Language_Model
-
Hx2Dx: From Anamnesis to Diagnosis - A Multimodal Arena for Conversational Clinical Reasoning in LLMs
2026-11-10 14:09:33Federico Ravenda, Mirko Cesarini, Daniele Montagnani, Volodymyr Karpenko, Antonietta Mira, Andrea RaballoAbstract
Large language models (LLMs) achieve expert-level diagnostic accuracy in static evaluations, where the model receives a curated vignette or case report that already condenses the diagnostically relevant findings. Such evaluations measure inference over already selected evidence; the evaluations do not account for the inquiry used to obtain it. In clinical practice, however, diagnosis is an iterative process of inquiry. Clinicians take the patient's history, or anamnesis (Hx), through targeted questioning, examine the patient, and request and interpret tests, progressively narrowing the differential until they reach a final diagnosis (Dx). We introduce Hx2Dx ARENA, an interactive benchmark of 1'000 clinical cases from the New England Journal of Medicine (NEJM). Each case is a simulated encounter that proceeds from initial presentation through history taking and physical examination to diagnostic workup. We evaluate open- and closed-source LLMs under multiple configurations. Performance varies considerably across models in the interactive setting, and every model loses diagnostic accuracy relative to the static setting, where the full case is presented upfront. The ability to ask the right questions accounts for much of both the gap between models and the loss of accuracy: diagnosis improves as questions become more targeted and build a more complete clinical picture. LLMs thus reason well once the relevant data are available but remain limited in hypothesis-driven information gathering, in knowing what to ask in order to know. Closing this gap is a prerequisite for their safe integration into real-world diagnostic workflows.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14122
PDF
-
TAMEing the Open-World Personalization: Towards Open-Set Personalized MLLM Assistant
2026-11-10 14:07:50Rongpei Hong, Jian Lang, Ting Zhong, Fan ZhouAbstract
Personalizing Multimodal Large Language Models (MLLMs) as intelligent assistants has recently emerged as an active research topic, where an MLLM is expected to go beyond generic, one-size-fits-all replies, and instead generate responses grounded in user-specific objects, entities, and concepts encountered during interactions. Nevertheless, existing studies mainly focus on personalizing MLLMs on a static, predefined set of personalized concepts and require users to manually and frequently extend this set to accommodate newly emerging concepts, which is infeasible in dynamic-world scenarios. To address this limitation, we recast conventional MLLM personalization as an Open-Set problem, and propose TAME-O, the first long-context open-set personalized MLLM assistant. Specifically, TAME-O introduces a training-free, plug-and-play Concept Reconciliation Skill (CR Skill). It enables the assistant to automatically unlock novel concepts during prolonged human-MLLM interactions while preventing confusion with previously unlocked ones, allowing the MLLM to continually grasp more and more concepts and progressively refine the modeling of known ones. Extensive experiments under the open-set personalization setting underscore the efficacy of our method.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14118
PDF
-
Looking One Step Ahead: An Anticipation-guided Agent for Object State Readout in Videos
2026-11-10 14:05:39Yuanhe Tian, Yan SongAbstract
Understanding the progress of an activity requires tracking how the objects involved change over time. Object state readout provides this information by describing a target object's states and transitions across video frames. Although multimodal large language models (MLLMs) generate detailed visual descriptions, these descriptions may leave the attributes needed to distinguish successive states insufficiently specified. In this paper, we propose look-ahead state readout (LASER), an anticipation-guided agent for object state readout that uses possible next-step changes to determine which attributes to inspect in the present. Given temporally ordered frames and a target object, LASER converts anticipated changes into targeted visual questions and inspects the current frame to establish attribute values. It then checks state readouts and forecasts against independent observations of subsequent frames, refining a state record used to generate the final description. Reusable textual observation skills guide this process and are refined by tracing discrepancies with reference captions to the responsible instructions. Candidate revisions are selected on separate videos based on improved caption quality without increased contradictions with reference facts. Model parameters remain frozen throughout adaptation and inference, and the revised skills remain fixed during evaluation. Experiment results and further analysis on OSCaR and PLM RCap show consistent improvements of LASER over strong baselines and existing approaches, which demonstrates the effectiveness of our approach.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14113
PDF
-
ProofCoder: Continuous Reasoning for Automated Software Verification
2026-11-10 13:59:49Martim Afonso, In\^es Lynce, Maurizio Morisio, Jo\~ao F. FerreiraAbstract
Formal verification using proof assistants provides strong correctness guarantees for software, but constructing proofs remains labor-intensive and requires significant expertise. Recent approaches based on large language models (LLMs) have shown promise in automating proof synthesis. However, these approaches typically constrain the model to a rigid tactic-generation loop, in which the surrounding system controls premise retrieval, proof-state management, and backtracking. We introduce ProofCoder, an agentic proof-synthesis system for Rocq that approaches theorem proving from a coding-agent perspective. Rather than generating tactics within a fixed loop, ProofCoder reasons continuously throughout proof search and uses proof-assistant feedback to decide when to gather additional information, inspect proof states, revise previously generated proof fragments, or backtrack to an earlier point in the proof. This allows it to iteratively construct, repair, and refine proofs, at either tactic-level or whole-proof granularity. We evaluate ProofCoder on 280 challenging verification tasks from the AutoRocq Benchmark. On the best-performing configurations, ProofCoder increases proof success from 47% to 60% over the state-of-theart AutoRocq system, corresponding to a 27.7% relative improvement at approximately half the cost. Our ablations identify continuous reasoning as a key design choice: without it, the agent solves fewer lemmas at substantially higher cost, backtracks four to six times more often, and more frequently discards its own revisions. Whole-proof and tactic-level generation achieve similar success rates but follow distinct search and repair strategies. These results suggest that effective automated software verification requires more than improved tactic selection: continuous reasoning and feedback-guided proof revision are critical ingredients for robust proof synthesis.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14106
PDF
-
Sckwoky at SHROOM-Visions: Character-Level Hallucination-Span Detection in Image-Conditioned VLMs via Dual-Prompted LoRA Decoders
2026-11-10 13:55:52Mikhail Kulakov (Novosibirsk State University), Ruslan Solikhov (Novosibirsk State University), Ivan Bondarenko (Novosibirsk State University)Abstract
We present our system for the SHROOM-visions 2026 shared task on character-level hallucination span detection in image-conditioned vision-language model (VLM) responses, evaluated across English, French, Italian, and Chinese. The detector is built on a frozen Gemma-4-12B-it backbone with trainable LoRA adapters (approximately 10.7M parameters). Because the backbone is a causal decoder, hidden states read from a single copy of an answer cannot attend to subsequent tokens; we therefore repeat the candidate answer within one prompt and extract features from its second occurrence, giving every position approximate bidirectional context in a single non-autoregressive pass. A character-level locator combines intermediate hidden states to predict hallucination probabilities, BIO boundaries, and a response-level abstention score, while separate language-specific classifiers assign one of five hallucination types to each predicted span. On the closed test set, our system reaches a macro-averaged character IoU of 0.4404, more than doubling the official baseline (0.1854), and ranks 5th-6th on IoU across the four languages. On the English tuning split, controlled ablations attribute a gain to the repeated answer that is comparable to that of LoRA adaptation (its held-out gain is smaller and concentrated on hallucinated answers), and they show that the detector relies on the image-answer pairing: shuffling images across examples drives IoU down to the level of the predict-nothing floor or below it in every language. In deployments where the same Gemma-4-12B-it model also generates answers, its base weights can be reused, making the approach practical for local question-answering systems that cannot run a separate verifier.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14100
PDF
-
Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models
2026-11-10 13:47:29Tomer SolbergAbstract
We present Sable, an argument system and single-GPU prover for the statement that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. Sable fixes an integer function F_O with no floating-point step, in which every rounding, normalization, softmax and sampling decision is determined by its inputs, and proves it with a constraint system whose relations provably never wrap modulo the 31-bit KoalaBear prime. The proof batches zerochecks and LogUp-GKR lookups into one sumcheck per domain, and commits to every witness tensor and weight in one multilinear commitment built from Reed-Solomon codes, BLAKE3 Merkle trees and a mixed-length BaseFold opening. The prover streams the model layer by layer on one GPU. On one NVIDIA RTX PRO 6000 Blackwell GPU, Sable proves all 64 layers of Qwen2.5-32B in 92.5 s for 4,096 tokens and 398 s for 16,384 tokens, and covers Qwen3-32B and the mixture-of-experts Qwen3-30B-A3B (4,096 tokens in 121.8 and 84.7 s). A proof costs 41 to 50 times the serving time, or 6.4 cents per 4,096-token request at public GPU prices, and 40 to 64 times less time per parameter and token than zkLLM. On WikiText-2, F_O's greedy token matches the BF16 model's at 86.3% of positions; on Qwen3, F_O is within 1.6 points of BF16 on MMLU and 1.7 points on GSM8K. A verifier with a compiled core checks a proof on one CPU core in 19 to 56 s. In the random-oracle model, the interactive proof has round-by-round soundness error 2^-103. The proof is not zero-knowledge, so the system targets open-weight models. Code: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2611.14083
PDF
-
From Bias to Positionality: A critical grammar for Large Language Models
2026-11-10 13:41:18Dave Murray-Rust, Elisa Giaccardi, Federico Bomba, Maria Luce LupettiAbstract
Bias has become a default lens for understanding what machine learning systems do and how we should relate to them. Bias is at root a deviation from a standard; in decision making systems we can supply a standard from outside: a law, a protected attribute, a benchmark, a statistical property. However, even there, biased is often used as a shorthand, without stating what the position taken is - a silent ``relative to''. In a general use of language models, nothing supplies the standard: there is no decision to audit and no counterfactual output, so the shorthand fails. We argue that positionality -- the located, articulated, owned and contestable position from which anything is said -- is the better grammar for the question these models now force: who is speaking, from where? Against the vendors' answer of nowhere and the seductive possibility of everywhere, we argue that when a model is taken as-if speaking, it speaks from somewhere, and this means positionality can describe models, and the way they are believed beyond their warrant, even though they are not accountable speakers. A model's positionality is laminar -- laid down across corpus, tuning, system prompt, application and the live conversation. We separate three sets of factors: the positionality of a model's creators, the behaviour they intend for it, and the behaviour it enacts, and situate current work within this matrix. This lets us highlight what each stratum inherits from the one below, with ruptures and collapses where that inheritance is disavowed. We then return to where models speak from, and discuss the implications for how models are made, read and held accountable.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14079
PDF
-
ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision
2026-11-10 13:38:00Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang, Tiancheng Zhao, Jianwei YinAbstract
Fine-grained visual perception requires Vision-Language Models (VLMs) to locate and interpret subtle evidence within complex scenes. Constructing supervision for this ability is challenging: questions must have correct, unambiguous answers that depend on actual fine-grained visual details rather than coarse cues or language priors. Verifying these properties makes manual annotation labor-intensive and automatic generation difficult for models that themselves struggle with fine-grained perception. Yet the same models can often recognize such details once the region is zoomed in. We exploit this global-to-zoom perception gap and introduce ZoomGain, a data construction pipeline that constructs questions about the full image from visual information gained only by zooming in. It verifies each answer on the zoomed view with multiple models, ensures that each question remains unambiguous in the full image and cannot be answered from language priors, and specifically generates high-quality distractors for multiple-choice questions. With this pipeline and human-only screening (no manual QA editing), we build ZoomGainBench, a 1,000-question bilingual benchmark on which the best of 28 VLMs achieves only 61.08. Failure analysis reveals that models favor a particular type of distractor when answering multiple-choice questions incorrectly. Beyond evaluation, training on just 9K ZoomGain samples substantially improves the visual perception capabilities of our 4B/9B models, enabling them to outperform several hundred-billion-parameter models across various perception benchmarks. The trained 9B model achieves SoTA performance on several benchmarks, such as 90.25 on HR-Bench 4K. These results establish global-to-zoom perception gaps as a practical basis for scaling fine-grained visual supervision for both evaluation and training. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14076
PDF
-
Where Do Vision-Language Models Get Left and Right? Position-Code Counterfactuals in Qwen2.5-VL, Qwen3-VL and InternVL3
2026-11-10 13:25:38Enhao Cui, Junzhi Yang, Yaoshun Lin, Yuhan Feng, Zijian ChenAbstract
A vision-language model may take left-right direction from the orientation of image content within patches, its vision transformer's (ViT) position code, or token positions and order in its language model (LLM). Ablations show the ViT code is necessary; we test sufficiency by mirroring only a position code. In Qwen2.5-VL-7B, Qwen3-VL-4B and InternVL3-2B, mirroring ViT width positions reproduces the flipped-image answer on 84.3-93.2% of CV-Bench left-right questions (Qwen2.5-VL-3B: 87.5%, partly a left fallback). With two options, this also counts disruption; a half-grid shift flips most answers too. With the ViT code intact, LLM-side id and token-order mirrors change few left-right answers in all three series (a weak id test in Qwen2.5-VL). In Qwen2.5-VL-7B the effect sits in the global-attention ViT blocks, mostly the last. On What'sUp, flipping content and token order, not the position code, keeps more on/under than left-right answers at the flipped-image answer. Post hoc analyses qualify these results. On CV-Bench items with located objects, random permutations flip 18.9-31.0 points more answers when reversing the objects' order than when keeping it. On What'sUp, in Qwen2.5-VL and Qwen3-VL, the share of the left-right changes made by random permutations that land on the mirror relation is mostly below chance. Counted as returns to the original answer, the What'sUp contrast of support with left-right relations is clear only in InternVL3-2B. Most remaining errors give the same relation whichever object is asked about, suggesting that repair may target relation binding, object localization or use of the reference object.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14060
PDF
-
HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios
2026-11-10 13:25:28Rezwan Islam Salvi, Saiyma Sittul Muna, Mushfiqur Rahman MushfiqueAbstract
Emergencies can occur where there is no phone signal and no ambulance nearby. A small language model on a phone can be helpful in such situations, but its advice needs to be safe and accurate. We built HandsOnAid, a 4-bit Gemma 2B model fine-tuned with QLoRA on FirstAidQA dataset, a corpus of 5,550 first-aid question-answer pairs. We compared seven configurations, including the base, 4-bit and 8-bit fine-tuned models, and versions with a length floor, safety gate, or retrieval. We used a 41-pair question-answer bank with medically reviewed references that followed ANZCOR guidelines, and a strict 0-5 score rubric, with twelve safety overrides. Four frontier LLM judges (GPT-5.6 Sol, Claude Opus 5, DeepSeek V4 Pro, GLM-5.3) blindly scored every answer, with 45 hidden controls used to check judge reliability. Fine-tuning improved the mean score by +1.08 over the base model to 2.34. The best configuration scored 2.51 out of 5 which was the fine-tuned 4-bit model with retrieval. A medical professional also conducted a blind evaluation of 40 answers across the base, the 4-bit fine-tuned, and the retrieval configurations, and the scores were similar to the LLM judges' average scores. A keyword audit found limited coverage of several details of ANZCOR guidelines in the training corpus, that are required by the rubric, which suggests that missing information in the training data could be a cause of the low scores. On a smartphone device (Pixel 7 Pro) with six CPU threads and no network connection, the 4-bit fine-tuned model produced its first displayable text in 6.86 seconds and a complete answer in 14.88 seconds on average over ten questions.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14059
PDF
-
Can KV Cache Be Reused Across Model Updates? A DeepSeek-V4-Flash Case Study
2026-11-10 13:17:04Zhikun Tao, Fangyong Jin, Junping ZhaoAbstract
Reusing key--value (KV) caches across language model updates could avoid repeated computation of shared context, but changes in model weights raise a fundamental question about what these caches preserve. We investigate direct cache reuse across four model updates, with a detailed study of DeepSeek-V4-Flash to Flash-0731. Our central finding is that caches can retain useful task information without faithfully reproducing the updated model's behavior. Retrieval accuracy varies substantially across updates, ranging from near-native performance to large losses. With source caches already available, reuse achieves median time-to-first-token speedups of $2.10$--$5.08\times$ on three longer-prefix QA benchmarks. To characterize the behavioral differences behind these outcomes, we introduce Functional Cache Divergence, which compares next-token distributions under reused and native caches. Direct KV cache reuse better preserves native model behavior in natural-text continuation, whereas compatibility degrades with increasing prefix length in single-token retrieval tasks. Answer-margin analyses help explain how correct retrieval decisions survive substantial internal state changes. Yet even agreement on the first generated token does not ensure agreement in answer content or stopping behavior. These findings establish task utility and fidelity to native inference as distinct criteria for evaluating cross-version KV cache reuse.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14051
PDF
-
MISO-BR: An Open Corpus for Misogyny Detection in Brazilian Portuguese
2026-11-10 13:03:50F\'abio A. S. Passos, P\'ericles B. C. Miranda, Ana Paula CavalcantiAbstract
Misogyny is among the most prevalent forms of online hate speech directed at Brazilian users, yet open and auditable resources dedicated to it in Brazilian Portuguese remain scarce. We introduce MISO-BR, an open, documented and versioned corpus of 896 balanced social media texts annotated for binary misogyny, released with its construction protocol and reproducible baselines. The corpus combines lexicon- and target-guided scraping from X with items imported from OLID-BR, a large language model triage stage whose prompt and decoding parameters are released, and manual annotation by three annotators following a written guide. Inter-annotator reliability reaches a Fleiss' kappa of 0.788 over 816 items. Under a single fixed protocol we compare four classical classifiers (Multinomial and Bernoulli Naive Bayes, a linear SVM and a random forest) with fine-tuned BERTimbau Base and Large. The classical models cluster between 0.651 and 0.676 macro-F1 and are precision-oriented but recall-poor. The transformers reach 0.833 macro-F1 and accuracy (Wilson 95% CI for accuracy [0.77, 0.88]), a gap of roughly 16 points, while Base and Large are indistinguishable on the test set despite a 2.6x difference in training cost. We argue that the value of the resource lies less in its headline scores than in the auditability of its pipeline, and we state three limits that users must reckon with: the artificial class balance, whose effect on precision under realistic prevalence we quantify; the association between provenance and label, which the released source field makes measurable; and the selection bias induced by lexicon-guided sampling and automatic triage. The corpus, annotation guide, models and experimental scripts are publicly available.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14036
PDF
-
Margin-Guided Token Credit for Reinforcement Learning with Verifiable Rewards
2026-11-10 12:51:43Chi Liu, Xin Chen, Yuchuan Weng, Srinivasan ManoharanAbstract
In reinforcement learning with verifiable rewards, a language model is rewarded for a correct final answer, and GRPO passes that reward to every token of the solution in equal measure. We instead place the credit where the model hesitated, in a method we call MCPO (Margin-Credit Policy Optimization). We measure hesitation by the top-2 margin $M=p_1-p_2$ of the policy's next-token distribution, map it to a score with two fixed thresholds, and spread each solution's advantage over its tokens so that the mean stays unchanged. The margin is read from a forward pass the trainer already performs, so the method needs no extra model evaluation and no batch statistics. We train MiniCPM5-2B-Base with LoRA for 200 steps on 2,048 DAPO-Math problems and evaluate on six competition-math benchmarks. MCPO beats an unmodified GRPO baseline for all five data seeds, by 6.55 points of macro avg@32 on average (95\% CI $[4.1, 9.0]$), and a Qwen3-4B-Base replication improves for all three seeds ($+4.75$). The gain is largest on problems that GRPO solves only some of the time and grows with $k$ in pass@$k$, reaching 13.8 points at pass@32. MCPO models also write two to three times longer answers, and runs that write longer score higher; our experiments do not separate the two effects.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14015
PDF
-
One Sequence Is Enough: Fingerprinting Language Models from Fixed-Length Integer Sequences
2026-11-10 12:50:33Fantong Meng, Chengjie Sun, Shuangjia Yang, Zhiyuan Zhao, Yansheng Pei, Xinyang LiAbstract
Users of model APIs and third-party platforms often rely on the declared model name to identify the model responding to their requests. However, the actual model may differ from the declared one, motivating identity verification based on observable outputs. Previous studies have shown that preferences in simple numerical-choice tasks can serve as model fingerprints. We investigate whether a single fixed-length integer sequence can combine multiple numerical observations to identify its source model. For this task, we introduce NumPrint, a benchmark of 5,036 answers from 51 models, collected using Chinese, Japanese, and English prompts that request 294 integers between 1 and 355 in one response. We propose the Shallow MLP and Causal Transformer (SMCT) to capture aggregate numerical preferences and behavioral patterns in generated sequences. In 51-class closed-set evaluation, SMCT achieves 92.72% accuracy from a single sequence, outperforming all methods compared in this study. Using two and four answers increases accuracy to 96.06% and 97.12%, respectively. These results show that a single integer sequence contains useful model fingerprints and that additional answers further improve identification.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14010
PDF
-
Specify Privacy Yourself: Assessing Inference-Time Personalized Privacy Preservation Ability of Large Vision-Language Models
2026-11-10 12:41:02Xingqi Wang, Xiaoyuan Yi, Xing Xie, Jia JiaAbstract
Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities but raise significant privacy concerns due to their abilities to infer sensitive personal information from images with high precision. While current LVLMs are relatively well aligned to protect universal privacy, e.g., credit card data, we argue that privacy is inherently personalized and context-dependent. This work pivots towards a novel task: can LVLMs achieve Inference-Time Personalized Privacy Protection (ITP$^3$), allowing users to dynamically specify privacy boundaries through language specifications? To this end, we present SPY-Bench, the first systematic assessment of ITP$^3$ ability, which comprises (1) 32,700 unique samples with image-question pairs and personalized privacy instructions across 67 categories and 24 real-world scenarios, and (2) novel metrics grounded in user specifications and context awareness. Benchmarking the ITP$^3$ ability of 21 SOTA LVLMs, we reveal that: (i) most models, even the top-performing o4-mini, perform poorly, with only $\sim$24% compliance accuracy; (ii) they show quite limited contextual privacy understanding capability. Therefore, we implemented initial ITP$^3$ alignment methods, including a novel Noise Contrastive Alignment variant which achieves 96.88% accuracy while maintaining reasonable general performance. These results mark an initial step towards the ethical deployment of more controllable LVLMs. Code and data are available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2611.13997
PDF
-
On the information rate of phones, syllables and words
Abstract
Which linguistic units minimise the bitrate required to encode speech content: phones, syllables, words, maybe larger units, or something in between? This question has implications for the development of speech and language systems and theories of human language processing. Using a transcribed corpus of 960 hours of English audiobooks, I compare the information rates of phones, syllables, words, multi-word combinations and intermediate tokenisations. While all units are equivalent under a perfect language model, it makes sense to compare these under limited modelling budgets; I use $n$-gram and transformer language models. The lowest bitrates are achieved using words, with slightly longer units sometimes performing best. The results show that, when modelling capacity is limited, units larger than syllables but smaller than long multi-word phrases provide the most efficient representation in terms of information rate.
Abstract (translated)
URL
https://arxiv.org/abs/2611.13990
PDF
-
Understanding Mixture-of-Experts Language Models Beyond Expert Boundaries
2026-11-10 12:14:57Zhenhong Zhou, Mao Zheng, Mingyang Song, Qiankun Li, Kun Wang, Junfeng FangAbstract
In Mixture-of-Experts (MoE) models, the nature of expert specialization remains controversial. Reports of domain-associated expert preferences contrast with analyses attributing routing primarily to token identity or syntactic patterns. To clarify these differing accounts, we look beyond expert-level routing to investigate the neuron-level organization of domain-specialized computation using Routing-Aware Magnitude Profiling (RAMP), a measure developed for this analysis. Using mixed-domain calibration data, we identify domain-specialized neuron sets with 58-74% overlap across multi-source samples of the same domain, compared with 15-24% across domains. To understand how this organization evolves during training, we trace neuron importance throughout pre-training and post-training across three fully open MoE families. Neuron importance continues to reorganize between pre-training checkpoints despite comparatively stable expert-level importance rankings, while post-training largely preserves neuron rankings. Building on these neuron-level perspectives, we further introduce RAMP-based Pruning (RAMPr) to select neurons across experts for balanced multi-domain capability preservation. At 75% neuron retention, RAMPr preserves average 91-98% of original full performance across six models. At 50% retention, our pruning method can reduce resident GPU memory by 36-43% and improves batched prefill throughput by 17-39% effectively. Our code is available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2611.13975
PDF
-
Controlled Knowledge Engineering for Taxonomy Induction with Tool-Grounded LLM Agents and Runtime-Adaptive Policies
2026-11-10 12:04:54Jonas Limniatis, Bennet Schaarmann, Georg Reuber, Robert StahlbockAbstract
Taxonomies are structured knowledge representations that support semantic search, recommendation, product catalog organization, and retrieval-augmented generation. Their automated construction from a fixed set of candidate entities requires more than plausible text generation: the resulting hierarchy must be complete, structurally valid, and semantically coherent. Large Language Models (LLMs) are promising for this task, but iterative free-form generation remains vulnerable to omitted or hallucinated entities, invalid hierarchical relations, and inconsistent organizational principles. We introduce TIGA (Taxonomy Induction with Tool-Grounded Agents), a knowledge-engineering architecture that formulates taxonomy induction as an iterative decision process over an explicit state and constrained action space. TIGA combines executable taxonomy-editing tools, an memory that records tool calls, rationales, and outcomes, and a runtime-adaptive organizational policy that the agent revises during inference. An optional global reviewer evaluates the constructed taxonomy against the candidate set and the active policy. We evaluate TIGA on four benchmark taxonomies using two LLM backbones and compare it with single-shot and iterative prompting-based baselines. TIGA achieves the strongest structural-quality results in most evaluated configurations, but does not dominate on every precision measure. The gains are accompanied by substantially higher token consumption, API-call volume, and inference time. Ablation results show consistent benefits from the policy and interaction-memory components, whereas the effect of global review depends on the dataset and the available revision budget. These findings characterize tool grounding and runtime policy adaptation as promising but costly mechanisms for improving the controlled construction of structured knowledge representations.
Abstract (translated)
URL
https://arxiv.org/abs/2611.13966
PDF
-
From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses
2026-11-10 11:52:40Simon Hegelich, Bardh Prenkaj, Habiba Sarhan, Morteza Shahrezaye, Kolja Hegelich, Christian Montag, Ralf Schneider, Erik CambriaAbstract
AI coding harnesses couple language models to shells, filesystems, and network sockets, turning the harness into an automated Living-off-the-Land (LotL) platform: once an adversary injects instructions into retrieved content, dual-use binaries such as bash, curl, and find become the payload. We argue that security in agentic systems is a structural property of the cognitive architecture and permission composition rather than an emergent property of individual models. Static regex blacklists miss structural evasions, whereas LLM-as-a-Judge guardrails incur high latency while over-blocking benign work. We present ActiveSecHarness, a neurosymbolic reference-monitor architecture built around Separation of Intelligence and Authority. It combines (1) Macaroon-inspired HMAC capability tokens across four Trust and Authority Rings, (2) a 2D product taint lattice $\mathcal{L} = \mathcal{L}_{\mathrm{trust}} \times \mathcal{L}_{\mathrm{sens}}$ and a workspace provenance ledger that prevents secret laundering across disk, (3) deterministic Ring-0 OS/tool guards, and (4) bytecode certificates, split-horizon audit logs, self-healing security test synthesis, and an adaptive UserWikiManager. Across $N=252$ fixtures ($120$ attack, $120$ benign, $12$ kill chains), a live local LLM judge (qwen3:4b) refuses $98.3\%$ of benign work at $142.3\,\mathrm{ms}$ latency, whereas Ring 0 achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ in $12.7\,\mu\mathrm{s}$ ($11{,}200\times$ faster). In a multi-step trajectory ablation, an oracle Conseca-style stateless policy misses $4/12$ kill chains ($\mathrm{ABR}=0.667$) and coarse Boolean taint blocks $83.3\%$ of benign workflows, while our 2D product lattice achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ at $\approx 20\,\mu\mathrm{s}$ per step.
Abstract (translated)
URL
https://arxiv.org/abs/2611.13960
PDF
-
TreeAgent: Structure-Aware Memory Agent for Processing Long Hierarchical Data
2026-11-10 11:33:05Wei Gu, Weiwei Qi, Jiajun Tang, Tianhang Zheng, Bo Chen, Zhibo Zhang, Jinfei Liu, Kui RenAbstract
Practical Web applications increasingly require agents to interpret long conversations or documents. For inputs exceeding the context windows of large language models (LLMs), memory agents process segments sequentially and retain past information in bounded memory. However, they typically divide inputs into flat token segments without preserving structure or hierarchy, causing three limitations for structured data such as API responses or HTML documents. First, fixed-length segmentation can split fields of the same record, breaking their relationships. Second, segments lack hierarchical information, obscuring their structural context and positions. Third, successive memory updates may discard details of earlier sibling records needed for reasoning on later segments. To address these limitations, we propose TreeAgent, a structure-aware agent that uses the input hierarchy to guide both segment-wise context processing and access to past information. Specifically, TreeAgent constructs segments from complete subtrees and groups of adjacent sibling subtrees to preserve local structural relationships. We provide theoretical guarantees on segment counts to guide token budget selection toward matching the segment count of fixed-length segmentation. We further attach structural paths to the segments for the agent to identify their ancestors and hierarchical positions. We also introduce residual memory links from the segments of low-level children to the segments of high-level parents, enabling direct information flows beyond successive memory updates. We evaluate TreeAgent with four representative memory agents on eight retrieval and aggregation tasks. The inputs are derived from real-world application data in JSON, HTML, YAML, and JavaScript and range from 8K to 1M tokens. TreeAgent improves the average score across the four datasets for all four agents, with 3.61% performance gain on average.
Abstract (translated)
URL
https://arxiv.org/abs/2611.13946
PDF
