Papers

Structured pruning of large language models (LLMs) offers hardware-efficient compression, yet existing methods require calibration data, gradient computation, or large auxiliary policy networks at pruning time. LILA (Latent-Informed Layer Analysis) scores neuron importance via the Kolmogorov–Smirnov (KS) distance between empirical singular value distributions of the full and neuron-ablated feed-forward network (FFN) weight matrix, providing a closed-form spectral rule requiring no training, calibration data, or auxiliary network. Without any fine-tuning, LILA surpasses PruneNet (45M-parameter RL policy) by 1.57~pp in zero-shot accuracy on LLaMA-2-7B at 25% sparsity, and outperforms WikiText-2-calibrated SliceGPT by up to 6.0~pp across all sparsity levels, while preserving the original architecture. After one epoch of LoRA recovery fine-tuning, LILA achieves highly competitive performance, matching the heavily calibrated SliceGPT baseline to within a 0.48~pp margin across LLaMA-2-7B and Phi-2, despite using zero calibration data. A Neural Tangent Kernel analysis confirms a 22$\times$ reduction in functional distortion versus random pruning, providing theoretical grounding for the spectral importance criterion. Finally, extending LILA to dynamically allocate sparsity budgets via KS-scores yields state-of-the-art generative preservation at moderate compression, while uncovering fundamental single-layer architectural bottlenecks at higher compression regimes.

How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments report repeated phrases within ten generations. Under a fixed sampling seed in vLLM, the fast loss of lexical diversity comes from the sampler. When vLLM serves a batch from one seeded sampling configuration, every request receives the same random draws, and a fixed seed replays them every generation. Fine-tuning raises the tokens that won, and the replayed draws let them win by more. Sharing across requests and replay across generations matter only together. Remove either one, by changing the shared seed every generation or by giving each request its own seed that repeats every generation, and the unique-4-gram fraction of two StableLM checkpoints stays near its starting value of about 0.98 through generation 3. Keep both, and the replayed shared seed takes seven checkpoints from five families to between 0.045 and 0.38 by then. Three generations of replay write the favoured phrases into the weights: decoded with one seed per request, the generation-3 weights of the replayed StableLM-2-1.6B chain recover most of their diversity, yet the phrase that filled every sample under the shared seed still opens 46% of them. Without replay, five checkpoints drift slowly, consistent with the gradual accumulation that theory describes, and three turn incoherent though their diversity scores stay high. One peer-reviewed model-collapse pipeline that fine-tunes Gemma-2-27B samples identical prompts under one seeded configuration, and three quarters of the rows it released for one iteration repeat nearly as often as one such batch copies them. A seed per request restores the fresh sample that stability analyses assume.

More and more of the text future language models learn from is written by a few of today's models. If one supplier writes most of a shared corpus, does it pull the models trained on it toward its own writing, or change how fast they drift? We retrain eight open models from their base weights on a shared pool of each other's text for five generations, varying the part written by one model, Phi-2, from an equal share to 90%. The models drift together toward a style with fewer function words, and none starts repeating itself. No share of Phi-2 brings the other models closer to its text than the equal share does. We split each ecosystem's separation from the equal-share one into a delay along its route and a departure from that route, both counted beyond the difference between two equal-share runs. With Phi-2 at 90%, delay outweighs departure 72 to 28 and 64 to 36 in two runs, and the ecosystem falls 2.7 and 2.5 generations behind. With Phi-2 at half the pool the two parts are about equal. When SmolLM2 or Qwen3-1.7B writes half instead, the ecosystem slows less or not at all. The departure leans toward Phi-2 more as its share grows, but more than toward every other model only at 90%. Human text filling a quarter or half of the pool slows the models along the same route.

Financial NLP has a standard workflow: validate a sentiment tool against human labels, then trust it to extract market signal. This assumes the two evaluations measure the same thing. We test that assumption in a setting where both can be measured at once: a corpus of securities class actions (2002-2025) linking 70,500 X messages to abnormal stock returns, with a single-annotator human labelled gold sample. Running five instruments (VADER, Loughran-McDonald, FinBERT, Twitter-RoBERTa, and an LLM annotator) through one identical pipeline, we find that the relationship between construct and predictive validity depends on the sampling convention and score representation. Under conventional method-specific sampling, human agreement aligns more closely with graded same-day associations than with one-day leads. On a fixed-n panel, however, agreement has similar graded rank correlations at both horizons, while the coarse ordering remains weak. Benchmark agreement therefore establishes semantic validity but does not by itself determine predictive rankings. In a conversation that is 17.6% spam, message volume predicts neither market damage nor settlement size.

Large Language Models (LLMs) are increasingly used to generate structured outputs, but their reliability remains unclear when those outputs must satisfy database-level constraints. We study this issue through database normalization, involving reasoning about functional dependencies, lossless join decompositions, and inter-table constraints. We introduce a Database Normalization Benchmark (DNBENCH), comprising 3,275 samples for evaluating LLM-driven database normalization from 1NF to BCNF. DNBENCH uses a three-axis protocol to measure semantic equivalence, structural accuracy, and logical validity. Across Single, Complex, and Real World levels, DNBENCH uncovers recurring failures in dependency inference, schema decomposition, and inter-table constraint reconstruction. We further propose Multi-Agent Reasoning for Schemas (MARS), which separates evidence extraction, violation diagnosis, and decomposition planning from schema generation and verification. MARS improves the DNB-SCORE by 82.0% over the single-prompt baseline. All artifacts will be released upon acceptance.

Human simultaneous interpreting (SI) is commonly assessed with analytic rubrics separating meaning transfer, delivery quality, and temporal synchrony, yet no automatic metric is designed for rubric-aligned segment-level SI evaluation. We construct a professionally annotated corpus of 1,101 SI segments with scores for meaning transfer (LQ), delivery quality (EXP), and perceived latency (LAT). We show that structured LLM prompting and scalar supervision collapse rubric dimensions, yielding near-zero correlation with human ratings and strong cross-dimension coupling. To isolate supervision structure under identical backbone capacity, we introduce dual regression heads on a LoRA-adapted COMET-KIWI encoder. On a held-out talk-level test set, the model achieves Pearson correlations of 0.388 (LQ) and 0.301 (EXP), improving over frozen COMET-KIWI. Given low absolute rater agreement, we interpret results relative to human consistency and target stable ranking signals for formative assessment.

In disinformation datasets, narratives are often understood as recurring interpretive patterns that group texts under narrative labels. Recent work formalized narrative mining as inductively inferring narrative labels from corpora, but its evaluation stays tied to predefined taxonomies, a closed-world setting that cannot capture narratives absent from the reference labels. We introduce a three-tier evaluation framework for unsupervised narrative label generation: recovery (against a corpus's own taxonomy), mining (against external label sets), and discovery (without predefined labels). Applying it, we compare clustering-based and graph-community-based pipelines across seven disinformation datasets, with human validation of discovery on two. The two families are complementary under automated metrics, but in a corpus with two prominent topics, clustering can reduce one topic to 2% of generated labels while graph-based pipelines stay balanced. Discovery validation also reveals many singletons (narrative labels derived from single claims, 30-62% of graph outputs), which clustering cannot produce. Annotators confirm many as recognizable disinformation narratives, suggesting that in open-world discovery the repetition assumed by narrative mining may be recognized outside the corpus, not within it. We release human-validated narrative candidate labels for the Climate Obstruction and PolyNarrative datasets to support taxonomy development and dataset extension.

This paper introduces the complete technical solution for the KuaiRP series of role-playing models. We aim to achieve four core objectives for a dedicated role-playing model: simplified prompt engineering, highly stable output quality, built-in domain world knowledge, and high-efficiency deployment with a small parameter size. However, effectively injecting deep domain knowledge often leads to a severe catastrophic forgetting of the model's general agent capabilities. To overcome this trade-off, we propose a multi-stage training pipeline. First, we design a standardized character template and construct an SFT data pipeline based on user behavior simulation and reverse profile filtering. Next, we utilize a rule-based composite reward function during the Reinforcement Learning (RL) phase to eliminate common degradation phenomena like length expansion and repetitive generation. Finally, to recover the general capabilities compromised during SFT and RL, we propose a novel self-distillation paradigm using Two-stage On-Policy Distillation (OPD) equipped with Cumulative-Divergence Decay (CDD). By using the domain-adapted model as the teacher and the original base model as the student, we effectively balance deep domain knowledge injection with the preservation of general agent capabilities. Experimental results demonstrate that the KuaiRP models not only match the current state-of-the-art proprietary models in role-playing fidelity within our target domains, but also successfully recover general agent capabilities, maintaining extremely low deployment costs.

Reproducibility is essential to scientific progress, yet the growing volume and complexity of scientific publications make exhaustive manual verification increasingly impractical. Although recent advances in large language model (LLM) agents enable automated experiment reproduction, existing evaluations largely focus on final repositories and are typically limited to machine learning (ML). We introduce AgentActionBench, a process-oriented benchmark for evaluating agent-based experiment reproduction across ML and AI4Science domains. Our framework uses an MCP-based Action Recorder to capture agents' behaviour throughout the reproduction process and evaluates the resulting traces with paper-specific rubrics. AgentActionBench contains 150 papers, including 120 ML papers and 30 AI4Science papers. A human-annotated subset covering 10% of the benchmark provides validation data, while model-assisted augmentation expands the full benchmark to more than 10,000 rubric items. Experimental results show that current systems remain limited, with execution as the primary bottleneck. Meanwhile, the strong Pearson and Spearman correlations between model-generated and human-annotated rubrics validate the reliability of our scalable rubric-generation approach.

Dispute mediation is essential for maintaining social harmony and resilience, yet developing skilled mediators is costly and time-consuming. Existing LLM-based mediation research remains limited by unrealistic task formulations, low-fidelity datasets, and coarse evaluation metrics that obscure turn-by-turn dynamics. To address these gaps, we introduce ProMediConv, a novel benchmarking framework that models mediation as a proactive, multi-stage, and party-aware dialogue process incorporating 11 mediation strategies and four party behavior pattern (BP) states. Using 972 complete real-world cases, we construct a high-fidelity mediation dataset with utterance-level annotations of strategies and BP states. Furthermore, to better assess agent impact, we propose MAD (Mean Attribute Difference), a fine-grained metric that captures BP shifts throughout the dialogue. Leveraging this framework, we establish a comprehensive benchmark by evaluating diverse models alongside our tailored baseline ProMediAgent. Extensive empirical analyses reveal critical behavioral phenomena and underscore the persistent challenges current models face in dynamic, multi-party mediation. Ultimately, ProMediConv provides a rigorous foundation and a vital quantitative standard for advancing AI-assisted conflict resolution. Our dataset and codebase are accessible at https://github.com/ZsWei66/ProMediConv_repo.

Neurosymbolic systems rely on mathematical solvers to guarantee reasoning correctness, yet solvers are fundamentally blind to whether a formal translation maintains strict reference-equivalence to a designated formalization. We formalize this vulnerability as Verdict-Preserving-Unfaithfulness (VPU): a failure mode where an incorrect encoding executes successfully and matches the expected verdict. We theoretically prove that structural, verdict-only verification heuristics are mathematically bounded to chance-level detection on these deceptively valid traces. To resolve this, we introduce Generative Verification (GenV), which distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score by repurposing the language model's native vocabulary space. Mechanistic analysis via decision-projected logit lenses and sparse autoencoders shows this generative readout natively extracts precise spatial error coordinates without explicit localization training. Empirically, our oracle-mined verifier (GenV+HN) achieves 0.961 AUROC in reference-equivalence verification, generalizes zero-shot across unseen translators and divergent formal styles, and yields an 11.3-point downstream accuracy gain in agentic test-time compute allocation.

Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.

Large language models learn similar behaviours, yet it remains unclear what structure they share or how to change one behaviour without disturbing others. The Fisher-Rao geometry of next-token probabilities connects these questions: behaviour determines this geometry up to output-preserving symmetries, whereas activation geometry depends on coordinates. Across transformer, state-space and recurrent models, output geometries agree more strongly than activation geometries, and shared geometry supports semantic-category transfer. Agreement with human word choices increases with predictive accuracy, scale and training, and improves further after model-only calibration. Token probabilities and read-out geometry jointly predict the spectrum and its effective dimension. Controlled language assignments show that geometry follows the language law across architectures. Pretraining corpus statistics predict held-out fact acquisition without recalibration, while randomised experiments show that deeper evidence substantially delays acquisition across every tested architecture and evidence construction. Finally, the geometry prescribes minimum-disturbance local interventions, predicts their relative cost, and supports reusable control: updates learned on donor prompts transfer to unseen prompts while better preserving behaviour on reference prompts than Euclidean control. The same geometric correction improves steering, editing, attribution, dictionary learning and fine-tuning.

Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize surface-level text spans. To address this, we present an information-weighted cross-entropy loss that rescales token-level contributions using TF-IDF statistics, emphasizing semantically informative tokens while down-weighting ubiquitous ones. Experiments on five decoder-only LLMs ranging from 1.1B to 13B parameters show consistent reductions in memorized substring length while preserving perplexity and downstream task performance. Under LoRA fine-tuning, TF-IDF reduces average substring memorization length by 14% across all five models. Under full-weight fine-tuning on TinyLLaMA 1.1B, the reduction reaches 58%. Our approach is architecture-agnostic and can be incorporated into existing training pipelines with less than 3% computational overhead, offering a lightweight and principled way to mitigate memorization without disrupting standard training dynamics.

A model first sees an image from one physical measurement experiment, such as how far a block coasted, and must answer a question about a new trial, such as whether the block will pass a target after a fixed push. The initial experiment may provide enough information to answer, or the model may need another measurement, such as the object's mass, friction, restitution, or spring stiffness. We study whether vision language models can decide when to answer immediately and, when more evidence is needed, which experiment to perform. Current physical reasoning benchmarks usually evaluate only the final answer, so they do not directly measure this decision-making ability. We introduce a controlled evaluation where each problem provides one measurement image and four possible physical worlds created by combining two possible masses and two possible values of another relevant property. The model must either stop and answer or select the cheapest additional experiment that can resolve the question. We construct matched problem pairs where changing either the observed measurement or the question changes the optimal action. Since all possible worlds and experiment costs are known, we can explicitly determine the optimal choice. Across six open models and 144 physical parameter sets, direct responses repeat the same action for 95.1% to 100% of image pairs even when the correct action changes. Brief reasoning improves action switching, but the best model makes both decisions correctly for only 5.9% of image pairs. Additional analysis reveals failures in measurement interpretation, physical reasoning, and response formatting. By evaluating evidence selection separately from final answers, our benchmark reveals limitations in physical reasoning that conventional answer accuracy can overlook.

We study K/V-cache interventions – transplanting a target-conditioned K/V trajectory into a source-persona generation – as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression – a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.

Verbalized confidence, long dismissed as overconfident, coarse, and prone to round-number clustering, is now the more robust soft-scoring mechanism for LLM-as-a-Judge on top-tier proprietary models. Across SummEval, AggreFact, and HelpSteer2, spanning up to 18 LLMs, we show that the standard advice to prefer log-probabilities no longer holds on post-2025 models, where verbalized confidence is the better signal. We call this a compatibility shift. On top of a standard verbalized-confidence baseline, we introduce two new ingredients: an overconfidence advisory and self-debate. Together they improve calibration, score-distribution spread, and robustness to task subjectivity. We further observe a generation effect: post-2025 models accommodate these two additions with little balanced-accuracy cost, whereas pre-2025 models pay a measurable penalty. Compared with logprob-based G-Eval, verbalized confidence is the more subjectivity-robust soft signal on GPT-family top-tier releases. The shift is invisible under accuracy-only reporting. Rather than defaulting to hard predictions, we recommend broader use of soft scoring in LLM-as-a-Judge. More broadly, verbalized confidence has moved from a weaker substitute for logprobs to a practical soft-scoring mechanism for contemporary LLM judges.

Multilingual large language models (mLLMs) contain a small fraction of feed-forward neurons that are sensitive to particular languages, commonly termed language-specific neurons. Existing work measures language specificity using the entropy of each neuron's language-wise probabilities of being active, where a neuron is considered active when its activation value is positive. However, this approach may not fully capture the multilingual nature of mLLMs, where language representations are distributional and mutually related. We propose Distribution-aware Language Neuron selection, which leverages pairwise relationships between per-language activation distributions over the full activation range, including negative values. Specifically, we quantify each neuron's language specificity by clustering languages using pairwise overlap coefficients between their activation distributions. Across two mLLMs and two held-out corpora, our identifier more effectively isolates language-specific causal effects, yielding up to 4.9$\times$ higher on-target language damage per neuron while preserving off-target language performance.

Recent multimodal sentiment analysis studies increasingly adopt text-centric fusion approaches to exploit the rich sentiment information inherent in the textual modality. However, these approaches often suffer from performance degradation during inference due to partially missing or noisy data in real-world scenarios, especially when sentiment-related cues are missing. To address this issue, we introduce a new completeness estimation approach that quantifies the degree of sentiment-relevant information preserved in incomplete data to guide the reconstruction of missing semantics. Furthermore, we propose a training strategy that stabilizes multi-task learning while jointly optimizing sentiment prediction and completeness estimation. Extensive experiments and in-depth analyses on three benchmark datasets demonstrate that the proposed approach enables more accurate semantic reconstruction, leading to more precise sentiment prediction.

Membership inference attacks (MIAs) try to determine whether a specific record was used to train a model, a privacy risk that matters in natural language processing (NLP), where training data can contain sensitive user text. This paper presents a controlled benchmark of membership inference vulnerability for text classification on the GLUE SST-2 sentiment dataset. A TF-IDF + Logistic Regression pipeline and a fine-tuned DistilBERT classifier are compared under a loss-threshold MIA, with utility measured by development accuracy and macro F1. DistilBERT reached 0.9466 accuracy and 0.9460 macro F1 against 0.8756 and 0.8727 for Logistic Regression, yet both models leaked membership signal (Attack AUC 0.5615 and 0.5800, respectively). Two mitigations were tested. Stronger regularization reduced leakage for Logistic Regression at a visible utility cost, whereas fine-tuning DistilBERT for 2 epochs instead of 3 reduced leakage with negligible accuracy loss. Lightweight training adjustments can improve the privacy-utility trade-off without complex defenses.

Turn-taking is a fundamental mechanism that governs when interlocutors speak and listen. Although Spoken Dialogue Systems (SDS) exploit a range of linguistic, acoustic, and non-verbal cues, they produce ill-timed responses in unscripted interaction. A central challenge is anticipating Transition Relevance Places (TRPs), or opportunities, not obligations, for a listener to take the floor. Human listeners do not wait for turn endings; as an utterance unfolds, they use expectations about its developing meaning to anticipate TRPs and decide whether to take the floor. We examine whether these evolving expectations can be modeled through semantic uncertainty – an LLM-derived measure of how strongly a turn so far constrains what may plausibly come next. To do so, we sample possible continuations of an ongoing turn and use changes in semantic dispersion to identify TRPs within turns. We evaluate this account on a dataset with TRP labels derived from real-time listener responses, rather than retrospective annotation. Our approach substantially outperforms prompt-based and fine-tuned text-only baselines, providing empirical support for the view that evolving semantic constraints inform perceived turn-taking opportunities in unscripted interaction.

Graph captions should help readers understand graph structure, rather than simply translate adjacency matrices into long textual edge lists. A useful graph caption abstracts connectivity into recognizable motifs, such as hubs, paths, cycles, cliques, and bridges, because these motifs provide compact structural units that are easier to read, compare, and recover. In this paper, we study motif-oriented graph captioning as a bidirectional graph-text translation task, where captions must both preserve enough topology for graph recovery and express the graph through concise motif-level descriptions. We show that direct prompting of GPT-5.1 often produces graph-recoverable captions by enumerating node-to-node connections, but these captions are verbose and can contain inconsistent motif interpretations. To address this gap, we introduce Structurally Speaking, a lightweight structured prompting protocol that guides translation between explicit connectivity and motif-level abstraction. Experiments on a synthetic motif-based dataset show that structured prompting produces shorter and more motif-consistent captions while maintaining comparable graph recovery. These results suggest that explicit topology-to-motif reasoning guidance can make LLM-generated graph captions more interpretable without model fine-tuning.

Auto-research agents have shown the potential to automate hypothesis generation, experiment execution, and iterative refinement. However, scaling this paradigm to industry-scale recommendation models introduces two challenges: (1) long feedback loops, where model training can take days, making serial iteration prohibitively slow and requiring parallel exploration across multiple research directions; and (2) system complexity, where large configurations, fragile infrastructure dependencies, and multi-day GPU jobs require robust and recoverable execution. We present Auto-RecSys, an autonomous research system for long-horizon experimentation on industry-scale recommendation models. Auto-RecSys addresses these challenges through three harness designs: (1) distributed asynchronous execution for running multiple experiments in parallel across servers, (2) centralized cross-server memory for persistent and recoverable execution across sessions and failures, and (3) cognitive-procedural separation, where natural-language skill files guide LLM reasoning while deterministic scripts enforce operational correctness. Auto-RecSys further employs a dual-loop self-evolving architecture: an Execution Evolution Loop in which model-specific playbooks accumulate operational knowledge by recording failed attempts and crystallizing successful pipelines, and an Idea Evolution Loop in which experimental outcomes inform subsequent ideation. Evaluated on recommendation models, Auto-RecSys significantly reduces the human time required per experiment cycle and improves execution reliability as its playbooks mature.

Structured phenotype extraction from radiology reports supports cohort construction, quality auditing, and clinical analytics, but practical deployment requires local inference and auditable predictions, while expert annotations remain scarce. Conventional labelers provide structured findings and assertion states but no supporting evidence, while API-hosted large language models may be unsuitable when clinical text cannot leave institutional infrastructure. We present CHESTPHENOT, a compact 0.5-3B language model that jointly extracts finding labels, three-class status (present/absent/uncertain), and verbatim supporting evidence spans. CHESTPHENOT is trained using hybrid CheXbert+72B silver supervision followed by supervised fine-tuning and lightweight GRPO refinement. Across three human-annotated gold sets spanning in-distribution, cross-taxonomy, and cross-institution evaluation, the 3B model remains below its CheXbert silver teacher in distribution but is competitive under distribution shift, significantly surpassing CheXbert on cross-institution detection (+2.0 F1). Task-specific training also enables the 3B model to match or exceed substantially larger prompted models on most detection and status comparisons. For evidence-grounded extraction, over 99% of final evidence spans are locatable in the source report, and the 3B model achieves 47.5 auditable-F1, outperforming Qwen2.5-7B one-shot prompting by 7.6 points and approaching Qwen2.5-72B. These results demonstrate that locally deployable models can provide competitive and directly auditable radiology-report extraction without relying on external inference APIs. Code and the full extraction/judge prompts will be made available at https://github.com/yukkai/ChestPheNoT.

Production text-to-SQL pipelines often end with an LLM-as-judge whose agreement with human annotators has never actually been measured. When we checked ours, the deployed gpt-4o-mini judge agreed with two-author gold at only Cohen's kappa = 0.04 on a disagreement-enriched set and 0.42 on a uniform-random spot-check, over-flagging 77.1% of the human-FAITHFUL cases in the enriched set. Most of its over-flags trace back to a single mechanism we call GRADE-HALLUCINATION. A self-hosted Qwen3.6-27B replacement (kappa = 0.72) lands in the same range as Claude Opus 4.7 (kappa = 0.71); the head-to-head is underpowered at n = 96, but for the deployment decision that hardly matters, since Qwen costs roughly 1/300 as much per call. Ensembling does not help for free. Pairing the weak judge with a stronger one degrades agreement, whereas three strong judges under unanimity routing reach kappa = 0.79 at 89.7% auto-coverage. Applied out-of-domain, the same audit recipe flags 25.5% of BIRD-financial's expert-authored gold SQLs as candidate gold-SQL issues under our annotation protocol. Code and pre-registration are at https://github.com/JamesL404/synca-audit.

In team collaboration scenarios, memory is heterogeneous and continually evolving. Team memories capture collective decisions, protocols, and current consensus, while individual memories preserve member-specific observations, execution traces, and intermediate progress. Existing memory-augmented systems typically retrieve from all stored memories as a flat pool, ranking them by semantic relevance, importance, or recency without modeling hierarchical structure or evolving validity. As a result, they often surface semantically relevant but outdated or conflicting memories, especially individual memories that no longer align with current team consensus, instead of prioritizing currently valid memories. This is particularly problematic when collaborative LLM agents answer user questions, since their responses should be grounded in valid memories. We propose HiCoMER, a framework for hierarchical collaborative memory management and validity-aware retrieval in LLM agents. HiCoMER first maintains the validity of team and individual memories and then retrieves memories that remain valid, rather than retrieving directly from all stored memories. It consists of three components: a Hierarchical Memory Conflict Updater, a Validity-Aware Memory Retriever, and a Memory-Grounded Answer Generator. To evaluate HiCoMER, we construct two new datasets for memory-grounded question answering in collaborative settings. Experiments on both datasets show that HiCoMER consistently outperforms strong baselines by reducing outdated retrieval, preserving current team consensus, and improving downstream QA quality.

In Transformer-based masked language models, attention is the primary mechanism for context mixing, but there are other ways to mix data across tokens. Recent attention-free mixers replace attention with fixed or hypernetwork-generated MLPs, alternating their dynamic, content-dependent weighting for computational simplicity. We build an alternative that gets the same property from a low-rank bottleneck autoencoder. We replace attention with a stack of autoencoder-based mixing modules, one operating over local neighborhoods, one over the full sequence, and one across attention heads, each compressing and reconstructing its input through a bottleneck, and its width is a hyperparameter rather than a training effect. In masked positions, we introduce an iterative refinement procedure that has two distinct steps. A pulling step that pulls an embedding representation toward a weighted average of its neighbors, and a correcting step that projects the result back to the learned manifold via an autoencoder. Our architecture achieves a significant portion of attention's performance at about $1.9 \times$ fewer FLOPs when pretrained on C4 and evaluated with parameter-matched BERT baselines. Our model equals parameter-matched BERT and TinyBERT baselines on the rarest-token frequency bucket using a frequency-aware training schedule that samples rare tokens more than uniformly for the masking tasks.

We present ScorePrompts, an interactive system in which users upload a score, receive natural-language descriptions of its musical structure, ask questions about specific passages, and inspect the corresponding analysis results in staff notation. Specialist MIR components first estimate harmony, tonality, cadences, formal boundaries, texture, and note-level roles, organizing their outputs at note, beat, measure, and piece levels. A schema-constrained language model converts these results into descriptions rather than inferring musical structure directly from raw MusicXML. For questions such as "What changes in measures 14-18?", a deterministic router selects the relevant measures and analytical levels and returns a concise response together with the underlying results and caveats. Verovio renders the score and links the returned information to cited measures and note-level attributes. The interface also exposes intermediate tables and disagreements between analytical levels. ScorePrompts is intended for exploratory score analysis and explanation, not score editing. The demo shows how existing analysis models, constrained language generation, Q&A retrieval, and notation-based visualization can provide natural-language access to symbolic music analysis while keeping intermediate results inspectable.

LLM search agents are often evaluated on final-answer accuracy, overlooking the process. Analyzing a search strategy requires understanding how credible evidence is retrieved to address question constraints. This valuable information is buried in raw search trajectories that are long and difficult to parse. We introduce SearchAtlas, a framework that converts search trajectories into structured graphs whose edges represent how evidence is propagated across the reasoning trace, from the query that retrieves it to the final answer. Our automated parsing pipeline achieves a mean edge F1 of 86.0% against human-annotated graphs and remains consistent across repeated runs. We analyze five search agents on three benchmarks, revealing systematic differences in search scale and evidence aggregation. SearchAtlas exposes fragmented answer support, question constraints that do not reach the answer, and unverified parametric knowledge entering the response. These process failures are strongly associated with incorrect answers, even more so than an LLM judge given either the raw trajectory or the ordered query list, suggesting that the constructed graphs provide useful interpretability. Moreover, an audit of cases in which process-diagnostic scores disagree with final-answer correctness shows that they capture information not reducible to answer accuracy.

Speech-based automatic detection of Alzheimer's disease (AD) provides a non-invasive and scalable approach to early cognitive screening. AD affects both lexical-semantic organization and speech production, including atypical pauses and word elongations. However, existing methods have yet to fully integrate these paralinguistic cues with linguistic content. We propose LLM-Anchored Paralinguistic Enrichment (LAPE), which enriches LLM-derived linguistic representations with paralinguistic cues through three coordinated innovations. The first is prosodic event textualization, which enables the LLM to model pauses and elongations jointly with lexical content by encoding them as explicit markers with bounded duration-aware repetition. The second is lexico-prosodic unitization and chunking, which preserves event identity and magnitude in both modalities by pooling only consecutive word units. The third is text-anchored paralinguistic fusion, which integrates local and utterance-level speech features by using NormGate to normalize and dynamically scale them relative to text. We evaluate LAPE on ADReSS and ADReSSo using participant-level cross-validation and leave-one-subject-out evaluation. LAPE achieves state-of-the-art performance across all four primary settings. Code will be released upon acceptance.