends tool usage through three components: Success-Triggered Pruning, Stuck-Triggered Pruning and Resampling, and Retry-Triggered Tool Suspension. These three components enable PruneTIR to mitigate the negative impact of erroneous tool calls and prevent LLMs from getting stuck in repeated failed resolution attempts, thereby improving overall LLM performance. Extensive experimental results demonstrate the effectiveness of PruneTIR, which significantly improves Pass@1 and efficiency while reducing the working context length for tool-capable LLMs.

Recent advancements in Neural Machine Translation (NMT) have significantly improved translation quality. However, the increasing size and complexity of state-of-the-art models present significant challenges for deployment on resource-limited devices. Knowledge distillation (KD) is a promising approach for compressing models, but its effectiveness diminishes when there is a large capacity gap between teacher and student models. To address this issue, we propose Evolving Knowledge Distillation (EKD), a progressive training framework in which the student model learns from a sequence of teachers with gradually increasing capacities. Experiments on IWSLT-14, WMT-17, and WMT-23 benchmarks show that EKD leads to consistent improvements at each stage. On IWSLT-14, the final student achieves a BLEU score of 34.24, narrowing the gap to the strongest teacher (34.32 BLEU) to just 0.08 BLEU. Similar trends are observed on other datasets. These results demonstrate that EKD effectively bridges the capacity gap, enabling compact models to achieve performance close to that of much larger teacher models.Code and models are available at https://github.com/agi-content-generation/EKD.

While recent self-training approaches have reduced reliance on human-labeled data for aligning LLMs, they still face critical limitations: (i) sensitivity to synthetic data quality, leading to instability and bias amplification in iterative training; (ii) ineffective optimization due to a diminishing gap between positive and negative responses over successive training iterations. In this paper, we propose Team-based self-Play with dual Adaptive Weighting (TPAW), a novel self-play algorithm designed to improve alignment in a fully self-supervised setting. TPAW adopts a team-based framework in which the current policy model both collaborates with and competes against historical checkpoints, promoting more stable and efficient optimization. To further enhance learning, we design two adaptive weighting mechanisms: (i) a response reweighting scheme that adjusts the importance of target responses, and (ii) a player weighting strategy that dynamically modulates each team member's contribution during training. Initialized from a SFT model, TPAW iteratively refines alignment without requiring additional human supervision. Experimental results demonstrate that TPAW consistently outperforms existing baselines across various base models and LLM benchmarks. Our code is publicly available at https://github.com/lab-klc/TPAW.

The implicit policy of maintaining relatively stable acceptance rates at top AI conferences, despite exponentially growing submissions, introduces a critical structural vulnerability. This position paper characterizes a new systemic threat we term Agentic Denominator Gaming, in which a malicious actor deploys AI agents to generate and submit a large volume of superficially plausible but low-quality papers. Crucially, their objective is not the acceptance of low-quality papers, but rather to inflate the submission denominator and overwhelm reviewing capacity. Under a relatively stable acceptance rate, this dilution can systematically increase the publication probability of a small, targeted set of legitimate papers. We analyze the practical feasibility of this threat and its broader consequences, including intensified reviewer burnout, degraded review quality, and the emergence of industrialized automated agent mills. Finally, we propose and evaluate a range of mitigation strategies, and argue that durable protection will require system-level policy and incentive reforms, rather than relying primarily on technical detection alone.

A vision-language model can look at a knot diagram and report what it sees, yet fail to act on that structure. KnotBench pairs an 858,318-image corpus from 1,951 prime-knot prototypes (crossing numbers 3 to 19) with a protocol whose answers are checked against Regina's canonical knot signature. Its 14 tasks span four families, equivalence judgment, move prediction, identification, and cross-modal grounding; an image-versus-symbol split locates failures along the perception-operation gap. We score Claude Opus 4.7 and GPT-5, each with and without thinking, under a 64K output-token budget matched on both vendors. Across 56 (task, model) cases, 15 sit at or below a random baseline and 8 of 14 tasks have a best score under 1.5x random. On diagram-to-symbol transcription, no model produces a strictly correct string, and permissive Regina decoding recovers the knot in 0 to 4 of 100 items. Thinking-mode reasoning lifts overall accuracy by 1.65 points for Claude and 9.25 points for GPT-5, narrowing the gap only modestly. Read together, the four families suggest current vision-language models hold features of a diagram but lack apparatus to simulate moves on those features.

Large language models (LLMs) are often evaluated based on their stated values, yet these do not reliably translate into their actions, a discrepancy termed "value-action gap." In this work, we argue that this gap persists even under explicit reasoning, revealing a deeper failure mode we call "Pseudo-Deliberation": the appearance of principled reasoning without corresponding behavioral alignment. To study this systematically, we introduce VALDI, a framework for measuring alignment between stated values and generated dialogue. VALDI includes 4,941 human-centered scenarios across five domains, three tasks that elicit value articulation, reasoning, and action, and five metrics for quantifying value adherence. Across both proprietary and open-source LLMs, we observe consistent misalignment between expressed values and downstream dialogues. To investigate intervention strategies, we propose VIVALDI, a multi-agent value auditor that intervenes at different stages of generation.

Recall presents a difficult choice: transformers have a linearly growing memory that slows each successive token, while linear RNNs typically have fixed costs but limited recall. We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. We release our code at https://github.com/featherless-ai/KVM-paper and trained models at https://huggingface.co/collections/featherless-ai/kvm-paper under the Apache 2.0 license.

Next-generation visual assistants, such as smart glasses, embodied agents, and always-on life-logging systems, must reason over an entire day or more of continuous visual experience. In ultra-long videos, relevant information is sparsely distributed across hours or days, making memory a fundamental challenge: models must accumulate information over time, recall prior states, track temporal order, and abstract recurring patterns. However, existing week-long video benchmarks are primarily designed for perception and recognition, such as moment localization or global summarization, rather than reasoning that requires integrating evidence across multiple days. To address this gap, we introduce EgoMemReason, a comprehensive benchmark for week-long egocentric video understanding through memory-driven reasoning. EgoMemReason evaluates three complementary memory types: entity memory, tracking how object states evolve and change across days; event memory, recalling and ordering activities separated by hours or days; and behavior memory, abstracting recurring patterns from sparse, repeated observations over the whole week period. EgoMemReason comprises 500 questions across three memory types and six core challenges, with an average of 5.1 video segments of evidence per question and 25.9 hours of memory backtracking. We evaluate EgoMemReason on 17 methods across MLLMs and agentic frameworks, revealing that even the best model achieves only 39.6% overall accuracy. Further analysis shows that the three memory types fail for distinct reasons and that performance degrades as evidence spans longer temporal horizons, revealing that long-horizon memory remains far from solved. We believe EgoMemReason establishes a strong foundation for evaluating and advancing long-context, memory-aware multimodal systems.

Production LLM coding agents drift over long sessions: they forget user-specified constraints, slip into mistakes the user already flagged, and confabulate prior agreements. White-box approaches such as persona vectors require model weights and so cannot be applied to closed APIs (Claude, GPT-4) that most users actually interact with. We present Nautilus Compass, a black-box persona drift detector and agent memory layer for production coding agents. The method operates entirely at the prompt-text layer: cosine similarity between user prompts and behavioral anchor texts, aggregated by a weighted top-k mean using BGE-m3 embeddings. Compass is, to our knowledge, the only public agent memory layer (among Mem0, Letta, Cognee, Zep, MemOS, smrti verified May 2026) that does not call an LLM at index time to extract facts or build a graph; raw conversation text is embedded directly. The system ships as a Claude Code plugin, an MCP 2024-11-05 A2A server (Cursor, Cline, Hermes), a CLI, and a REST API on one daemon, with a Merkle-chained audit log for tamper-evident anchor updates. On a held-out test set built from real Claude Code session traces and labeled by an independent LLM judge, Compass reaches ROC AUC 0.83 for drift detection. The embedded retrieval pipeline scores 56.6% on LongMemEval-S v0.8 and 44.4% on EverMemBench-Dynamic (n=500), topping the four published EverMemBench Table 4 baselines. LongMemEval-S 56.6% is ~30 points below recent white-box leaders (90+%); we treat that as the architectural ceiling of the no-extraction design. End-to-end reproduction cost is $3.50 (~14x cheaper than GPT-4o-judged stacks). A paired cross-vendor behavior A/B accompanies these numbers as preliminary system-level evidence. Code, anchors, frozen test data, and audit-log tooling are MIT-licensed at github.com/chunxiaoxx/nautilus-compass.

The Metacognitive Probe is an exploratory five-task, 15-slot diagnostic that decomposes an LLM's confidence behaviour into five behaviourally-distinct dimensions: confidence calibration (T1-CC), epistemic vigilance (T2-EV), knowledge boundary (T3-KB), calibration range (T4-CR), and reasoning-chain validation (T5-RCV). It is evaluated on N=8 frontier models and N=69 humans. The instrument is motivated by Flavell (1979) and Nelson and Narens (1990) but operates on observable confidence-correctness alignment; it is not a validated cross-species metacognition scale, and the pre-specified human developmental hypothesis was falsified. Composite benchmarks (MMLU, BIG-Bench, HELM, GPQA) ask whether a model produces a correct response. They are silent on whether the model knows when its response is wrong. A model can score 80 on a composite calibration benchmark and still be wildly overconfident in narrow pockets the aggregate cannot surface. The Metacognitive Probe surfaces those pockets. Our headline is a 47-point within-model dissociation in Gemini 2.5 Flash: panel-best within-task calibration (T1-CC = 88; Spearman rho = +0.551, 95% CI [+0.14, +0.80], p = 0.005) and panel-worst cross-task difficulty prediction (T4-CR = 41; sigma_conf = 1.4 across twelve factoids).

Sentiment analysis has been of long-standing interest in psychotherapy research. Recently, the Transformer deep learning architecture has produced text-based sentiment analysis models that are highly accurate and context-aware. These models have been explored as proxies for emotion measurement instruments in psychotherapy, but not investigated as stand-alone psychometric tools. Using proposed utterance-level and session-level sentiment features derived from a fine-grained sentiment model on a large corpus of psychotherapy sessions (N = 751), we investigate the distribution of session aggregated sentiment scores. Further, we characterize the relationship of these features to individual components and the overall score of the OQ-45 instrument and find that this sentiment feature is most strongly correlated to components related to emotional valence in directionally intuitive ways. Finally, we report that there are statistically significant differences between the sentiment distributions for patients flagged as at risk of deterioration or dropping out of care via either the OQ Rational or Empirical outcome models. These correlations to a fully-validated psychometric instrument demonstrate that these proposed sentiment features are, at least, adjunctive measures of client distress and deterioration.

While language models remain frozen at their training state, the world evolves continuously. Knowledge editing has emerged as a key alternative to full retraining, but its deployment is bottlenecked by the erosion of core capabilities: mathematical and programmatic reasoning collapse while encyclopedic recall remains intact. We trace this asymmetric degradation to a distributional mismatch. Covariance-based editors preserve only the subspaces spanned by their reference corpus, but fail to capture the operative distribution shaped by post-training such as SFT and DPO. Static external corpora, including Wikipedia and even the original pretraining mixture, cannot recover this shifted manifold. We propose Moir, which estimates the preservation covariance $C$ directly from the model itself by sampling from its own decoding distribution. Seeding generation with a single random vocabulary token bypasses the instruction-following templates that otherwise dominate sampled outputs, exposing the broader subspaces the model has internalized. Moir requires no external data and serves as a drop-in component for any covariance-based editor, a practical advantage given that the pre- and post-training corpora of most modern LLMs are not publicly accessible. Across OLMo-2, Llama-3.1, and Qwen-3 (7-8B), under both MEMIT and AlphaEdit and in batch and sequential regimes, Moir consistently extends preservation in the most vulnerable domains, most strikingly on Qwen3-8B after 20,000 AlphaEdit batch edits, it retains 79.9% GSM8K accuracy compared to 10.9% with the Wikipedia baseline. These results suggest that aligning the preservation distribution with the model's operative distribution is a key factor in non-destructive editing, and that the model itself may be the most accessible source of that distribution for deployed systems.

Recent advances in large language models and their widespread adoption have prompted claims that natural language could entirely replace formal languages, such as programming languages for software design. In this position paper, we argue that this perspective overlooks fundamental linguistic properties of natural language, specifically that it is optimized for underspecification in open-ended contexts. We introduce a formal framework centered on *task specificity*, defining it as the information-theoretic reduction of uncertainty in an output space – such as all possible images – given a user's specific requirements. We prove a *specificity crossover theorem*, showing the existence of a threshold beyond which the cost to express formal requirements into natural language exceeds the cost of direct formal specification. By analyzing case studies across modalities, such as image generation, code synthesis, and audio production, we demonstrate that natural language excels at low specificity tasks, while formal languages are advantageous on tasks with stricter requirements. We conclude that natural and formal languages are complementary tools and advocate the development of hybrid systems that allow users to move across the specificity spectrum.

Materials science literature analysis requires simultaneous attention to composition, processing, characterization, and property relationships, yet conventional retrieval-augmented generation pipelines struggle to reconcile heterogeneous tasks within a single retrieve-then-generate architecture. Here we present AlphaAgent, a skill-driven agent framework that decouples retrieval-based question answering from paper-level report generation through explicit skill contracts. A dedicated retrieval skill rewrites user requests into material-specific search intents, queries a curated index of more than 300,000 papers from the Journal Citation Reports Metallurgy and Metallurgical Engineering category, and reformulates queries when initial evidence is insufficient. A separate report-generation skill parses full-text PDFs to produce structured per-paper analytical reports and cross-paper summaries. In a blind evaluation on 40 materials-science questions, half of which required deep analytical reasoning, AlphaAgent substantially outperformed a baseline system matched for underlying model, document index, and retrieval scale, with the largest gains in mechanistic explanation and awareness of credibility boundaries. These results indicate that explicit task separation, refined retrieval intent, and evidence-aware generation improve large-language-model-based literature analysis for materials research.

We present LLM-INSTRUCT, the winning system for the UZH Shared Task at ArgMining 2026 on paragraph-level argument mining in UN and UNESCO resolutions. The task requires paragraph-type classification, prediction of a subset of 141 official tags, and directed relation prediction under a strict JSON schema setting using only open-weight models up to 8B parameters. We frame the task as constrained structured prediction. The system first narrows the candidate tag space with metadata-aware dense retrieval, then applies constrained decoding with per-dimension caps, escalates only uncertain cases to a three-agent debate branch, and finally validates the output schema. On the official leaderboard, LLM-INSTRUCT ranked 1st overall, with 1st in F1 and 5th in LLM-as-a-Judge. During development, our configuration search further improved Task 1b Micro-F1 from 35.83% to 40.08% while keeping the internal Task 2 score at 4.421. The main lesson is simple: reducing the decision space before generation improves both accuracy and submission robustness. Our code and supporting scripts are publicly available at: https://github.com/LLM-Instruct-at-UZH-Shared-Task-2026/Method

Large language models are increasingly used to simulate diverse human opinions in open-ended tasks such as synthetic surveys, focus group modeling, and public opinion prediction. However, LLM outputs exhibit systematic opinion homogenization. Practitioners have explored various interventions to increase diversity, but the landscape remains fragmented: different methods are evaluated in isolation with incomparable metrics, and in practice they are typically deployed and upgraded simultaneously, making it difficult to attribute gains to specific components. To advance a more scientific understanding of LLM output diversity, we design a factorial experiment that separates two primary intervention dimensions: input conditioning (operationalized through persona depth) and interaction architecture. We evaluate all conditions on 100 real-user open-ended questions across 7 models, measuring diversity with multiple complementary metrics. Our findings challenge several common assumptions. First, more persona detail does not monotonically increase diversity. The initial step of persona conditioning already captures the majority of the gain, while further elaboration with demographic detail does not consistently improve and can reduce diversity on some models. Second, rather than seeking a single best interaction architecture, we find that different architectures explore largely non-overlapping opinion regions. Combining multiple architectures yields broader coverage than optimizing any one. Third, commonly attempted low-cost alternatives such as raising sampling temperature and adding diversity instructions produce negligible effects compared to structured interventions. Overall, our work demonstrates that diversity is not a product of scaling along any single dimension, but is highly sensitive to the structural form and combination of interventions.

Artificial agents now generate behavior rich enough to invite trust, surprise, and concern, yet our evaluation tools still privilege capability scores over psychological structure. This paper argues that the philosophical impasse between two symmetrical errors (Artificial Mind Blindness, which dismisses psychological organization in non-biological systems, and Artificial Mind Projection, which infers human-like inner life from fluent behavior alone) can be circumvented not by resolving the consciousness question, but by introducing a disciplined measurement layer beneath it. Drawing on Michael Levin's continuum view of cognition as goal-directed competency across substrates, and on the methodological repertoire of mathematical psychology (Item Response Theory, Signal Detection Theory, Bayesian cognitive modeling, calibration analysis, cognitive-bias batteries), the paper develops Machine Psychometrics as a measurement science of latent behavioral, metacognitive, communicative, and self-modeling dispositions in artificial agents. Its operational core is the Machine Mindprint: a multidimensional, domain-bounded, versioned profile spanning calibration, source integrity, suggestibility resistance, context stability, expressive alignment, tool integrity, drift monitoring, and distributional grounding. A complementary Trust Protocol turns Mindprints into deployment decisions through probe batteries, perturbation testing, reliability and validity analysis, and longitudinal monitoring across high-stakes domains. The philosophical contribution is a third stance, Artificial Mind Discipline, that neither anthropomorphizes nor dismisses, neither presupposes consciousness nor forecloses it. The aim is not to humanize artificial agents, but to understand them precisely because they are not human, through measurement before judgment.

Defenses against indirect prompt injection (IPI) in tool-using LLM agents share two structural weaknesses. First, they all attempt to prevent attacks rather than detect the compromises that slip through. Second, they have only been evaluated in English, leaving users of low-resource languages such as Kurdish and Arabic without tested protection. This paper addresses both gaps with AgentShield, a deception-based detection framework that places three layers of traps inside the agent's tool interface: fake tools, fake credentials, and allowlisted parameters. The same trap triggers serve as high-precision labels for a self-supervised classifier. An LLM agent that follows an attacker's hidden instruction almost always touches one of these traps, which gives both a real-time compromise signal and a zero-FP label for training a downstream detector without manual annotation. Across 176 cross-lingual attack prompts and four LLMs from three providers, and because modern LLMs already refuse most IPI attempts on their own (attack success rate <= 10%), AgentShield's job is to catch the attacks that do slip through. On commercial models, it catches 90.7%-100% of such successful attacks, with zero false alarms on 485 normal-use tests. It survives a systematic adaptive-attack evaluation with zero evasion on commercial models, and the self-supervised classifier transfers across models and languages without retraining.

User simulators are increasingly leveraged to build interactive AI assistants, yet how to measure the quality of these simulators remains an open question. In this work, we show how simulator quality can be quantified in terms of its downstream utility: how an LLM assistant trained with this user simulator performs in the wild when interacting with real humans. In a controlled experiment where only the user simulator varies, we train LLM assistants via reinforcement learning against a spectrum of simulators, from an LLM prompted to role-play a user to one fine-tuned on human utterances from WildChat. As evaluation, we measure pairwise win rates in a user study with 283 participants and on WildBench, a benchmark derived from real human–AI conversations. Training against the role-playing LLM yields an assistant statistically indistinguishable from the initial assistant in our user study (51% win rate), whereas training against the fine-tuned simulator yields significant gains (58% over the initial and 57% over the one trained against role-playing). Closer inspection reveals three further patterns: methods for making role-playing LLMs more realistic (e.g., persona conditioning) improve trained assistants but do not close the gap to the fine-tuned simulator; scaling the simulator's model size benefits the fine-tuned simulator but yields no gain for role-playing ones; and assistants trained against role-playing simulators fail to generalize when paired with other simulators at test time, while the one trained against fine-tuned simulator does. Together, these results argue for grounding user simulators in real human behavior and measuring their quality by their downstream effect on real users.

This paper presents our systems and results for the Hope Speech Detection in Code-Mixed Tulu Language shared task at the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages (DravidianLangTech-2026). We trained an XLM-RoBERTa-based text classification system for detecting hope speech in code-mixed Tulu social media comments. We compared this organically adapted hope speech detection model with our baseline model. On the development set, the organically adapted model outperformed the baseline system. While our submitted systems performed more modestly on the official test set, these results suggest that further adapting XLM-RoBERTa on organically collected Tulu social media text containing code-mixed and mixed-script variation can improve hope speech detection in code-mixed Tulu.

Large Language Models exhibit mode collapse, producing homogeneous outputs that fail to explore valid solution spaces. We present QD-LLM, a framework for parameter-efficient neuroevolution that evolves prompt embeddings, compact neural interfaces (~32K parameters) that steer generation in frozen LLMs (70B+ parameters), within a Quality-Diversity (QD) optimization framework. Our contributions: (1) evolved prompt embeddings via gradient-free optimization enabling behavioral steering without model fine-tuning; (2) hybrid behavior characterization combining semantic and explicit features with formal coverage bounds (Theorem 1) under validated near-independence (NMI $= 0.08 \pm 0.02$); (3) co-evolutionary variation operators including targeted behavioral mutation via finite-difference gradient estimation. On HumanEval (164 problems), MBPP, and creative writing benchmarks, QD-LLM achieves 46.4% higher coverage and 41.4% higher QD-Score than QDAIF ($p<0.001$, 30 runs, Vargha-Delaney $A=0.94$). We demonstrate downstream utility: diverse archives improve test generation (34% more edge cases) and fine-tuning data quality (8.3% accuracy gain). We validate across open-source LLMs (Llama-3-70B, Mistral-Large) with full embedding access, establishing prompt embedding evolution as an effective paradigm bridging neuroevolution and modern LLMs.