Papers

How do artificial neural networks bind concepts to form complex semantic structures? Here, we propose a simple neural code, whereby the existence and the type of relations between entities are represented by the distance and the direction between their embeddings, respectively. We test this hypothesis in a variety of Large Language Models (LLMs), each input with natural-language descriptions of minimalist tasks from five different domains: arithmetic, visual scenes, family trees, metro maps and social interactions. Results show that the true semantic structures can be linearly recovered with a Polar Probe targeting a subspace of LLMs' layer activations. Second, this code emerges mostly in middle layers and improves with LLM performance. Third, these Polar Probes successfully generalize to new entities and relation types, but degrades with the size of the semantic structure. Finally, the quality of the polar representation correlates with the LLM's ability to answer questions about the semantic structure. Together, these findings suggest that LLMs learn to build complex semantic structures by binding representations with a simple geometrical principle.

Geospatial foundation models compress multispectral observations into dense embeddings increasingly used in natural-language environmental reasoning systems. A single planetary-scale model, e.g. Google AlphaEarth, handles broad characterization well but may compromise on specialized hydrologic signals. Such generalist models are also often inaccessible, expensive, and require large-scale compute. We propose Mini-JEPAs: a fleet of small sensor-specialized Joint Embedding Predictive Architecture (JEPA) foundation models consulted by a routing agent for specialized questions. We pretrained five 22M-parameter Mini-JEPAs sharing an identical Vision Transformer backbone, JEPA recipe, and 64-d output space, using Sentinel-2 optical, Sentinel-1 SAR, MODIS thermal, multi-temporal Sentinel-2 phenology, and a topography-soil stack. Each Mini-JEPA reconstructs the variable matched to its sensor, with cross-validated $R^2$ reaching 0.97 for elevation, 0.97 for temperature, and 0.81 for precipitation. The five manifolds differ in geometric structure, with global participation ratios from 8.9 to 20.2 and local intrinsic dimensionalities from 2.3 to 9.0. Joint topography-soil and phenology models add predictive value beyond AlphaEarth alone for soil moisture, aridity, and precipitation ($\Delta R^2$ up to 0.031). A router LLM reads per-modality references and selects appropriate sensors with a perfect hit rate over a curated question set. In paired LLM-as-Judge evaluation, dual retrieval over AlphaEarth and the routed fleet outperforms AlphaEarth alone on physics-matched questions (Cohen's $d = 1.10$, $p = 0.031$). Locally-trained Mini-JEPAs can be operationalized for hydrologic intelligence with modest compute.

An AI system for professional floor plan design must precisely control room dimensions and areas while respecting the desired connectivity between rooms and maintaining functional and aesthetic quality. Existing generative approaches focus primarily on respecting the requested connectivity between rooms, but do not support generating floor plans that respect numerical constraints. We introduce a text-based floor plan generation approach that fine-tunes a large language model (LLM) on real plans and then applies reinforcement learning with verifiable rewards (RLVR) to improve adherence to topological and numerical constraints while discouraging invalid or overlapping outputs. Furthermore, we design a set of constraint adherence metrics to systematically measure how generated floor plans align with user-defined constraints. Our model generates floor plans that satisfy user-defined connectivity and numerical constraints and outperforms existing methods on Realism, Compatibility, and Diversity metrics. Across all tasks, our approach achieves at least a 94% relative reduction in Compatibility compared with existing methods. Our results demonstrate that LLMs can effectively handle constraints in this setting, suggesting broader applications for text-based generative modeling.

Biomedical retrieval-augmented large language models (LLMs) often face evidence that is incomplete, misleading, or internally contradictory, yet evaluation usually emphasizes answer accuracy under helpful context rather than reliability under conflict. Using HealthContradict, we evaluate six open-weight LLMs under five controlled evidence conditions: no retrieved context, correct-only context, incorrect-only context, and two mixed conditions containing both correct and contradictory documents in opposite orders. In this conflicting-evidence order contrast, where the same two documents are both present and only their order is reversed, accuracy drops for every model and 11.4%–25.2% of predictions flip. To support abstention in these difficult cases, we also evaluate a conflict-aware abstention score that combines model confidence with a detector of evidence conflict. In the two hardest conditions, this score improves selective accuracy over confidence-only, with mean gains of 7.2–33.4 points in incorrect-only (`IC') and 3.6–14.4 points in incorrect-first conflicting (`ICC') conditions across 75%, 50%, and 25% coverage. These results show that conflicting biomedical evidence is both an uncertainty and robustness problem and motivate evaluation and abstention methods that explicitly account for evidence disagreement.

Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths. In this setting, correctness is no longer tied to a single reasoning trace but to the aggregation rule over a pool of candidate paths, making aggregation uncertainty the central challenge. This issue is critical where confidently incorrect answers are far more costly than abstentions. We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty. Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control. This approach leads to finite-sample guarantees on the confident-error rate–the probability that the system answers and is wrong. We further identify score separability as the key condition under which abstention provably improves selective accuracy, and derive closed-form expressions that predict accuracy gains from calibration data alone. The method is fully inference-time, and requires no retraining. Across four benchmarks, four open-source models, and three score classes, realized confident-error rates are consistent with the prescribed targets up to calibration-split and test-set variability. Our method achieves $90.1\%$ selective accuracy on GSM8K by abstaining on less than $5\%$ of problems, compared with $82\%$ accuracy under majority-voting baseline.

Large Language Models (LLMs) trained on web-scale corpora inherently absorb toxic patterns from their training data. This leads to toxic degeneration where even innocuous prompts can trigger harmful outputs. This phenomenon poses significant risks for real-world deployments. Thus, necessitating effective mitigation strategies that should maintain model utility while ensuring safety. In this comprehensive replication study, we evaluate the efficacy of DExperts (Decoding-time Experts), which is an inference-time mitigation technique that steers generation without requiring model retraining. We structured our research into three systematic phases: (1) establishing baseline toxicity measurements using RealToxicityPrompts on standard GPT-2 models; then (2) implementing and evaluating DExperts to mitigate explicit toxicity; and finally (3) stress-testing the method against implicit hate speech using the adversarial ToxiGen dataset. Our empirical results confirm that while DExperts achieves near-perfect safety rates (100%) on explicit toxicity benchmarks, it exhibits brittleness against adversarial, implicit hate speech, with safety rates dropping to 98.5%. Furthermore, we quantify a critical trade-off. The method introduces a 10x latency penalty (from 0.2s to 2.0s per generation), posing challenges for real-time deployment scenarios. This study contributes to the growing body of work on AI safety by highlighting the robustness gap between explicit and implicit toxicity mitigation. We emphasize the need for more sophisticated approaches that generalize across diverse hate-speech patterns without incurring prohibitive computational costs.

Code agents must both reason over long-horizon repository state and obey strict tool-use protocols. In paired Instruct/Thinking checkpoints, these capabilities are complementary but misaligned. The Instruct model is concise and tool-disciplined, whereas the Thinking model offers stronger planning and recovery behavior but often over-deliberates and degrades agent performance. We present CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing), a training-free parameter-editing method that treats the Thinking-Instruct delta as a directional pool of candidate reasoning edits for the Instruct backbone. CRANE combines magnitude thresholding to denoise the delta, a Conservative Taylor Gate to retain edits that are jointly beneficial for reasoning transfer and tool-use preservation, and Graduated Sigmoidal Projection to suppress format-critical update directions. By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass1/pass5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.

Large language models (LLMs) have revolutionized natural language processing. Understanding their internal mechanisms is crucial for developing more interpretable and optimized architectures. Mechanistic interpretability has led to the development of various methods for assessing layer relevance, with cosine similarity being a widely used tool in the field. On this work, we demonstrate that cosine similarity is a poor proxy for the actual performance degradation caused by layer removal. Our theoretical analysis shows that a layer can exhibit an arbitrarily low cosine similarity score while still being crucial to the model's performance. On the other hand, empirical evidence from a range of LLMs confirms that the correlation between cosine similarity and actual performance degradation is often weak or moderate, leading to misleading interpretations of a transformer's internal mechanisms. We propose a more robust metric for assessing layer relevance: the actual drop in model accuracy resulting from the removal of a layer. Even though it is a computationally costly metric, this approach offers a more accurate picture of layer importance, allowing for more informed pruning strategies and lightweight models. Our findings have significant implications for the development of interpretable LLMs and highlight the need to move beyond cosine similarity in assessing layer relevance.

Distilling reasoning traces from strong large language models into smaller ones is a promising route to improve intelligence in resource-constrained settings. Existing approaches face a fundamental trade-off: offline distillation from teacher-generated traces provides high-quality, sample-efficient supervision but suffers from distributional drift: during training, the student model conditions on teacher-generated prefixes, whereas during inference the student autoregresses on self-generated prefixes, leading to compounding errors over long reasoning trajectories. Meanwhile, on-policy or self-distillation methods better match the student's inference-time distribution, but require costly online sampling and often produce low-quality traces in early training. We propose a principled offline reasoning distillation framework that preserves the efficiency and supervision quality of offline teacher-generated data while correcting teacher-student distribution drift. It adaptively emphasizes teacher supervision that is better aligned with the student's on-policy distribution. Evaluations on mathematical reasoning benchmarks of GSM8K, MATH, MATH500, and harder held-out competition-style tasks, including AMC, AIME, and OlympiadBench, show that our method improves reasoning accuracy over prior offline distillation algorithms and yields more stable reasoning traces while preserving instruction-following capabilities. Our work shows that lightweight, distribution-correction-aware training can substantially strengthen offline reasoning distillation without online rollouts.

Early-stage Parkinson's disease (EarlyPD) detection from speech is clinically meaningful yet underexplored, and published results are hard to compare because studies differ in datasets, languages, tasks, evaluation protocols, and EarlyPD definitions. To address this issue, we propose the first benchmark for speech-based EarlyPD detection, with a speaker-independent split designed for fair and replicable cross-method evaluation on researcher-accessible datasets. The benchmark covers three common speech tasks and evaluates methods under different training-resource settings. We also present multi-dimensional evaluation breakdowns by dataset, aggregation level, gender, and disease stage to support fine-grained comparisons and clinical adoption. Our results provide a replicable reference and actionable insights, encouraging the adoption of this publicly available benchmark to advance robust and clinically meaningful speech-based EarlyPD detection.

While synthetic data generation with large language models (LLMs) is widely used in post-training pipelines, existing approaches typically generate full outputs before applying quality filters, leading to substantial token waste on samples that are ultimately discarded. To address this, we propose Multi-Stage In-Flight Rejection (MSIFR), a lightweight, training-free framework that detects and terminates low-quality generation trajectories at intermediate checkpoints before they reach full completion. MSIFR decomposes the generation process into sequential stages and applies fast rule-based validators to identify arithmetic inconsistencies, hallucination patterns, and formatting violations, enabling early rejection of faulty samples. We formalize in-flight rejection as a sequential decision process and show that any non-trivial discard policy reduces expected token consumption, with stage-wise savings increasing when rejection occurs earlier in the generation pipeline. We further demonstrate that conditional utility estimates form a martingale, ensuring that early, in-flight rejection does not bias the expected utility of retained samples. Across five instruction-tuned models and seven reasoning benchmarks, MSIFR reduces token consumption by 11%-77% as a standalone method, and up to 78.2% when combined with early-exit methods, while preserving or improving evaluation accuracy. These results confirm that MSIFR provides a practical mechanism for improving the efficiency of LLM-based synthetic data generation without additional training or architectural changes.

Most existing dialogue systems are user-driven, primarily designed to fulfill user requests. However, in many critical real-world scenarios, a conversational agent must proactively extract information to achieve its own objectives rather than merely respond. To address this gap, we introduce Inquisitive Conversational Agents (ICAs) and develop an ICA specifically tailored to U.S. Supreme Court oral arguments. We propose a Dual Hierarchical Reinforcement Learning framework featuring two cooperating RL agents, each with its own policy, to coordinate strategic dialogue management and fine-grained utterance generation. By learning when and how to ask probing questions, the agent emulates judicial questioning patterns and systematically uncovers crucial information to fulfill its legal objectives. Evaluations on a U.S. Supreme Court dataset show that our method outperforms various baselines across multiple metrics. It represents an important first step toward broader high-stakes, domain-specific applications.

Parameter-Efficient Fine-Tuning (PEFT) is widely used for adapting Large Language Models (LLMs) for various tasks. Recently, there has been an increasing demand for fine-tuning a single LLM for multiple tasks because it requires overall less data for fine-tuning thanks to the common features shared among tasks. More importantly, LLMs are resource demanding and deploying a single model for multiple tasks facilitates resource consolidation and consumes significantly less resources compared to deploying individual large model for each task. Existing PEFT methods like LoRA and Prefix Tuning are designed to adapt LLMs to a specific task. LoRA and its variation focus on aligning the model itself for tasks, overlooking the importance of prompt tuning in multi-task learning while Prefix Tuning only adopts a simple architecture to optimize prompts, which limits the adaption capabilities for multi-task. To enable efficient fine-tuning for multi-task learning, it is important to co-optimize prompt optimization and model adaptation. In this work, we propose a Parameter-Efficient Multi-task Learning (\PM), which employs a neural architecture engineering method for optimizing the continuous prompts while also performing low-rank adaption for model weights. We prototype PEML by creating an automated framework for optimizing the continuous prompts and adapting model weights. We evaluate PEML against state-of-the-arts multi-task learning methods MTL-LoRA, MultiLoRa, C-Poly, and MoE, on the GLUE, SuperGLUE, Massive Multitask Language Understanding, and commonsense reasoning benchmarks. The evaluation results present an average accuracy improvement of up to 6.67%, with individual tasks showing peak gains of up to 10.75%.

The application of Large Language Models to Question Answering has shown great promise, but important challenges such as hallucinations and erroneous reasoning arise when using these models, particularly in knowledge-intensive, domain-specific tasks. To address these issues, we introduce Derivation Prompting, a novel prompting technique for the generation step of the Retrieval-Augmented Generation framework. Inspired by logic derivations, this method involves deriving conclusions from initial hypotheses through the systematic application of predefined rules. It constructs a derivation tree that is interpretable and adds control over the generation process. We applied this method in a specific case study, significantly reducing unacceptable answers compared to traditional RAG and long-context window methods.

The growing adoption of large language models in legal practice brings both significant promise and serious risk. Legal professionals stand to benefit from AI that can reason over contracts, draft documents, and analyze sources at scale, yet the high-stakes nature of legal work demands a level of rigor that current AI systems do not provide. The central problem is not simply that LLMs hallucinate facts and references; it is that they systematically draw inferences that go beyond what the source text actually supports, presenting assumption-laden conclusions as if they were logically grounded. This proposal presents a neuro-symbolic approach to legal AI that combines the expressive power of large language models with the rigor of formal verification, aiming to make AI-assisted legal reasoning both capable and trustworthy, thus reducing the burden of manual verification without sacrificing the accountability that legal practice demands.

Trackable improvement in multimodal physics reasoning rests on a training-and-evaluation system that is itself rarely verified: the corpora a model trains on, the reward it is optimized against, and the benchmarks and judges that score it. We audit this system end to end and find that standard construction practices systematically distort measurement: contamination slips past n-gram deduplication, translation degrades problems, saturated multiple-choice formats overstate capability, partial-credit training rewards are easier to exploit than to earn, and open-ended grading silently depends on the choice of judge. Left unverified, these distortions inflate reported progress and leak test knowledge into training. We answer with a released verifier system: a three-stage contamination audit that certifies the train/test boundary behind an audited multimodal training corpus and a held-out olympiad benchmark; a binary answer verifier that supplies the reinforcement-learning training reward; and an answer-judging harness that brackets every open-ended score between a deterministic strict layer and a large-language-model liberal layer. All per-record verdicts are released and cross-checked against an independent open-weight judge, whose substitution shifts absolute scores but preserves the sign of every base-to-trained lift. Training against the binary answer verifier confirms the certified corpus supports training: a reference recipe lifts an 8B open-source base by 18.3 points on the held-out benchmark across three seeds. The simple binary reward also beats a dense partial-credit variant on three of four open-ended benchmarks, tying the fourth. All verifiers, verdicts, and datasets are public.

Under modern test-time compute and agentic paradigms, language models process ever-longer sequences. Efficient text generation with transformer architectures is increasingly constrained by the Key-Value cache memory footprint and bandwidth. To address this limitation, we introduce Self-Pruned Key-Value Attention (SP-KV), a mechanism designed to predict future KV utility in order to reduce the size of the long-term KV cache. This strategy operates at a fine granularity: a lightweight utility predictor scores each key-value pair, and while recent KVs are always available via a local window, older pairs are written in the cache and used in global attention only if their predicted utility surpasses a given threshold. The LLM and the utility predictor are trained jointly end-to-end exclusively through next-token prediction loss, and are adapted from pretrained LLM checkpoints. Rather than enforcing a fixed compression ratio, SP-KV performs dynamic sparsification: the mechanism adapts to the input and typically reduces the KV cache size by a factor of $3$ to $10\times$, longer sequences often being more compressible. This leads to vast improvements in memory usage and decoding speed, with little to no degradation of validation loss nor performance on a broad set of downstream tasks. Beyond serving as an effective KV-cache reduction mechanism, our method reveals structured layer- and head-specific sparsity patterns that we can use to guide the design of hybrid local-global attention architectures.

In current Large Language Models we can trust the production of smoothly flowing prose on the basis of the principles of machine learning. However, there is no comparably principled basis to justify trust in the content of the text produced. It appears to be conventional wisdom that addressing this issue by adding more principled reasoning is not computationally affordable. Here we propose a principled method of reasoning that is efficient enough to be practical for large language models. Further, the method allows the retention of much of the currently used software and hardware base. Our method for improving the functioning of large language models consists of a first stage of preprocessing that recodes the data to a Unary Relational Integracode that is more explicit about the relationships among the objects described in the text, followed as a second stage by a standard but possibly streamlined machine learning process that then also learns to predict these relationships. The method may be viewed as realizing a world model and applying beyond natural language, to vision and actions, for example, where the multiple properties of an object referred to in an input are brought together explicitly, rather than remaining distributed in the various references to it in the input. We articulate its advantages in terms of Robust Logic, a system for performing principled chaining on learned, and hence uncertain, information. We show that this recoding has the surprising and fortuitous property that, while succinct, it makes the task of learning a core subset of relational rules that hold in the world described in the training data polynomial time learnable in a defined sense, the polynomial depending on the complexity of the rule. This gives support for sound reasoning within each single call of the learned classifier as well as between multiple calls.

Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, we propose a novel value-based framework that employs GraphRAG to convert principles into value-based instructions and steer the agent to behave as expected by retrieving the suitable instruction upon a specific conversation context. To evaluate the ratio of expected behaviors, we define the expected behaviors from two famous theories, Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. By experimenting with our method on the benchmark of DAILYDILEMMAS, our method exhibits significant performance gains compared to prompt-based baselines, including ECoT, Plan-and-Solve, and Metacognitive prompting. Our method provides a basis for the emergence of self-emotion in AI systems.

Speculative decoding has become a widely adopted technique for accelerating large language model (LLM) inference by drafting multiple candidate tokens and verifying them with a target model in parallel. Its efficiency, however, critically depends on the average accepted length $\tau$, i.e., how many draft tokens survive each verification step. In this work, we identify a new mechanism-level vulnerability in model-based speculative decoding: the drafter is trained to approximate the target model distribution, but this approximation is inevitably imperfect. Such a drafter-target mismatch creates a hidden attack surface where small perturbations can preserve the target model's visible behavior while substantially reducing draft-token acceptability. We propose Mistletoe, a stealthy acceleration-collapse attack against speculative decoding. Mistletoe directly targets the acceptance mechanism of speculative decoding. It jointly optimizes a degradation objective that decreases drafter-target agreement and a semantic-preservation objective that constrains the target model's output distribution. To resolve the conflict between these objectives, we introduce a null-space projection mechanism, where degradation gradients are projected away from the local semantic-preserving direction, suppressing draft acceptance while minimizing semantic drift. Experiments on various speculative decoding systems show that Mistletoe substantially reduces average accepted length $\tau$, collapses speedup, and lowers averaged token throughput, while preserving output quality and perplexity. Our work highlights that speculative decoding introduces a mechanism-level attack surface beyond existing output robustness, calling for more robust designs of LLM acceleration systems.

Sparse Mixture-of-Experts (MoE) layers route tokens through a handful of experts, and learning-free compression of these layers reduces inference cost without retraining. A subtle obstruction blocks every existing compressor in this family: three experts can each be pairwise compatible yet form an irreducible cycle when merged together, so any score that ranks experts on pairwise signals is structurally blind to which triples are jointly mergeable. We show the obstruction is a precise mathematical object, the harmonic kernel of the simplicial Laplacian on a 2-complex whose vertices are experts, whose edges carry KL merge barriers, and whose faces carry triplet barriers; Hodge-decomposing the edge-barrier signal isolates the kernel exactly. We turn the diagnostic into a selection objective: HodgeCover greedily covers the harmonic-critical edges and triplet-critical triangles, and a hybrid variant of HodgeCover pairs it with off-the-shelf weight pruning on survivors. On three open-weight Sparse MoE backbones under aggressive expert reduction, HodgeCover matches state-of-the-art learning-free baselines on the expert-reduction axis, leads on the aggressive-compression frontier of the hybrid axis, and uniquely balances retained mass across all four Hodge components. These results show that exposing the harmonic kernel of a learned MoE structure changes which compressor wins at the regime that matters most.

We present VectraYX-Nano, a 41.95M-parameter decoder-only language model trained from scratch in Spanish for cybersecurity, with a Latin-American regional focus and native tool invocation via the Model Context Protocol (MCP). The model has four contributions. (i) Corpus: VectraYX-Sec-ES, a 170M-token Spanish corpus assembled by an eight-VM distributed pipeline at ~$25 USD of cloud compute and split into three curriculum phases (conversational 42M, cybersecurity 118M, offensive tooling 10M). (ii) Architecture: a 42M Transformer decoder with GQA, QK-Norm, RMSNorm, SwiGLU, RoPE and z-loss, paired with a domain-balanced 16,384-token byte-fallback BPE. (iii) Curriculum with replay across the three phases yields a monotonic loss descent (9.80 -> 3.17 -> 3.00 -> 2.16); after SFT (loss 1.74) the v2 bootstrap-ablation reference attains a conversational gate of 0.775 +/- 0.043 on B5 over N=4 seeds, and a controlled Phase-2 replay sweep over {0,5,10,25,50}% saturates B5 at >=25% replay. (iv) Two empirical findings, both N=4. A controlled bootstrap-corpus ablation across v2 (OpenSubs), v4 (mC4-ES), and v6 (60/25/15 OpenSubs/mC4/Wiki) exposes a loss-versus-register inversion: lower-perplexity bootstraps yield measurably worse conversational behavior (v2 > v4 > v6 on B5 at every paired seed). The B4 (tool-selection) floor of 0.000 is a corpus-density artifact, not a capacity gate: rebalancing the SFT mixture to tool-use ratio 1:21 yields VectraYX-Nano v7, the released headline configuration, reaching B4 = 0.230 +/- 0.052 at 42M while retaining B1 = 0.332 +/- 0.005 and B5 = 0.725 +/- 0.130; a LoRA replication on a 260M from-scratch mid-tier reaches 0.445 +/- 0.201. The released GGUF is 96 MB in F16, runs sub-second TTFT on commodity hardware under llama.cpp, and is, to our knowledge, the first published Spanish-native cybersecurity LLM with end-to-end MCP integration.

Diffusion language models are a promising alternative to autoregressive models, yet post-training methods for them largely adapt reward-maximizing objectives. We identify a central failure mode in this setting we call trajectory locking: sampled reward-driven updates over-concentrate probability mass onto a narrow set of denoising paths, reducing coverage of alternative correct solutions under repeated sampling. To address this, we propose TraFL (Trajectory Flow baLancing), a trajectory-balance objective that trains the policy toward a reward-tilted target distribution anchored to a frozen reference model. We make this practical for diffusion language models with a diffusion-compatible sequence-level surrogate and a learned prompt-dependent normalization. Across mathematical reasoning and code generation benchmarks, TraFL is the only evaluated post-training method that improves over the base model in every benchmark-length setting, with gains that persist as the sampling budget increases. The improvements transfer to held-out evaluations: TraFL stays above the base model on Minerva Math and is the strongest method on every LiveCodeBench difficulty split.

Multilingual knowledge editing (MKE) remains challenging because language-specific edits interfere with one another, even when locate-then-edit methods work well in monolingual settings. This paper focuses on three issues: the effectiveness of vector merging methods for MKE, the extent to which Task Singular Vectors for Merging (TSVM) can reduce multilingual interference, and the influence of the weight scaling factor and rank compression ratio on performance. We evaluate six merging variants with two popular backbone large language models, two base knowledge editing methods, and 12 languages on the MzsRE benchmark under a large-scale batch-editing setting. Our results show that vector summation with shared covariance is the most reliable overall strategy, whereas simple summation without shared covariance performs poorly. TSVM improves performance in some settings, but its ability to mitigate multilingual interference is limited. We also find that performance is sensitive to both weight scale and rank ratio, with larger-than-default scaling and relatively low rank often yielding better results. These findings clarify the practical strengths and limits of current vector merging methods for MKE and provide guidance for future multilingual knowledge editing research.

This paper introduces WARDEN, an early language model system capable of transcribing and translating Wardaman, an endangered Australian indigenous language into English. The significant challenge we face is the lack of large-scale training data: in fact, we only have 6 hours of annotated audio. Therefore, while it is common practice to train a single model for transcription and translation using large datasets (like English to French), this practice is no longer viable in the Wardaman to English context. To tackle the low-resource challenge, we design WARDEN to have separate transcription and translation models: WARDEN first turns a Wardaman audio input into phonemic transcription, and then the transcription into English translation. Further, we propose two useful techniques to enhance performance. For transcription, we initialize the Wardaman token from Sundanese, a language that shares similar phonemes with Wardaman, to accelerate fine-tuning of the transcription model. For translation, we compile a Wardaman-English dictionary from expert annotations, and provide this domain-specific knowledge to a large language model (LLM) to reason and decide the final output. We empirically demonstrate that this two-stage design works better than data-hungry unified approaches in extremely low data settings. Using a mere 6 hours of annotated data, WARDEN outperforms larger open-source and proprietary models and establishes a strong baseline. Data and code are available.

Voice agents are increasingly deployed across enterprise applications. However, no existing benchmark jointly addresses realistic conversation simulation and comprehensive voice-specific evaluation. We present EVA-Bench, an end-to-end evaluation framework that addresses both. On the simulation side, EVA-Bench orchestrates dynamic bot-to-bot audio conversations with automatic simulation validation that detects user simulator error and appropriately regenerates conversations before scoring. On the measurement side, EVA-Bench introduces two composite metrics: EVA-A (Accuracy) and EVA-X (Experience). EVA-Bench includes 213 scenarios across three enterprise domains, a controlled perturbation suite for accent and noise robustness, and multi-trial measurements that distinguish peak from reliable capability. Across 12 systems spanning all three architectures, we find: (1) no system simultaneously exceeds 0.5 on both EVA-A pass@1 and EVA-X pass@1; (2) peak and reliable performance diverge substantially (median pass@k–pass^k gap of 0.44 on EVA-A); and (3) accent and noise perturbations expose substantial robustness gaps, with effects varying across architectures, systems, and metrics (mean $\Delta$ up to 0.314). We release EVA-Bench under an open-source license.

Multi-agent LLM systems usually collaborate by exchanging natural-language messages. This interface is simple and interpretable, but it forces each sender's intermediate computation to be serialized into tokens and then reprocessed by the receiver, thereby increasing the generated-token cost, prefill overhead, and KV-cache memory. We study an alternative communication interface: instead of appending a sender's message to the receiver's context, compile the sender's hidden states into a transient, receiver-specific weight perturbation. We introduce TFlow (Thought Flow), a weight-space communication framework for a known and fixed receiver architecture. For each query, frozen role-prompted sender agents process the input, and a learned parameter generator maps their internal activations into low-rank LoRA perturbations targeting the receiver's modules. These perturbations are fused and applied only during the receiver's generation, enabling instance-level adaptation without permanently changing the model or enlarging the receiver's text context. With three Qwen3-4B agents, TFlow improves over a standalone receiver by up to 8.5 accuracy points across five benchmarks while reducing processed tokens by up to 32.69%. Compared with a text-based three-agent baseline, it reduces total processed tokens by up to 83.27% and the wall-clock inference time by up to 4.6$\times$, while maintaining competitive accuracy on four of five benchmarks. These results suggest that transient low-rank weight perturbations can serve as an executable communication medium for efficient multi-agent LLM collaboration.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

Arguments are a fundamental aspect of human reasoning, in which claims are supported, challenged, and weighed against one another. We present an end-to-end large language model (LLM)-based system for reconstructing arguments from natural language text into abstract argument graphs. The system follows a multi-stage pipeline that progressively identifies argumentative components, selects relevant elements, and uncovers their logical relations. These elements are represented as directed acyclic graphs consisting of two component types (premises and conclusions) and three relation types (support, attack, and undercut). We conduct two complementary experiments to evaluate the system. First, we perform a manual evaluation on arguments drawn from an argumentation theory textbook to assess the system's ability to recover argumentative structure. Second, we conduct a quantitative evaluation on benchmark datasets, allowing comparison with prior work by mapping our outputs to established annotation schemes. Results show that the system can adequately recover argumentative structures and, when adapted to different annotation schemes, achieve reasonable performance across benchmark datasets. These findings highlight the potential of LLM-based pipelines for scalable argument mining.

Large language models hallucinate during multi-step reasoning, but most existing detectors operate at the trace level: they assign one confidence score to a full output, fail to localize the first error, and often require multiple sampled completions. We frame hallucination instead as a property of the hidden-state trajectory produced during a single forward pass. Correct reasoning moves through a stable manifold of locally coherent transitions; a first error appears as a localized excursion in transport cost away from this manifold. We operationalize this view with a label-conditioned teacher that builds a trace-specific contrastive PCA lens and scores each step with seven geometric transition features, and a deployable BiLSTM student distilled from the teacher that operates on raw hidden states without inference-time labels. We prove that contrastive PCA is the optimal projection for a transport-separation objective between first error and correct states, and that single-pass first error localization holds whenever the first error creates a positive transport margin over preceding correct transitions. On ProcessBench, PRM800K, HaluEval, and TruthfulQA, both models outperform entropy-based, probing-based, and attention-based baselines in-domain; the teacher transfers stably across language models and datasets, while the student collapses under shift, a gap our distillation theory predicts. These results recast step-level hallucination detection as a problem of trajectory dynamics and identify the central obstacle to deployment: preserving the contrastive transport margin under distribution shift.