每天知道
你的领域
出了什么

As reasoning models emit chains of thought tens of thousands of tokens long, KV cache increasingly becomes a deployment bottleneck. Existing cache eviction methods rank tokens by attention weight, which is a noisy importance proxy in long reasoning traces, and prohibits the use of fused kernels in production inference by forcing the model to materialize the attention matrix. In this work, we instead score tokens with a metric we term the epiphany score: the change in the model's internal representation, read directly from the forward pass with no attention matrix and negligible extra state. Our resulting cache eviction method, EpiKV, requires no training, classifier, or custom kernel, and can be used directly in FlashAttention inference stacks unchanged — scaling to a 16x longer feasible context than attention-based scoring. upper-mid layers negatively) and remove a positional trend with a causal rolling z-score. At a 4096-token cache EpiKV reaches 72% on MATH-500, matching the strongest attention-based baseline (ThinKV 71%, H2O 67%); a lag-normalized KV variant reaches 37% on AIME-2024 at 8192 tokens against the best of them (33%), at up to 2.8x the speed.

Mixture-of-Experts (MoE) architectures scale large language models (LLMs) to hundreds of billions of parameters. Serving a single MoE model requires multiple GPUs operating in parallel, typically through tensor parallelism (TP) or expert parallelism (EP). The optimal choice depends on the number of in-flight requests: TP is faster at low concurrency, whereas EP wins at high concurrency. Production workloads cross this boundary continually: online serving sees bursty arrivals that subside into quiet periods, and reinforcement-learning rollouts begin as a high-concurrency burst that decays into a long tail of stragglers. Pinning either layout therefore forfeits performance when the workload crosses to the other side. We present Moebius, a serving system that switches between EP and TP at runtime without restarting the engine or dropping in-flight requests. Our key insight is that EP and TP are two layouts of one model, not two models: they compute the same function over byte-identical expert weights and KV cache, so a switch changes only which rank owns each slice. Moving those owner-changed slices is the sole irreducible cost, and modern high-bandwidth GPU interconnects make it fast enough to do between decode steps without draining in-flight requests. Moebius preserves each parallelism's runtime resident, and reshards the single copy of expert weights and KV cache at fixed addresses with fused GPU-to-GPU transfer kernels. On 8x H200 GPUs serving Qwen3-235B-A22B, Moebius matches the better static parallelism at every operating point, and beats it on RL rollouts by 1.16-1.25x across steps. Each switch completes in 215-434 ms, and Moebius holds both layouts resident with only 2.4% memory overhead.

Reasoning capability has advanced rapidly in large language models (LLMs), leading to an increasing size of key-value (KV) cache in both prefilling and decoding stages. Existing KV cache compression methods mainly rely on attention weights to estimate token importance. While attention effectively captures contextual relevance, it overlooks complementary information-theoretic signals related to predictive uncertainty and token informativeness. In this paper, we revisit token importance from a forward-looking perspective and introduce Forward Influence, a metric that measures how compressed tokens affect future contexts. Our analysis reveals that tokens selected by attention scores mainly influence nearby contexts, whereas tokens associated with high predictive uncertainty exhibit substantially stronger influence on distant future contexts. Based on the observation, we propose InfoKV, an entropy-aware KV cache compression framework that incorporates information-theoretic signals. It combines token-level predictive uncertainty with layer-wise representation evolution and integrates the resulting entropy scores with attention scores during reasoning. Experiments on long-context reasoning benchmarks with Llama-3.1, Llama-3.2, and DeepSeek-R1 demonstrate that InfoKV consistently outperforms existing attention-based KV compression methods in both long prefilling and decoding scenarios.

Autoregressive large language model (LLM) serving is increasingly limited by key-value (KV) cache movement rather than dense matrix multiplication. Modern paged-attention systems reduce KV-cache fragmentation and mature kernels such as FlashInfer provide highly optimized native-paged decode attention. However, the best single-kernel implementation is not always the best serving schedule: low-active long-context decode can under-utilize commodity GPUs, while mixed sequence lengths introduce a tension between many exact-length launches and coarse padded batches. We present PersistentKV, a native block-table decode attention engine and page-aware scheduling study for grouped-query attention (GQA). PersistentKV maps work by KV-head group, is designed to reuse K,V tiles across grouped query heads, supports native page tables, and adds a compact workqueue schedule that executes only non-empty row-KV-head-sequence-split tasks. On an RTX 3060 with FP16, page size 16, Hq=32, Hkv=8, d=128, and identical correctness tolerance against FlashInfer, a calibrated adaptive policy selects FlashInfer for small active batches, PersistentKV sequence splitting for B1 long-context steps, and PersistentKV workqueue scheduling for B8 long-context steps. With thresholds and split counts fixed on calibration traces, one held-out trace seed improves synchronized wall throughput by 1.063-1.265x on B8 bimodal, uniform, and Zipf-like workloads and by 1.399x on a B1 bucketed trace. On the B4 bimodal boundary case, the policy avoids the PersistentKV regression by selecting FlashInfer. These results identify a concrete systems niche for adaptive page-aware decode scheduling and show that work assignment, not only attention math, is a decisive serving-system variable.

Low-bit floating-point formats and semi-structured sparsity are increasingly supported by modern accelerators, yet combining them for LLM activation compression remains challenging: activations contain input-dependent outliers that dominate block scales in FP4 quantization, and directly applying N:M sparsity masks discards moderate values, coupling sparsification loss with quantization error. We introduce SharQ, a training-free inference method that bridges activation sparsity and FP4 quantization through an online sparse--dense decomposition. For each activation tensor, SharQ generates an input-adaptive N:M mask to extract an outlier-dominated sparse backbone, quantizes it to FP4, and defines a dense residual relative to the quantized sparse backbone rather than the unquantized sparse values. A sparse FP4 GEMM processes the backbone while a dense FP4 GEMM compensates for both mask-induced activation loss and sparse-path quantization error. The two paths share a single FP4 weight payload with path-specific scale views, and a fused preparation kernel absorbs mask generation, residual construction, and layer normalization into one operator. SharQ requires no calibration data, retraining, or model-specific tuning. Evaluated on Llama-3.1-8B, Qwen2.5-7B, Qwen3-30B-A3B, and Qwen3-VL-8B, SharQ recovers 43--63% of the NVFP4-to-FP16 accuracy gap across language and vision-language tasks, and generalizes across NVFP4, HiF4, and MXFP4 formats. On an RTX 5090, SharQ delivers 2.2--2.4$\times$ latency reduction over FP16 and 1.2--1.4$\times$ throughput improvement over FP8 in language model serving, and up to 1.58$\times$ speedup on Wan2.2-T2V-A14B video generation when combined with SageAttention. Our code is available at https://github.com/actypedef/SharQ.

Language-model representations provide structured, high-dimensional annotations of naturalistic language stimuli and can serve as informative neural predictors during comprehension. We analyzed locked derived data from Brain Treebank, MEG-MASC, and Podcast ECoG with eight frozen language models, blocked encoding models, and matched temporal, nuisance, and representation-capacity controls. Positive held-out prediction and gains over low-level baselines were widespread in source-level summaries. Across Brain Treebank and Podcast ECoG, 67 of 432 evaluable rows met a controlled predictive-only criterion, and model-side feature ablations changed prediction scores in most evaluable source rows. Brain-derived, timing-linked, acoustic, and implanted-signal controls confirmed component-level sensitivity of the analysis pipeline. These findings show that language-model-derived quantities can annotate neural activity during natural speech and text comprehension. Participant-level matched-control advantages were localized rather than uniform, response-profile and feature-specificity contrasts bounded representational or computational interpretations, and complete co-indexed integrated interpretation will require future jointly indexed coverage. Together, the analyses identify language-model features as useful neural predictors and separate predictive usefulness from claims about shared neural organization or language-processing computations.

Concolic testing combines concrete execution with symbolic constraint solving, but Python programs expose recurring limits. Library calls can cause symbolic variables to downgrade to concrete values. Regular expressions, checksums, parsers, and other semantic operations can be hard to solve, and exploration can plateau on already covered paths. We present ConcoLixir, a reactive LLM extension for Python concolic execution. The LLM acts as a discovery oracle, not a replacement for the solver or a correctness oracle. It generates initial seeds, proposes concrete inputs after solver failures, and targets uncovered code when coverage stalls. Each candidate is executed concolically, and only observed coverage and collected path constraints guide later exploration. Across synthetic, real-world, and library targets, ConcoLixir improves mean line coverage over the baseline concolic tester without an LLM oracle by 8.6, 15.1, and 17.0 percentage points. The gains are strongest near semantic barriers and library boundaries, and the full evaluation costs \$1.63 in API charges. These results show that bounded LLM discovery can complement symbolic reasoning without replacing it.

Large language models can serve as capable long-horizon agents, but their out-of-distribution (OOD) generalization remains weak. We identify a key source of this failure as task insensitivity: when faced with similar but distinct tasks, models might apply patterns learned during training and fail to solve the task at hand. We show that models often continue with actions aligned with the original task even when the instruction is semantically corrupted and cannot be directly answered. We further find that, when we replace the task description in a trained prompt with another similar but distinct task, the model may still output the same action. This behavior is accompanied by a consistent training-time attention drift away from task tokens and toward local observations, suggesting an optimization bias toward shortcuts. To mitigate this problem, we propose Task-Perturbed NLL Optimization, a lightweight contrastive regularizer that explicitly encourages action dependence on the task instruction. Extensive evaluations show that our intervention improves task sensitivity and OOD generalization while preserving more stable attention to task tokens.

Multimodal large language models (MLLMs) applied to Medical Visual Question Answering (VQA) tend to produce overconfident outputs regardless of actual correctness, and existing verbalized confidence calibration methods, developed primarily for text only LLMs, do not account for the multimodal nature of medical image understanding. This work proposes a training based framework that finetunes MLLMs to improve their calibration using a composite loss function combining a Brier style calibration term, an anchor regularizer that prevents confidence collapse toward extreme values, a contrastive image text alignment term, and a KL based model stabilization term. The alignment signal is derived from a $2 \times 2$ factorial perturbation design that crosses image presence with text integrity, probing the reliance of the model on visual modality input versus language priors. Finally, a top K KL divergence regularizer is used to protect the answering ability of the model during finetuning. Across three Medical VQA benchmarks and two architectures (MedGemma 4B IT and Qwen2 VL 7B Instruct), our method reduces calibration error by 60% or more, and improves discrimination by 26% or more, while preserving predictive accuracy. On average across benchmarks, the technique outperforms prompting based, sampling based, and training based approaches, and ablation experiments confirm that each component of the loss function is indeed necessary for improving the calibration. All code for the experiments is publicly available.

Efficient sampling of molecular systems at thermodynamic equilibrium is a hallmark challenge in statistical physics. This challenge has driven the development of Boltzmann Generators (BGs), which allow rapid generation of uncorrelated equilibrium samples by combining a generative model with exact likelihoods and an importance sampling correction. However, modern BGs predominantly rely on normalizing flows (NFs), which either suffer from limited expressivity due to strict invertibility constraints (discrete time) or computationally expensive likelihoods (continuous time). In this paper, we propose Autoregressive Boltzmann Generators (ArBG) — a novel autoregressive modelling framework — that overcomes these limitations by departing from the flow-based BG paradigm. ArBG circumvents the topological constraints of flows and enables sequential inference-time interventions, while offering enhanced scalability by leveraging architectures effective in Large Language Models. We empirically demonstrate that ArBG leads to significant improvements over flow-based models across all benchmarks, but particularly in larger peptide systems such as the 10-residue Chignolin. Furthermore, we introduce Robin, a 132 million parameter transferable model trained with the ArBG framework which improves over the previous state-of-the-art, reducing the zero-shot energy error, E-W$_2$, on 8-residue systems by over 60$%$. The code can be found at the following link: https://github.com/danyalrehman/autobg.

Scientific discovery is fundamentally an optimization problem, defined by a vast "state space" of theories and experiments, and an evaluation criterion based on quality, novelty, and validity. Large language models (LLMs) have enabled automated exploration of this space, but we argue that simultaneous modification of the evaluation criteria is equally important. Here, we propose formalizing research as meta-optimization, where the optimization objective itself is also being optimized. Our key contribution is "consensus objective aggregation," where LLM-generated objective functions are combined via correlation-weighted voting, yielding a stable, self-correcting evaluation criterion that evolves as understanding deepens. We apply this framework to algorithm discovery for 3-SAT problems based on digital MemComputing machines, reducing the baseline scaling with problem size $N$ from $\sim N^{2.51}$ to $\sim N^{1.33}$ and delivering a $\sim 67\times$ speedup on the largest instances tested. As a problem-agnostic framework, we hope this approach will considerably aid scientific discovery.

We present a novel approach for applying Large Language Models (LLMs) to threat assessment in the context of foreign peacekeeping missions. Building on the PINPOINT project and its use case, the EU Monitoring Mission in Georgia, we combine an interdisciplinary risk-model with OSINT-based media collection and LLM-supported threat extraction. The proposed workflow maps media contents to mission-relevant threats, extracts structured information and applies several additional LLM-based processing steps to improve relevance and grounding. An evaluation of threats extracted from media documents shows high agreement between automatically generated results and human judgment for core aspects such as threat and mission relevance. These results indicate that LLMs provide a promising approach to support analysts in the context of peacekeeping missions.

Translation cascades for reasoning translate the query from another language to English, reason in English, and translate the answer back to the original language. This is a competitive approach to multilingual reasoning, but structurally lossy, since each stage discards information later stages may need, including cues for cultural grounding, register, and disambiguation. We examine the benefits of a simple and training-free intervention: a context-aware translation cascade, which additionally provides the original question, the English translated question, and the reasoning trace to the context of the final translation module. We evaluate gains across nine multilingual benchmarks including various task types, three backbone models, and 285 high-, mid-, and low-resource languages, and demonstrate strong gains for open-ended generation across models and resource regimes. We show that the original language question carries most of the beneficial context. Our study emphasizes the need to better design information flow in machine translation cascades for mitigating error propagation, and provides a simple and actionable default strategy: preserve the original user question until the end of the pipeline.

Static verification tools can assure industrial scale software, but require significant human labor to write specifications. This is particularly true of static verifiers based on separation logic (SL verifiers), which excel at verifying heapmanipulating programs, but require many complex auxiliary specifications to reason about heap structure. Recent work applies large language models (LLMs) to generate code, tests, and proofs, including specifications for verifiers, but mostly targeting non-SL verifiers. To address this gap, this paper thoroughly evaluates how well LLMs perform when prompted to generate specifications for verifying 303 C functions with the SL verifier VeriFast. We explored eight prompting approaches, ten LLMs, and three input types in two stages. Quantitative and qualitative analyses are used to assess the LLM-generated code and specifications for functional behavior, verifiability and errors. The results show that LLMs preserve functional behavior in source code and specifications (both over 91%), but achieve modest verification success (31.4%). Using Gemini 2.5 Pro and providing formal contracts lead to higher success rates in our setting. Moreover, most errors (94%) come from LLMs' mistakes in the domainspecific knowledge of SL verifiers such as VeriFast. These findings provide guidance for optimizing LLM-generated specifications for SL verifiers.

Modern software development increasingly involves the use of large language models (LLMs) to generate code. Despite their rapid advancement, LLMs remain prone to errors and hallucinations, emphasizing the importance of careful code inspection. However, in practice, developers' trust in LLM-generated code and their willingness to review it thoroughly may differ from these recommendations. How developers actually behave when reviewing LLM-generated code remains largely unexplored. In this study, we conduct a Wizard-of-Oz experiment to examine how software engineers behave when code is explicitly labeled as LLM-generated during a code review task. We collect both behavioral data and participant feedback through eye-tracking and exit interviews. Combining Bayesian data analysis with qualitative analysis, we found that while the thoroughness of code review did not change for participants, they spent more time fixating on LLM-labelled code, indicating that the label itself influences attention. Practitioners also adapted their review strategy for LLM-labelled code by assessing the code based on specific criteria (e.g., logical correctness), or using the prompt to guide their review. These findings inform LLM-based tool design on labelling while incorporating the prompt as a software artifact. Our study reveals a gap between reviewers' intentions and actual reviewing behaviour, highlighting the need for software companies to revisit their AI policies (particularly regarding LLM-assisted development) to better support developers in reviewing LLM-generated code.

Security misconfigurations in Terraform Infrastructure-as-Code are a growing risk in cloud deployments, and large language models are increasingly used as automated repair agents. Existing evaluations often treat a repair as successful when the targeted static-analysis finding disappears, without checking planning validity, behavioral change, or security intent. This paper presents TerraProbe, a five-layer oracle framework for evaluating LLM-assisted Terraform security repair. We apply TerraProbe to 288 first-pass repairs generated by gemini-2.5-flash-lite, GPT-4o, and Claude 3.5 Sonnet across 68 real-world TerraDS modules and 28 controlled injected-defect modules. The results show that targeted Checkov removal overstates repair success. Although targeted removal reaches 83.3 percent for the primary model, full-scanner cleanliness drops to 10.4 percent, Terraform planning succeeds for 39.6 percent, and plan comparison is reachable for 38.5 percent. Human adjudication further shows that 71.4 percent of plan-compared real-world repairs are deceptive fixes that pass automated checks while leaving the underlying vulnerability in place. This pattern is statistically indistinguishable across the three models, with deceptive-fix rates from 57.1 percent to 71.4 percent and pairwise Fisher exact p-values above 0.10. The paper introduces a four-dimensional taxonomy of deceptive fixes, validated with Cohen kappa of 0.78 and Krippendorff alpha of 0.76. IAM permission analysis confirms that wildcard Resource grants persist in all nine CKV2 AWS 11 deceptive-fix cases. TerraProbe contributes an evaluation methodology, a replication package, and the Multi-Layer Oracle Evaluation framework for distinguishing intent-aligned security repairs from scanner-passing false successes.

Weight-space regularization methods such as Elastic Weight Consolidation (EWC) are the standard approach to catastrophic forgetting in continual learning. However, those methods tend to underperform when applied to large language models. We argue that such underperformance can be partly explained by the ``polysemantic'' nature of large language models: per-weight importance estimates utilized by EWC-style regularization are too coarse and cannot isolate the knowledge that needs protection. In this paper, we propose regularizing instead in the model's activation space, using pretrained Sparse Autoencoders (SAEs) as a monosemantic feature dictionary. From the perspective of constrained optimization, we derive a new loss function that uses the SAE feature dictionary to explicitly balance stability and plasticity, and show that EWC is a special case in the one-sided weight-space penalty setting. Unlike replay-based methods that store or revisit examples from earlier tasks, our method requires no previous-task data after mask construction: current-task data is used to compute a compact SAE feature mask, and only this mask is retained for later training. Further, since the feature space has significantly lower dimensionality than the parameter space, the proposed method is more memory efficient. On the TRACE and MedCL continual learning benchmarks, the method achieves the strongest result among approaches without introducing task-specific architectural components, also surpassing traditional weight-space regularization methods like EWC. Beyond performance comparisons, we provide empirical evidence for the polysemanticity thesis: task-relevant representations are linearly separable in the SAE feature basis but indistinguishable from chance in the weight basis, and weight-space protection is nearly non-selective at the concept level.

Language models (LMs) capture large amounts of factual knowledge applicable to a wide range of tasks, motivating the view of their parameters as a knowledge base. An important property of knowledge bases is that different queries for the same fact return consistent results, drawing on a single source of truth. We investigate whether LMs satisfy this property through behavioral and mechanistic analyses. Our results suggest that they encode knowledge in a task-specific manner. Behaviorally, facts acquired on one task frequently fail to co-emerge on others during training. Parameter localization experiments suggest a mechanistic explanation, revealing distinct parameter subsets underlying different tasks for the same fact. Finally, we show that chain-of-thought reasoning draws part of its effectiveness from engaging task-specific parameters beyond those tied to the evaluation task. Our findings suggest that what the model knows and how it is asked are intertwined in parameter space, undermining the "knowledge base" analogy and carrying implications for the reliability and controllability of factual knowledge in LMs.

Verifying the eligibility of securities as collateral is a key responsibility of the German Central Bank. However, manually verifying these assets against legal and financial criteria within lengthy, semi-structured, and often bilingual prospectuses is a resource-intensive task. While previous efforts utilized traditional Named Entity Recognition (NER) for information extraction, these methods can struggle with OCR noise, linguistic variance, and rigid span-based constraints, and the need for manually annotated training data for each relevant annotation type. In this paper, we present the first case study applying Large Language Models (LLMs) to the eligibility examination process, shifting the paradigm toward a generative Information Extraction pipeline. Our approach decomposes the task into extraction, normalization, and interpretation, allowing for greater flexibility in handling noisy text and interleaved German-English content. We further introduce a value-based evaluation methodology using LLM-as-a-judge, which offers a more semantic assessment than location-based metrics. Our results demonstrate that LLM-based systems achieve high precision (up to 91%) in document-level eligibility, exhibiting a conservative operating profile that minimizes false acceptance.

Multimodal large language models (MLLMs) have achieved strong multimodal reasoning capabilities, but their efficiency is limited by the large number of visual tokens, which introduces substantial computational overhead. Visual token pruning offers a natural solution, yet existing methods are imperfect: attention-based criteria tend to retain redundant tokens, while diversity-based criteria are often agnostic to user instructions. Even methods that combine multiple criteria still lack a principled formulation of the intrinsic objective of token pruning. In this paper, we revisit visual token pruning from a first-principles perspective and formulate it as constructing Token Optimal Preservation Sets. Through a top-down information-theoretic analysis, we identify three fundamental principles for effective token selection: Task Relevance, Information Coverage, and Semantic Diversity. Based on these principles, we propose TOPS, a training-free and model-agnostic pruning module that can be applied to various MLLMs. Extensive experiments on 7 MLLM backbones and 14 benchmarks demonstrate that TOPS outperforms prior methods under diverse pruning settings. Notably, on LLaVA-NeXT, TOPS removes 77.8% of visual tokens while preserving 100.0% and 100.6% performance on its 7B and 13B models, respectively, suggesting that pruning redundant visual tokens can sometimes mitigate hallucination and inspire future lightweight MLLM design.

Diffusion language models offer a promising alternative to autoregressive models due to their potential for parallel and iterative generation. However, existing approaches use a single network for both context representation and iterative denoising, forcing one model to serve both roles and limiting its capacity for either role. We propose TwoTower, a block-wise autoregressive diffusion model that decouples these roles into two towers: a frozen AR context tower that causally processes clean tokens, and a trainable diffusion denoiser tower with bidirectional block attention that refines noisy blocks via cross-attention to the context. Built on Nemotron-3-Nano-30B-A3B, an open-weight 30B hybrid Mamba-Transformer MoE model, and trained on approximately 2.1T tokens, Nemotron-TwoTower retains 98.7% of the autoregressive baseline's quality while offering 2.42X higher wall-clock generation throughput. We release the code and model weights at https://huggingface.co/collections/nvidia/nemotron-twotower.

The engineering of adaptive user interfaces has traditionally relied on either rule-based systems encoding designer intuitions about user needs or machine learning approaches requiring substantial historical data before achieving effective personalization. We present a technical architecture that leverages Large Language Models as behavioral synthesis engines to enable immediate adaptation from sparse, heterogeneous user signals. Our system integrates three distinct behavioral channels, i) explicit micro-feedback on individual interface elements, ii) spatial priority inferred from manual widget reorganization through drag-and-drop interaction, iii) and attentional investment measured through dwell time during hover events, within a structured prompt engineering framework that continuously regenerates dashboard layouts while maintaining explanatory coherence. The architecture addresses the technical challenge of translating low-level interaction patterns into high-level design decisions through a layered prompt construction methodology that separates temporal context determination, behavioral signal extraction, explicit preference enforcement, and user profile synthesis. The approach combines manually specified behavioral interpretations and temporal heuristics with LLM-mediated synthesis, enabling the reconciliation of multiple simultaneous signals that would be difficult to encode through explicit rules alone. We demonstrate the system through an instantiation in the personal health monitoring domain, including an analytical evaluation of adaptation behavior across multiple scenarios and a working implementation managing fourteen distinct health metrics across seven widget visualization modalities. The evaluation compares profile-driven initialization, multi-signal behavioral adaptation, and presents the resulting interfaces through representative post-adaptation screenshots.

Transcatheter Aortic Valve Replacement (TAVR) planning requires meticulous multimodal reasoning. However, adapting Multimodal Large Language Models (MLLMs) to this high-stakes domain is severely impeded by diagnostic hallucinations, where generated text lacks anatomical grounding. To address this, TAVR-VLM is introduced: a novel framework featuring Risk-Conditioned Causal Grounding Attention (R-CGA) that instantiates a model-internal ``Risk $\rightarrow$ Region $\rightarrow$ Word'' structural grounding pathway. R-CGA compresses multimodal inputs into a causal risk bottleneck, purifying dense visual features into a global risk mask. During autoregressive generation, a support-projected causal consistency objective constrains token-level grounding within the risk-defined support mask. Evaluated on $\text{M}^3\text{TAVR}$, a comprehensive 1,482-patient cohort, TAVR-VLM establishes a new state-of-the-art. It achieves an AUROC of 0.896, boosts CIDEr to 0.936, and drastically reduces the hallucination rate to 8.1%, thereby improving interpretability for evidence-based surgical AI.

Large language models (LLMs) are increasingly being integrated into mental health support tools and other psychologically sensitive conversational applications. In such settings, behavioral stability and consistency are important for trustworthy human-AI interaction. However, semantically similar concerns can be presented through different contextual framings, potentially eliciting different model responses. Such framing-sensitive variability may challenge user expectations regarding system behavior and complicate the assessment of AI reliability. While prior studies have primarily examined such effects at the behavioral level, less is known about how framing-related variation is reflected in the internal representations of aligned language models. In this work, we investigate these effects using controlled matched prompts spanning multiple contextual framing conditions across several instruction-tuned model families. Across architectures, framing systematically alters interpretive response tendencies. Layer-wise probing analyses show that behavior-associated information remains decodable throughout transformer depth, with architecture-dependent variation in decoding strength. Moreover, held-out framing probes remained consistently above chance across architectures despite strong lexical baselines. Activation steering experiments further suggest that framing-associated representational directions can partially modulate downstream behavioral outcomes. Finally, these findings indicate that robustness to contextual variation may represent an important consideration when evaluating the consistency and trustworthiness of conversational AI systems deployed in mental-health-oriented interactions.

Large language models (LLMs) are increasingly used to screen and rank job applicants, creating incentives for candidates to strategically manipulate algorithmic hiring systems. We study prompt injection in automated résumé screening, defined as subtle self-promotional text that introduces no new qualifications but is designed to influence LLM evaluations. Using controlled experiments, we show that prompt injection reliably improves applicant rankings when résumé quality is homogeneous and few candidates inject. However, its effectiveness rapidly diminishes as more candidates inject, collapsing when manipulation becomes widespread. When candidate quality is heterogeneous, prompt injection is less effective on average, but can occasionally allow lower-quality candidates to outrank higher-quality ones, raising fairness concerns. Overall, LLM-based screening is most vulnerable when manipulation is rare and candidate quality differences are small. Code and resources are publicly available at: https://github.com/preetb1199/Prompt_Injection_ACL26

Supervised fine-tuning (SFT) on a small, high-quality set of long reasoning traces is an effective approach for eliciting strong reasoning capabilities in Large Language Models (LLMs). However, existing methods for curating high-quality SFT data rely heavily on strong reasoning models to filter examples based on diversity and difficulty, making the curation process costly while often yielding suboptimal data quality. In this work, we show that diverse and challenging reasoning examples can be identified using only the initial reasoning tokens. Specifically, we demonstrate that difficult problems can be reliably detected based on the loss of the first 100 reasoning tokens evaluated at a randomly perturbed checkpoint of the pretrained model. We further show that examples exhibiting similar loss patterns over their first 1k reasoning tokens across a small number of perturbed checkpoints extrapolating along the fine-tuning trajectory provably induce similar gradients. We validate our approach through extensive experiments on fine-tuning Qwen2.5-7B and Llama3.1-8B models on the M23K medical reasoning and OpenThoughts-Math datasets. Our method outperforms existing baselines by up to 1.7% while being 91% more token efficient.

Synthesizing human motion from textual descriptions is essential for immersive digital applications, yet existing methods face a persistent trade-off between semantic fidelity and physical realism. Large language model (LLM)-based approaches can interpret diverse open-vocabulary instructions and compose high-level action plans, but they often generate motions that violate physical constraints. Physics-aware models improve realism through simulation or control, but they struggle with semantic complexity, fine-grained instructions, and novel concepts. To address this gap, we propose In-Context Model Predictive Generation (ICMPG), a framework that integrates language-model planning with inference-time physical feedback. ICMPG reformulates motion synthesis as a Model Predictive Control (MPC)-like process with two modules. The Context-Aware Motion Generation (CAMG) module uses an LLM as a planner to decompose textual commands and generate candidate motion sequences from motion tokens. The Model Predictive Generation (MPG) module evaluates these candidates through physical simulation and semantic alignment, estimates a composite reward, and selects the best sequence to guide subsequent generation steps. Unlike open-loop generation, this closed-loop refinement enables ICMPG to adapt motions to both the input semantics and the simulated physical environment without task-specific policy retraining. Extensive experiments across standard and zero-shot open-vocabulary settings show that ICMPG generalizes robustly to diverse commands and produces motions that are more physically plausible and semantically faithful than representative baselines on the evaluated benchmarks. The framework bridges semantic interpretation and physical simulation while remaining flexible enough to incorporate different LLM backbones, enabling more versatile and controllable text-driven motion synthesis.

Chain-of-thought (CoT) reasoning is widely used in language-model agents, but prior work has shown that verbalized CoT is not always faithful and may instead reflect post-hoc reasoning, which means the model already knows the answer before reasoning. We therefore ask what CoT training is actually improving: is the model getting better at changing its action through generated reasoning, or is it getting better at predicting the action directly from the prompt? We study this question by comparing prompt actions (predicting action without CoT) with CoT actions (predicting action with CoT). Across checkpoints, prompt-action quality improves substantially. While interacting with the environment, the relative advantage of CoT actions over prompt actions remains similar, showing that CoT training does not widen the advantage of CoT reasoning, and it helps to improve the quality of prompt actions. We further find that later checkpoints are less likely to revise the action in response to CoT, suggesting greater reliance on the prompt. Motivated by these patterns, we selectively mask action-token supervision on a fraction of training examples. This intervention improves out-of-domain generalization.

Enhancing the analysis of service feedback is essential for public sector organizations, particularly tax administrations, where trust and compliance depend on fair and effective service delivery. As feedback volumes grow, identifying emerging service quality issues and potential disparities across diverse populations becomes increasingly challenging. Traditional approaches often rely on manual review or static expert-defined indicators, limiting scalability and the ability to capture complex patterns in textual feedback. This paper presents a novel methodology that integrates large language models (LLMs), statistical techniques, and human-AI collaboration to improve multilingual customer feedback analysis. The primary objective is to detect emerging service quality topics that may also reveal potential inequities in service delivery. Our framework combines fine-tuned, quantized LLMs with expert oversight to produce accurate, computationally efficient, and context-aware analyses. The proposed approach was evaluated using similarity analysis and assessments from experienced tax officers, demonstrating stronger alignment with expert judgments than baseline models. By incorporating a human-in-the-loop framework, the methodology reduces LLM fabrication while improving the reliability and relevance of generated insights. The results demonstrate the practicality of combining LLMs with human expertise to support scalable, evidence-based decision-making in public sector organizations. This work contributes to the development of responsible AI systems that enhance service quality, responsiveness, fairness, and public trust through more effective analysis of multilingual customer feedback.

Multilingual large language models often produce inconsistent reasoning and answers for semantically equivalent prompts in different languages. Prior work suggests that intermediate representations can be relatively language-agnostic, but generation becomes increasingly language-specific as models commit to discrete output tokens. This is problematic because language-specific lexical choices can cause semantically equivalent reasoning paths to diverge across languages. These divergences motivate searching for a cross-lingual alignment signal that is less tied to any single vocabulary item or script. We propose SOLAR, an auxiliary objective for supervised fine-tuning that aligns soft-token representations across languages, using English as a pivot. Soft tokens are probability-weighted mixtures over the vocabulary embeddings, yielding continuous representations that can aggregate information from semantically related tokens across languages. We then align each non-English soft-token summary to its English counterpart in the shared embedding space. Across four multilingual reasoning benchmarks, SOLAR improves accuracy by up to +17.7 points over the base model and +3.8 over standard supervised fine-tuning, with the largest gains on low-resource languages. SOLAR also strengthens final-layer cross-lingual similarity and substantially reduces language-cluster separability, suggesting that aligning soft-token representations helps preserve shared semantic structure during multilingual reasoning.