Papers

Autoregressive generation in Large Language Models (LLMs) is constrained by the memory and computational demands of attention mechanisms. Sparse attention methods mitigate this cost by selecting only high-probability entries of the attention matrix. We observe that in many such methods, this renders the probability-value multiplication negligible, shifting the bottleneck to the query-key step. Key heads can therefore be reduced to accelerate inference, while retaining more value heads preserves capacity with limited additional decoding cost. We introduce Sparse Asymmetric Group-Query Attention (SAGA), which decouples key and value head counts to exploit this principle, and pair it with approximate top-N (Atop-N) attention, a simple sparse attention method designed to study the interaction between sparsity and head-count asymmetry. We formalize the benefits of this asymmetry theoretically and validate them empirically through latency measurements and quality evaluations on models up to 1.5B parameters. Together, SAGA and Atop-N achieve end-to-end decoding speedups exceeding $2\times$ over our full-attention GQA baseline at long contexts. Models trained from scratch with SAGA nearly match the quality of comparable GQA variants on the evaluated benchmarks. To facilitate adoption, we introduce an efficient fine-tuning method that converts pretrained models to the SAGA architecture, enabling practitioners to benefit from our approach without costly retraining.

Structured prediction tasks pose unique challenges for in-context learning (ICL): their compositional outputs require modeling fine-grained, token-level patterns that sentence-level approaches fail to capture, and their task-specific annotation conventions are human-defined artifacts that cannot be acquired through pretraining alone. We propose Verb-ICL, a selective annotation framework for ICL-based structured prediction that addresses both challenges. Verb-ICL first selects representative examples using a token-level coverage strategy that captures local semantic patterns critical for structured prediction, then generates actionable error feedback that codifies task-specific annotation guidelines and incorporates this feedback into ICL demonstrations. We evaluate Verb-ICL on six structured prediction datasets spanning information extraction and semantic parsing. Experiments with recent LLMs show that Verb-ICL consistently outperforms strong selective annotation baselines under low-resource settings and continues to provide gains as the annotation budget increases. Extended analyses demonstrate that the generated feedback is predominantly useful across a four-category quality taxonomy, generalizes as task-level guidance beyond instance-specific corrections, and improves performance regardless of the underlying selection strategy.

In a Transformer, token mixing is the step that lets each token draw information from other tokens, and it dominates the cost of encoding long sequences. Self-attention does this mixing very well: every token weighs every other token by content, which gives strong contextual modeling. That all-pairs comparison is also why its cost grows quadratically with sequence length. We introduce WNet, a Transformer encoder that replaces self-attention with token mixing based on the discrete wavelet transform (DWT). Three attention-free mixers recombine the scales: by linear fusion, by learned gating, or by letting each token choose its scales. A hybrid adds self-attention in the last layer only. A receptive-field analysis shows that wavelet mixers built from two-tap filters, such as Haar, never relate tokens outside fixed blocks, however deep the network, even when the filters are learned. Longer filters reach the whole sequence within two layers. We pre-train every model with masked language modeling on a fixed-token subset of C4 and fine-tune on GLUE, using one controlled setup with size-matched BERT and FNet baselines and a control that cannot mix tokens. The token-gated mixer trains as fast as attention at 256 tokens and 2.7 times faster at 4,096.

A verifier for an agent faces rules of two kinds: the ones a fixed check can settle and the ones that require a judge. A team that derives its own checks fixes that split up front. Where the requirements come from outside, as in finance, healthcare and law, the agent enforces rules it did not write, so the split falls to runtime, recurring for every predicate of every rule on every action at a rate no reviewer can audit. Every escalation scheme assumes a model can make that decision itself, that it is self-decidable. Across six corpora, including the EU AI Act, FINRA guidance and a deployed credit agent, we collect roughly 22,000 labels from four models built by three labs. They agree almost perfectly where the answer is obvious and collapse on regulatory text; their errors run in opposite directions, so no model can be trusted as the conservative choice; and on the deployed agent's own rule-set they err together, over-claiming that a fixed check will do, the direction that never gets escalated. We introduce CoVer (corroborate-then-verify), which treats unanimity as a nomination, admitting a predicate only when the check synthesized for it survives intervention, reading fields the agent cannot write and holding under deterministic rewording. That gate rejects most of what corroboration wrongly admits, at a cost in coverage we report rather than tune away. The obvious alternative, agreement with a reference judge, certifies nothing: it climbs from 30% to 77% across calibration bands while the genuinely decidable share does not move, because a judge drawn from the population under indictment ratifies the blind spot it shares. Self-decidability is not a capability to elicit from a model but a boundary the verifier must construct.

Agents are entering finance, healthcare and law, sectors where a violation leaves no lexical signature and carries real penalties. Whether an omission is material, or a disclosure sufficient, depends on what the response left out. Probing such an obligation means finding responses one minimal edit from flipping compliance, and every probe costs a generation and two adjudications, so the binding constraint on regulatory red-teaming is sample efficiency, not volume. We introduce Penumbra, an adversarial search that walks from a verified anchor under an expanding edit budget until a two-evaluator committee changes its verdict, and emits the two adjacent responses that straddle the change. Allocation is adaptive, and the objective is coverage of the defeat surface: distinct (obligation x defeat mode) cells resolved per candidate. At matched budget, adaptive allocation reaches uniform allocation's full-budget coverage on 59% of the candidates; at equal records it covers 1.43x the defeat modes of naive enumeration, and the gain is confined to the axis it targets. On 60 screened obligations of a financial advisory constitution, Penumbra returns 144 pairs, each a compliant and a violating response that the committee places on opposite sides of the boundary, differing by a handful of words where a model asked for both directly produces texts sharing almost nothing. A second constitution, for clinical triage, reproduces this on 18 obligations: 49 pairs at the same tightness, with the same modes hardest. Pairs like these show where an obligation's own terms stop deciding, which is what an agent deployed under it must be tested against, and the search finds them at a cost that scales with the boundary, not the text.

Self-supervised learning (SSL) is standard for speech representation learning, but mainstream models are designed around single-speaker audio, limiting their usefulness in multi-speakers scenarios. We present SepRQ, an open-source SSL framework that replaces masked prediction with a pseudo-source-separation objective over frozen random-projection codebooks. By adopting a novel mask-free, multiresolution approach, SepRQ achieves state-of-the-art performance in Speaker Diarization and Speech Separation on the SUPERB benchmark, surpassing WavLM and other cocktail-party derived SSLs at both Base and Large scales, while requiring only 85.68M inference parameters. SepRQ also demonstrates strong performance across target-speaker tasks requiring enrollment (such as Target-Speaker Automatic Speech Recognition), and on the challenging multi-domain DIHARD 3 diarization dataset. Notably, we report strong separation capabilities on three-speaker mixtures (WSJ0-3Mix), where current SSL literature struggles. While cocktail-party SSLs remain scarce and closed-source, limited to C-HuBERT and the enrollment-based SA-WavLM, we open-source SepRQ to the community.

Answering questions over imperfect tables requires handling errors that can affect the answer. We investigate two challenges for large language models (LLMs): whether error discovery depends on where errors appear in a table, and whether providing their locations is sufficient for accurate question answering (QA). Using human-reviewed instances from RADAR-T, we conduct controlled studies across three LLMs by varying row order and comparing original, error-marked, and repaired tables. First, reordering rows changes error discovery even when the table contents and gold answer remain unchanged. During direct inspection, LLMs are more likely to discover all rows containing relevant errors when these rows appear later in the table or are grouped more closely together. Second, providing verified error locations alone is insufficient for accurate QA: with code execution, accuracy on repaired tables exceeds that on error-marked tables by 39.0-59.1 percentage points across the three LLMs. As a practical application of these findings, we combine error discovery across shuffled table views with explicit guidance for verifying and handling the reported errors in a simple workflow, Geometry-Balanced Discovery and Intervention (GBDI). On RADAR-T, GBDI improves QA accuracy by 3.8-18.5 percentage points over a code-agent baseline across five LLMs (paired 95% confidence intervals exclude zero for four), at the cost of additional inference. These results highlight the importance of both reliable error discovery and effective error handling in QA over imperfect tables. Code is available at https://github.com/645-t/GBDI-ICLR-2027.

Multi-agent debate can improve reasoning, yet often fails to beat simple majority voting. We argue that successful debate requires two distinct mechanisms: proposal supply must surface a correct answer, and readout must identify that answer when voting misses it. We formalize the first requirement through recoverable headroom, which measures cases where a correct proposal is available but the majority answer is wrong. For the second, we develop Latent Verification Debate (LVD), an accounting model in which candidate proposals receive answer-specific verification evidence before final generation. Controlled fixed-proposal interventions estimate this latent effect in equivalent peer-support units and show that correct evidence changes answer probabilities and generated decisions while proposal supply remains fixed. To improve proposal supply, we construct societies from neural-thicket agents using labeled and label-free coverage objectives. Across two backbones and matched-budget reasoning benchmarks, coverage-selected societies increase complementary proposal supply and improve aggregate accuracy in repeated stochastic evaluations. Round-level controls further show that interaction provides gains beyond applying the same finalizer directly to the initial proposals. These results identify proposal coverage and truth-sensitive evidence use as complementary conditions for debate to outperform voting. Code is available at https://github.com/Wang-ML-Lab/when-debate-helps.

Activation steering has proven effective for controlling the behavior of Large Language Models (LLMs) at inference time, but its application to SpeechLLMs remains new, and training-free steering approaches for such models are still largely unexplored. We propose a training-free Contrastive Activation Addition (CAA) protocol that derives steering vectors for common speech tasks (e.g. transcription) in SpeechLLMs from a small number of labeled utterances. We showcase that adding these vectors in the representation space, at inference time, enforces better the targeted speech task. We further show that, when combined with prompting, these vectors yield to consistent improvement over prompting alone on most evaluated tasks such as Automatic Speech Recognition (ASR) or Emotion Recognition (ER), and transfer to out-of-domain data. We additionally demonstrate the usefulness of script-normalization directions to enforce the target script of a specific language.

Large Language Models (LLMs) can adopt distinct personas to tune their semantics, expertise, and perspective to different users and tasks. Precise control over these traits is critical to ensure safety and reliability in model behavior. Existing methods like activation steering and prompt-based persona induction reduce a persona to a single dominant direction, missing the finer, nested traits that emerge only once that dominant signal is factored out. We introduce modulation as a setting where the persona context is already embedded in the content being manipulated, requiring control methods to amplify or suppress a trait already present rather than inject it from scratch. PaSS is an inference-time control paradigm that models personas as multi-dimensional subspaces in a model's latent space without supervised contrastive examples. The persona subspaces are extracted via iterative concept erasure and applied to modulate persona-guided generation without retraining. To extract this subspace, we use Iterative Nullspace Projections (INLP) to linearly and iteratively isolate persona-specific directions. We causally evaluate six personas against diverse tasks like MATH-500, TinyAlpaca, GSM8K, and IFEval, showing that discriminative, iterative subspace extraction captures diverse traits underlying a given persona, enabling stronger and larger modulation than single-direction additive methods, while maintaining content fidelity. We further study individual peeled directions within each subspace to uncover the distinct aspects of persona behavior they encode. Overall, we show that persona subspaces offer a controllable, interpretable, and generalizable framework for modulating LLM behavior without sacrificing task performance.

Detecting responses that retrieval-augmented generation does not ground in its context trades speed against accuracy: surface checks miss paraphrased fabrication, sampling-based methods cost extra generations. Hidden-state probes sit between the two, but every existing one reads the generating model's own activations, so a change of generator invalidates the detector and a closed-weight generator is out of reach. This paper removes that coupling. The Grounding Probe is logistic regression over the mean-pooled middle-layer hidden states of an observer language model that reads the context, question, and response in one forward pass and generates nothing, with the recipe it needs: pool over response tokens, read a middle layer, and control capacity, which closes the train-test AUROC gap from 0.087-0.202 to 0.009-0.013. Asking the observer outright, rather than reading its hidden state, costs at least +0.166 AUROC in every one of four models. Fitted on 15,090 annotated responses it reaches 0.879-0.894 AUROC on RAGTruth test across four observers, and 0.924 AUROC with 0.820 F1@0.5 averaged with a supervised span detector, 0.060 above that detector alone. One probe holds across six generators, and hold-out controls, including one in which no evaluation prompt appears in training, bound the cost of removing a generator at about 0.02 AUROC. Code, probes, and predictions are released.

Increasing architectural complexity is often assumed to improve health misinformation detection, yet reported gains are difficult to interpret when studies use different corpora, preprocessing pipelines, data splits, and leakage controls. This study provides a controlled cross-corpus benchmark of five compact neural architectures (1D-CNN, LSTM, BiLSTM, CNN-LSTM, and CNN-BiLSTM), a soft-voting neural ensemble, and three classical machine-learning baselines using COVID19-FNIR and CONSTRAINT. Exact-text duplicate controls were applied before modelling; all neural systems used a common preprocessing and optimisation protocol, and neural results were repeated across three random seeds. On COVID19-FNIR, the deep ensemble achieved a mean macro-F1 of 0.9963 and ROC-AUC of 0.9994, while individual neural models ranged from 0.9945 to 0.9957 macro-F1. On CONSTRAINT, the ensemble achieved macro-F1 of 0.9272 and ROC-AUC of 0.9811, whereas a linear SVM achieved macro-F1 of 0.9574 and ROC-AUC of 0.9931. Architecture rankings changed across corpora, and simple sparse linear models remained highly competitive. The findings show that model complexity does not provide a stable performance advantage and that benchmark construction can dominate architecture choice. The study contributes a reproducible, leakage-controlled basis for evidence-driven model selection in health misinformation classification

Large language models (LLMs) increasingly mediate human communication, from drafting emails to summarizing scientific reports, yet whether they faithfully preserve a speaker's position remains largely untested. We model AI-mediated communication as a two-step generation-extraction pipeline: one LLM produces an argument from a specified stance, and a second LLM extracts the stance from that argument. We represent the pipeline as a probabilistic state transition over five Likert-type stance categories and define the stance preservation rate (SPR) as the average probability that the extracted stance matches the initial stance. Across 112 debate propositions, none of the nine LLMs tested exceeded an SPR of 0.7 under the default configuration. Three drift patterns accounted for most of the drift: polarization, deviation from neutrality, and flipping. Among the mitigation strategies tested, including in-context learning, multiple extraction with shuffled options, assertion, and reflection, only adding medium reasoning effort to a reflection prompt for GPT-5.4 substantially improved the SPR, to 0.775, yet polarization remained the largest pattern, with 0.119 of the transition mass. An exploratory comparison with human extraction on a single proposition suggests that drift arises at both the generation and the extraction stage. These results point to a fidelity gap in AI-mediated communication, with implications for journalism, policy deliberation, scientific communication, and other domains where opinion-laden messages pass through language models.

Chain-of-Thought (CoT) faithfulness detectors are widely used to audit reasoning models, yet a detector is itself a predictor whose verdicts are treated as stable properties. We ask whether a detector is faithful to itself under distribution shift. We formalize meta-faithfulness as an invariance principle: a valid detector must return identical verdicts on traces that differ only by transformations preserving ground-truth faithfulness. We prove three results: (i) no detector using only intervention-response profiles can separate faithful from epiphenomenal mechanisms with identical signatures; (ii) any detector relying on shift-sensitive features violates invariance at a rate independent of its in-distribution accuracy; (iii) an asymptotic certified selective-risk guarantee enables confident abstention. We operationalize the principle in FaithShift, a stress-test protocol spanning ten shift axes, and propose SIFT, a hidden-state trajectory detector trained with cross-environment invariance objectives and certified abstention. Across 14,996 traces, four domains, and eight models, three findings emerge. First, transfer collapse is real: all existing detectors show gaps $\geq 0.15$ AUROC. Second, the dominant bottleneck is sampling stochasticity, not shift: over 80% of detector instability stems from random seed variation, falsifying our preregistered prediction that shift-attributable violations exceed 0.25. Third, SIFT cuts invariance violations by 64% over the best single-seed baseline, but a four-seed ensemble of any detector narrows the margin to 0.01 (indistinguishable at matched coverage, $p=0.21$), and SIFT needs a 51% abstention rate. Cross-model transfer degrades from within-family to cross-family to open-weight-to-API, partly closed by multi-model training. We offer a framework for auditing auditors: the real barrier is detector variance, not distribution shift.

Is agentic memory robust against stealthy steganographic attacks? We carry out a large-scale red-teaming exercise to test whether agents can encode attacker-controlled strings in one session and recover them in another without triggering safety oversight. Following SHADE-Arena-style tasks, we embed malicious side tasks to encode secret strings using steganography within otherwise benign tasks and evaluate them using independent task-completion and safety oversight. We test 14,000 attack trials spanning 91 model-attack configurations, 13 models, and 7 steganographic schemes: Acrostic, Mesostic, Telestich, Word-choice, Capitalization, Position-based, and Homoglyph. 41.2% of trials successfully encode the target payload in the generated response, while 29.8% survive the memory-write step with the encoding intact. Overall, 25.4% of trials yield a recoverable payload at retrieval, and 20.1% achieve exact-match recovery, including synthetic API keys and cryptographic material never exposed to the receiver. Among the exact-match transmissions, 71.4% also pass primary task-completion scoring and are independently judged safe by the oversight model, corresponding to 14.4% of all trials in which a successful covert transmission would appear to be an ordinary, benign interaction under task-level evaluation. Our results demonstrate that agentic memory can function as a persistent cross-session covert channel. The results further show that the principal bottleneck occurs at memory persistence rather than retrieval: once a steganographic payload survives the memory-write stage, a substantial fraction remains recoverable. We therefore argue that memory integrity, information-flow control, and covert-channel detection should be explicit security requirements for agentic systems.

Multi-modal large language models (MLLMs) achieve strong modality understanding by pairing a large language model (LLM) with an encoder for a target modality such as vision, video, or audio. However, improving an MLLM's capability for a given modality typically requires additional training on large modality-specific datasets, incurring substantial data collection and compute costs. Model merging offers an alternative, but it is often infeasible for data-scarce, large per-sample size, or domain-specific modalities (e.g., audio and video), where same-modality model variants are rarely available. In this work, we characterize an intriguing asymmetric phenomenon: merging a well-aligned, data-rich source-modality MLLM into a data-scarce target-modality MLLM substantially improves the target on its own benchmarks. Our theoretical and empirical analyses show that this gain stems from enhanced alignment between modality-specific and textual tokens, induced by the stronger donor modality. Specifically, we derive a mutual-information lower bound that is monotonic in alignment-related quantities and strongly correlated with downstream MLLM performance. Building on this principle, we propose Directional Cross-modal Alignment Transfer (DCAT), a novel framework that transfers textual alignment from a strong, well-aligned source (donor) modality to a weak target (recipient) modality, boosting target-modality performance without further fine-tuning. We further show that the alignment-enhancing objective admits a closed-form weight-space solution computed from only a small calibration set. DCAT outperforms existing model-merging methods, offering an efficient path toward cross-modal alignment transfer. Project page with code is available at \url{https://seohoiki3215.github.io/DCAT_project_page}

Activation probes can predict safety-relevant properties of language models with high area under the receiver-operating-characteristic curve (AUROC), but deployed agent monitors make thresholded alarm decisions under tight false-alarm budgets. These are different estimands. We introduce an Operational Validity Contract that fixes a monitor's target, observability, identity, timing, intervention unit, comparator, calibration, and cost. We formalize risk at the semantic request or trajectory level: when one task contains repeated alarm opportunities, row-level AUROC and false positive rate do not identify semantic-unit any-alarm risk. A confidence-certified threshold also requires enough independent negative units, a tie-safe rule, and transport to deployment. Across Models Under Pressure, LASR refusal prediction, immutable AgentDojo, and a prospectively protocol-frozen ST-WebAgentBench replication, joint activation-observable monitors reach AUROC 0.957 and 0.935 on the first two benchmarks, yet their locked 5%/10% detection rates are only .642/.742 and .719/.782, respectively. On AgentDojo, the secondary mean-activation rollout monitor reaches AUROC 0.922 but detects none of 38 positive semantic cases at the locked 5% operating point; thresholds intended for 10% false alarms realize 18.7-20.0% on test. Because the test misses its prospectively frozen 40-positive support gate, we label it support-insufficient. On ST-WebAgentBench, activation reaches AUROC .874, but 23 independent calibration negatives cannot identify even a 10% controller; the locked policy abstains rather than reporting its mechanical zero FPR as a success. An exploratory counterexample also lowers full AUROC while improving realized 10% utility. The fail-closed compiler caps MUP and LASR at restricted predictive value and AgentDojo and ST-Web at representation accessibility; no setting reaches alarm-policy validity.

Purpose: To develop and evaluate an open-weight large language model (LLM) pipeline that converts an entire archive of free-text radiology reports into structured reports without human oversight. Materials and Methods: In this retrospective study, a pipeline with 150 hierarchically organized templates was developed at one center and tested at a second center on reports from 2010 to 2025. The open-weight model gpt-oss-120B selects the template in three constrained-decoding steps and fills it on one local graphics processing unit. Template selection was scored against expert labels on 914 randomly sampled reports of five modalities, structuring quality on 920 radiography and CT reports corrected field by field by five residents. The pipeline then processed the complete archive of the second center. Proportions are reported with Wilson 95% confidence intervals (CIs). Results: An optimal template set was selected for 74.4% of reports (680 of 914; 95% CI: 71.5%, 77.1%) and an appropriate set for 82.3% (752 of 914; 95% CI: 79.7%, 84.6%), 87.7% for single-region and 54.1% for multi-region reports. Macro semantic textual similarity between output and corrected reference was 0.95 for radiography and 0.97 for CT, residents left 88.7% of 24,638 fields unchanged, and unsupported content was flagged in 1.0% and 1.5% of reports. Of 2,186,982 archive reports, 96.5% received structured output, 2,401,544 structured reports, at 1,258 reports per hour on one graphics processing unit. Conclusion: An open-weight LLM pipeline structured a complete multimodality report archive without human oversight with high content fidelity. Multi-region reports remained the main source of template errors.

Answer correctness is encoded as a recoverable geometric direction in the hidden states of language models. We show that the mean displacement from incorrect to correct answer representations, computed at approximately 70% of model depth from fifty labeled examples with no parameter updates, yields a scoring direction that outperforms zero-shot log-probability scoring by up to +32.0 percentage points on factual benchmarks (ARC-Challenge and MMLU) and by +38.1 to +51.8 percentage points on TruthfulQA, across five models spanning 1B to 8B parameters in three architecture families (Llama, Qwen, Gemma). The method requires one forward pass and one dot product per candidate; no generation is performed at inference. Applied as a hallucination detector on individual (question, answer) pairs, the recovered direction achieves 0.693~AUROC versus 0.578 for log-probability scoring. We additionally find that correctness directions for factual reasoning, domain knowledge, and calibrated truthfulness are near-orthogonal in representation space, revealing that language models allocate geometrically independent subspaces to qualitatively distinct notions of correct answer, with architecture-dependent variation in the degree of separation. This structure explains the observed transfer pattern—the direction calibrated on factual questions transfers within task type but not across it—and suggests that LLM calibration failures may reflect a routing problem: the model's internal representation contains more correctness signal than its output behaviour exploits.

LLM-agent security has produced a dense landscape of defenses - prompt hardening, content filters, permission gates, sandboxes - yet no framework tells a deployer what a defense actually guarantees, or where that guarantee comes from. We apply Verification Autonomy Levels (VAL) - L0: LLM self-declaration; L1: deterministic rules; L2: objective ground truth; L3/L4: decidable completeness; L5: impossible - to 22 agent-security defenses; the taxonomy is falsifiable (10/10 prediction hits on frozen cards, flagged). We run the first controlled deployment-value comparison: at equal budget, a VAL-guided stack (confirmation gate + schema sandbox) versus a mainstream intuition stack (prompt hardening + keyword filter), 50 scenarios, 12 attack variants, adaptive/white-box/PAIR escalation (~7,000 testbed calls; ~10,000 harness calls on AgentDojo/JADE). The VAL stack holds 0.000 attack success at 1.000 benign success (0.5% ASR at 79.7% utility on AgentDojo banking vs 4.3% undefended); the intuition stack reaches 0.000 ASR but kills all benign actions - security by model-behavior luck, not structure. Across testbeds of rising attack-surface hardness the intuition stack's zero drifts (0->1.9%->6.2%, n=16 on JADE) while the VAL stack's holds within its ODD (0->0->0), its only breach a disclosed out-of-ODD password gap (0.5%). The same zero, two different guarantees: zero is an outcome, not a guarantee.

Radiology report generation (RRG) aims to convert medical images into diagnostic texts to assist in clinical decision-making and alleviate the workload of physicians. Although existing methods have made extensive progress in cross-modal interaction and the incorporation of external priors, the distribution shift of underlying representations and the undifferentiated rigid coupling of heterogeneous information cause weak visual abnormality cues to be easily diluted by massive text priors and generation inertia during decoding. To overcome this bottleneck, inspired by cognitive science, we propose a novel two-stage Homogeneous Semantic Alignment and Hierarchical Expert Routing (HSA-HER) framework. First, the model introduces an explicit homogeneous distribution constraint in the underlying latent space to effectively eliminate the cross-modal distribution shift between visual and textual features, thereby extracting purified visual features as semantic anchors that accurately align with diseases. Second, for heterogeneous clinical evidence composed of visual features, local entities, and global retrievals, we design a hierarchical expert routing mechanism guided by these disease semantic anchors. This mechanism abandons the undifferentiated rigid coupling paradigm. Specifically, it dynamically activates expert networks to perform targeted mining and semantic reconstruction on multi-source evidence, and adaptively allocates fusion weights. Extensive experiments on three mainstream benchmark datasets demonstrate that HSA-HER achieves state-of-the-art performance, accurately depicting complex imaging details and key diagnostic information.

Nowadays, emojis are often replacing words. Yet computational systems still oversimplify them. Most existing approaches treat emojis as static sentiment indicators and overlook their emotional distributions. In this study, we propose an emoji-aware emotion analysis framework based on a Twitter (X) dataset of 100,000 emoji-containing replies collected between 2020-2025. After preprocessing and text cleaning, we applied text-to-emotion classification to detect five primary emotions (Happy, Angry, Sad, Fear, and Surprise) for each message. By aggregating emotion scores across contexts in which each emoji appears, we estimate emoji-emotion association distributions and construct an emoji-emotion ranking system reflecting relative emotional dominance. Furthermore, we project emojis into the Russell valence-arousal space to enable continuous affective interpretation. Our results demonstrate that emojis exhibit probabilistic, context-sensitive emotional profiles rather than fixed sentiment polarities.

Understanding parameter functions helps elucidate the internal mechanisms of large language models (LLMs). However, how to connect parameters from different modules to verifiable output effects and further characterize the relationship between their functional organization and model capability remains to be explored. To this end, we introduce the downstream vocabulary lens (DV-Lens), a parameter-level interpretability framework that links native parameter directions to their downstream vocabulary responses. Specifically, we first estimate module-specific downstream Jacobians over a reference prompt set for attention query, key, value, and output (Q/K/V/O) projections and feed-forward networks (FFNs). Second, we use these mappings to project native parameter columns into the final vocabulary space, obtaining signed readouts that characterize their average local output responses. Third, we group parameter columns by their vocabulary readouts and introduce downstream vocabulary complexity (DV-Complexity), which quantifies within-group structural variation using normalized reconstruction residuals of the original weights. At the parameter level, randomized controls and finite-difference tests show that DV-Lens readouts capture non-random vocabulary structure and predict local logit changes with 98.0% coordinate-orientation agreement across 720 cases from nine models. These readouts further guide parameter ablation, steering, and swapping across 21 models, shifting target-token probabilities in the predicted directions under controlled conditions. At the model level, the joint-parameter score of DV-Complexity achieves a Spearman correlation of 0.904 with benchmark-based capability rankings across 48 language models. Together, these results provide intervention-based evidence for DV-Lens interpretations and reveal an association between DV-Complexity and model capability.

Artificial intelligence enables computational agents to represent political preferences and take part in collective decision-making. In this thesis, I investigate the opportunities and challenges of digital twins (DTs) based on Large Language Models (LLMs) as intermediaries in augmented democracy, focusing on individual preference representation, collective representation of political organizations, and the vulnerability of those representations to attackers. First, using data from an online experiment in Brazil, I examine whether personalized DTs can predict citizens' preferences for unseen policy proposals. Second, I extend the DT framework from individuals to political organizations. Using Swiss parliamentary data, I build topic-specific knowledge graphs from lawmakers' legislative records and connect them to LLM-based lawmaker agents, which are organized into party-level DTs representing collective positions. Agentic deliberation among these agents tests whether aggregated party representations capture a broader range of intra-party perspectives than official party communications. Finally, I study the vulnerability and robustness of LLM-mediated deliberation against prompt-injection attacks that amplify viewpoints, suppress opinions, or redirect consensus. Using data from a 2023 deliberative experiment in the United Kingdom, I analyze how attack effectiveness varies with the distribution of opinions and rhetorical strategies, and evaluate a pipeline combining injection detection, structured opinion representations, and reinforcement learning to improve resistance. These findings characterize the opportunities and challenges of LLM-based digital twins in augmented democracy, stressing accurate preference representation, faithful aggregation, and robustness to strategic interaction.

Large language models (LLMs) increasingly serve as autonomous agents that invoke external tools. However, this capability introduces tool hallucination, selecting incorrect tools or generating invalid calls. Existing mitigation methods report substantial improvements, yet we identify a previously overlooked failure mode that we term Hallucination Escape. These methods reduce hallucination on the tool configuration they are tuned on but increase it on other configurations, canceling out the gain. We further investigate this phenomenon and find that hallucination rises sharply when a model's intrinsic tool-use tendencies conflict with the current tool configuration, and that existing methods reinforce rather than suppress these tendencies, which in turn contributes to hallucination escape. Building on these findings, we propose EscapeGuard, a training-free inference-time method that combines conflict-aware gating with configuration-derived attention enhancement to mitigate tool hallucination while preventing hallucination escape. Across six benchmarks on various models, EscapeGuard reduces tool-selection hallucination by 9.0 pp and suppresses hallucination escape, lowering the cross-configuration mean by 23.7 pp and achieving an 89.1% net improvement in paired-query evaluation. We hope this work can encourage evaluation beyond a single tool configuration and pave the way for more reliable tool-using LLM agents.

Post-training quantization to the GGUF format's mixed-precision K-quants is commonly how open-weight language models reach consumer hardware, yet its effect on fine-grained lexical competence is uncharacterized. We audit 27 quantized artifacts across 13 families and four architecture backbones, 0.35B-14B parameters, evaluated down their published ladder to Q2_K (about 2.6 bits per weight), on 429 frequency-validated rare English words under two probes: surface inclusion of a prompt-supplied word and its one-sentence definition, scored by a tiered multi-synonym matcher, its error measured by a blind LLM-judge census of every definition, with human verification. Three regimes emerge at Q2: total collapse into unusable builds, severe semantic dissociation in sub-2B models, and mostly robust preservation above about 3B. In every sub-2B artifact, definitions fall 20-67% below the artifact's baseline, typically several times the inclusion loss. Two controls separate rarity from task difficulty: within the rare set, loss rises with rarity in six of seven sub-2B artifacts, and on a 100-word common-word set rare words lose more than common words in all eight, significantly in six. Tokenizer vocabulary size does not predict the damage (Spearman rho=0.12); parameter count dominates (rho=0.72), confirmed within five of six same-tokenizer families. Q4_K_M remains lexically clean at >=1B. The damage is frequency-graded, provider-dependent, and not calibrated by WikiText-2 perplexity: across nine artifact-matched ladders, near-identical Q2 penalties (44.7%/47.6%) separate an artifact keeping its definitions (3.6%) from one losing them (43.6%). Aggressively quantized small models can keep generating fluent text while no longer knowing what it means, risking hardware-constrained deployments in domains where semantics carries consequences. Validation must be per artifact.

Public vs. private universities is a debatable issue, and it creates polarization on social media in Bangladesh. Debate on quality, jobs, and prestige is passionate among the students, parents, and graduates, the majority of whom speak Bengali, a low-resource language. To measure this polarization, this paper introduces a manually annotated dataset of 4,060 Bengali comments labeled as Pro-Public, Pro-Private, or Neutral. We evaluated the quality of our annotations by Fleiss's Kappa agreement that was 0.89, corresponding to a high agreement among annotators. The classical ML (SVM, Random Forest, XGBoost), BiLSTM network, hybrid BanglaBERT+XGBoost models and the state-of-the-art zero-shot LLMs (Claude Sonnet 4, DeepSeek-V3.1, Llama 4 Maverick, Kimi K2 Thinking, Qwen3-235B Thinking) models are evaluated. The accuracy of BanglaBERT+XGBoost is 91.81% and macro F1 score is 91.70%, which is higher than all the supervised baselines. The zero-shot Llama 4 Maverick Thinking achieves a macro F1 of 0.931 (overall accuracy of 93.31%) without any fine-tuning. All machine learning (ML), deep machine learning (DL) and transformer models were outperformed by the zero-shot Llama 4 Maverick model. Polarization also is evident, in some ways more clearly in the Pro-Private comments, which emphasize modern facilities and timely graduation, versus the Pro-Public comments, which emphasize affordability and government jobs. Our findings open new directions for analyzing social media polarization in low-resource languages.

General turn-taking behavior in real-time dialogue systems requires deciding whether to keep listening or start responding while listening, and whether to continue or stop while speaking. Existing turn detectors use heterogeneous, task-specific label spaces and are often trained on limited annotations or evaluated on isolated utterances, making them difficult to use as a unified causal controller with comprehensive context. We propose XTurnix, a compact text-based model that formulates turn control as two binary decisions conditioned on the AI's current listening or speaking state and predicts a single control token from the complete dialogue history. XTurnix is pretrained on 5.5 million causal action examples automatically derived from timestamped two-speaker transcripts, then fine-tuned on synthetic multi-turn examples with a flatter distribution across the four state-action labels. We evaluate XTurnix on four public benchmarks and a balanced self-curated benchmark. Across the public benchmarks, XTurnix achieves the best results on all SemanticVAD and LiveKit splits, ties the native Smart-Turn model on Smart-Turn Bench, and achieves the highest incomplete-turn accuracy on Easy-Turn. On the self-curated benchmark, it reaches 89.06% accuracy, more than 20 percentage points above the strongest third-party baseline at 68.75%, while maintaining F1 scores between 84.21% and 90.63% across all four categories. These results demonstrate unified listening- and speaking-state turn control in a single compact model. Code is available at https://github.com/xcc-zach/xturnix, with an interactive demo at https://huggingface.co/spaces/xcczach/xturnix-demo.

Glitch tokens are anomalous vocabulary entries that can cause large language models (LLMs) to produce outputs inconsistent with their inputs. Existing repair methods require access to model internals, making them impractical for frozen checkpoints. We investigate whether glitch tokens can be repaired outside the model by optimizing the input tokenization. An empirical study on BPE merge-rule deletion reveals that (1)deleting a glitch token's merge rule can fix a substantial fraction of failures, yet disrupting normal tokens sharing intermediate merge nodes causes the overall glitch rate to rise, and (2)different decomposition granularities yield non-monotonic fix rates while collateral damage on normal tokens grows monotonically. Motivated by these findings, we propose GlitchPatch, a repair framework for frozen language models based on local retokenization, consisting of two stages: the offline stage uses Behavioral Path Optimization (BPO) to find the behaviorally optimal replacement token sequence for each glitch token and compiles validated replacements into a rule table; the online stage substitutes only the IDs of matched glitch tokens in the canonical token sequence, with no modification to model parameters or internal states. Experiments on ten models spanning six tokenizer families show that GlitchPatch achieves an 85.10% mean fix rate, outperforming the strongest baseline by 14.37 percentage points, and reduces the average glitch rate from 14.88% to 2.27%. GlitchPatch achieves a 0.00% RR in full-vocabulary evaluation and leaves rule-unmatched inputs unchanged by design. We further evaluate the practical impact of repair from the perspectives of time cost, language understanding, and capability, supporting its deployment feasibility. We hope this work provides a practical option for improving tokenizer reliability.

Data quality now matters as much as compute for training language models. Much training data comes from human annotation of text, and interpretive annotation has no ground truth that could settle what is "accurate". Two lines of work respond to this. One combines annotators into a "ground truth" and measures how well they agree with each other; the other treats their disagreement as a signal. Both compare different people at one point in time. We measure something else: how well one reader reproduces their own reading of the same text over time. One expert human reader and three LLM families segmented three Sumerian myths and labelled the causal function of each segment with one of seven states. Across runs months apart, the human cut the text in much the same places but named the segments differently, in every myth. The models show no such consistent pattern: their gap between the two layers is positive in some myths and negative in others, and its size varies. The human's label changes are not random: the runs go through much the same functions but start them one step apart, while model runs start them at the same places. We argue that this pattern is a usable measure of data quality and a contamination check: a "human" annotation whose labels are as stable as its boundaries, and whose functions start in sync, looks like a model's.