Papers

Neologisms, emerging terms in meaning or form, can serve as new vehicles for toxic expression, like "country girl" as a stigmatizing label targeting feminism. Such toxic neologisms appear benign but have evolved into toxic usage in public consensus, posing challenges to moderation systems and remaining underexplored. In this paper, we investigate how to detect implicit toxicity expressed via neologisms. We first propose a taxonomy that captures the origins and consensus-verification criteria of toxic neologisms, followed by the construction of a lexicon spanning widely observed risk categories. To capture toxicity grounded in public consensus, we introduce SeTox, a search-augmented framework that enables static large language models (LLMs) to incorporate real-time web context for neologism toxicity detection. Experiments show that SeTox, even with 3B-scale models, outperforms recent large-scale models, demonstrating its scalability to incorporate real-world knowledge for toxic neologism detection. Disclaimer: this paper has offensive contents that may be disturbing to some readers.

Parameter-efficient fine-tuning enables large language models to adapt to downstream tasks with substantially lower computational and storage cost, and Low-Rank Adaptation (LoRA) is among its most widely used techniques. However, vanilla LoRA assigns a uniform rank to all adapted modules, while existing adaptive methods either incur additional optimization overhead or rely on static weights and local gradients that do not capture task-conditioned representation changes. We propose RSRA, a training-free rank allocator that estimates where adaptation capacity is most needed through forward-only representation sensitivity probing on a small calibration set. Specifically, RSRA uses Spectral Effective Rank to allocate capacity across layers, measures module-wise hidden-state displacement under standardized virtual low-rank updates with the Frechet Distance, and combines both signals through hierarchical normalization to produce a task-aware rank configuration before fine-tuning. Across commonsense reasoning and natural language understanding benchmarks with Qwen3-4B and Mistral-7B, RSRA achieves the highest average performance in all three reported model-benchmark settings and a 1.48x-1.93x speedup in allocation time over the fastest competing pre-allocation method. When integrated with DoRA, LoRA-FA, and PiSSA, RSRA improves 15 of the 18 evaluated combinations and increases the average performance of all three PEFT methods.

While recent advances in large language models have enabled end-to-end automated manuscript generation, existing systems suffer from three critical deficiencies: (i) generated claims are not deterministically grounded in verifiable literature, (ii) experimental results are frequently fabricated rather than executed, and (iii) there exists no standardized, multi-dimensional framework to assess whether AI-generated manuscripts meet the quality and rigor required for real-world publication. We present Prompt-to-Paper, a multi-agent framework that directly addresses this evaluation gap through three integrated innovations. First, a deterministic retrieval-augmented generation pipeline with section-aware relevance scoring and snowball citation expansion grounds every claim in a verifiable corpus of 60–100 papers. Second, an autonomous coding agent executes real computational biology experiments replacing synthetic outputs with genuine numerical results. Third, an eight-dimensional automated quality scorer, benchmarked with approximate reference statistics from published papers and augmented with explicit hallucination penalties, provides standardized, reproducible quality assessments. The quality-driven improvement loop uses a context-rich reviser that routes each iteration to one of three researcher actions and fires a deep research cycle every ten iterations to re-run experiments and re-manuscript from stronger outputs. We validate the system on five bioinformatics case studies; all five cases compiled submission-formatted PDFs with zero out-of-range citations. The improvement loop raises manuscript quality by an average of +17.96 points on a 0–100 scale (maximum +26.04. As partial external checks, a human reviewer scored the five manuscripts at an average of 7.0 out of 10. Complete manuscripts are produced at approximately 0.31 USD per paper.

We present an empirical study of structural routing failure in large language models (LLMs) over a formally verified algebraic corpus. The task requires selecting the correct proof-mechanism label from a fixed closed template set for compact mathematical objects drawn from the FiberRing formalization in Lean 4, where each item is anchored to a Lean-verified artifact and assigned a label from the corresponding certificate family. Our central finding is a mechanism-level routing ceiling: under blind conditions, gpt-oss-120b achieves 80.3% template accuracy on 22 FiberRing items (n=66; temperature=0, seed=0), while Llama 3.3 70B reaches 68.2%. Exposing a mechanism-bearing Lean verdict/witness cue (Condition A2) raises accuracy to 90.9% and 81.8% – gaps of +10.6 and +13.6 pp termed cue-induced routing uplift. The dominant failure is a CRT-to-ring-equivalence misroute: gpt-oss-120b misroutes 7 of 12 CRT items (58.3%) blind, zero under A2. A cross-model dissociation in Llama is notable: verdict accuracy is identical in both conditions (95.5%), while template accuracy improves 13.6 pp – confirming that truth inference and proof-mechanism classification are separable capacities. A cross-corpus extension (Set B; 6 POM/CollisionKernel items, 72 evaluations) provides a small cross-module check: CRT-granularity compression reappears with different labels, and an inverse cross-model dissociation emerges. These findings extend the router hypothesis (Cazares 2026) to formal algebraic structures. The full pipeline, manifest, and results are at https://github.com/bytepro-ai/fiber-routing-eval.

How concepts are represented in neural networks is a fundamental question in machine learning. The dominant view treats concept representations as stationary geometric objects. Yet concepts appear in context, and context transforms them. Drawing from neural population geometry, we formalize concept representations as point-cloud manifolds and contextual transformations as vector fields, and instantiate this framework in large language models. Across six model families of varying scales, we find that context moves each concept differently. The variance in these displacements is semantically organized, correlating with lexical concreteness and density. Importantly, both the concepts being transformed and this variance structure are shared across models: displacement structure transported from one model predicts held-out displacements in others significantly above chance. Together, these findings show that models share a common geometry not only in how concepts are represented, but more importantly in how context transforms them, a structure with richer organization than prior work has recognized.

Generic statements like "tigers are striped" and "cars have radios" communicate information that is, in general, true. However, while the first statement is true in principle, the second is true only statistically. People are exquisitely sensitive to this principled-vs-statistical distinction. It has been argued that this ability to distinguish between something being true by virtue of it being a category member versus being true because of mere statistical regularity, is a general property of people's conceptual machinery and cannot itself be learned. We investigate whether the distinction between principled and statistical properties can be learned from language itself. If so, it raises the possibility that language experience can bootstrap core conceptual distinctions and that it is possible to learn sophisticated causal models directly from language. We find that language models are all sensitive to statistical prevalence, but struggle with representing the principled-vs-statistical distinction controlling for prevalence. Until GPT-4, which succeeds.

Efik, a tonal language spoken by about 3 million second language speakers and 1.5 million native speakers in Southeastern Nigeria, remains underrepresented in speech synthesis research. We present the first documented end-to-end text-to-speech study for Efik, introducing a curated single speaker corpus of 2,632 utterances totaling three hours and a comparative evaluation of four neural models (VITS, MMS-TTS, SpeechT5, and Orpheus-TTS) under low resource conditions. Native speakers evaluated the systems using MOS, Nat-MOS, and A-MOS. MMS-TTS achieved the highest MOS of 3.80 +/- 0.63 and produced more stable long form speech, though tonal errors persisted. Other models showed greater tonal and prosodic inconsistencies. These results provide a reproducible baseline and highlight the need for larger corpora and tone aware modeling for tonal African languages.

Emergent misalignment (EM) — the broad misbehaviour a language model acquires after fine-tuning on narrow harmful data — is mediated in Qwen2.5 models by a latent persona direction, and that direction is causal in open weights. Transplanting it into a model that shares only pretraining with its source induces broad EM (2.83 $\pm$ 0.26% misaligned against a random-direction floor of $\sim$1.1%), and ablating a model's own direction roughly halves an overt inducer's broadcast (21% to 10%). The transplant doubles as a measurement method, causally assaying directions that a source model represents but cannot itself express. Whether a fine-tune recruits this persona depends on method and capacity, and since low-rank PEFT is the cheaper regime at scale, the recruiting method is also the economical one. On Qwen2.5-32B, LoRA at low ranks on insecure code recruits it (3.4% misaligned) while full SFT on identical data does not (0.3%) and moves against the persona axis (drift–persona cosine $+0.17$ at rank 1 to $-0.10$), the far-inducer, high-capacity exception consistent with a representational-distance $\times$ capacity account. The persona's causal role is itself conditional. Steering a bad-medical SFT run away from the direction during training raises the broadcast from ${\sim}24\%$ to ${\sim}50\%$ while matched random controls stay at or below baseline, replicated across three training seeds, so removing the direction is no blanket recipe. Because recruitment is a loss-reducing shortcut that capacity renders redundant, it can be screened for and prevented in the tested instances. Persona loss-relevance at the SFT solution orders four inducers' broadcasts rank-perfectly within Qwen2.5, inoculation removes recruitment selectively (4.75% to 0.0%, code coherence 65% to 87%), and fine-tuning orthogonal to the single behaviour-derived axis reduces it persona-specifically.

Diffusion language models (dLLMs) commit multiple tokens per denoising step by decoding each selected position independently from a shared context. When these positions are dependent, this factorization introduces an error captured by conditional total correlation, which confidence-based selection cannot infer from marginal probabilities alone. We propose CoCommit, a marker-gated coordination pass that delays commitment. After the usual bundle selection, a learned marker identifies the commit set, and the backbone's last n layers are re-applied to coordinate the marked positions before greedy argmax writes the tokens. This approximates joint-mode decoding while reusing existing weights, requiring only one partial forward pass and no auxiliary model. On LLaDA 2.1 with LoRA adapters and greedy inference, joint commitment improves five of the seven evaluated benchmarks over the released factorized decoder. The largest gains occur on code and reasoning tasks, while the remaining tasks are near parity.

The rapid integration of large language model-based agents into recommender systems has driven a shift from static, ranking-based pipelines toward autonomous and interactive systems that can reason, plan, and act. This survey provides a comprehensive overview of this emerging landscape by introducing a unified taxonomy grounded in the level of autonomy and three core paradigms of agentic recommender systems: agent-assisted recommendation, agent-as-recommender, and agent-as-user-simulator. The autonomy framework organizes existing methods along increasing capabilities in proactivity, context awareness, interaction flexibility, and adaptivity. Building on this framework, the survey analyzes how each paradigm adopts different agentic architectures and how agents enhance key components such as profiles, memory, tool use, workflows, and optimization mechanisms. We further examine evaluation methodologies for agentic recommendation, covering automated metrics, LLM-based judging, and simulation-based assessment, and discuss their limitations in capturing reasoning quality, user experience, and system behavior. Beyond existing evaluation protocols, we further discuss unresolved issues in evaluating agentic recommender systems, including trajectory-level assessment, agent contribution analysis, and calibration of user simulation. Lastly, the survey outlines open challenges in lifelong user modeling, contextual abstraction, multimodal alignment, controllability, trustworthiness, privacy, scalability, and efficiency. Together, these analyses establish a unified foundation for understanding the current progress of agentic recommender systems and highlight promising opportunities for developing more autonomous, reliable, and human-aligned recommendation agents.

Large language models (LLMs) are increasingly deployed in question answering (QA) systems, yet they may generate hallucinated or misaligned responses without reliable confidence estimates. Uncertainty quantification (UQ) offers a natural basis for selective answering, where a system answers only when its prediction is deemed reliable and abstains otherwise. However, existing uncertainty scores for LLMs are often heuristic: a threshold chosen on such scores does not, by itself, provide statistical guarantees on the error rate among accepted answers. We propose CIC, a confidence-interval-based calibration framework that converts arbitrary uncertainty scores into risk-controlled selective answering rules. Given a held-out calibration set, CIC evaluates each generated response using an application-specific alignment criterion and associates it with an uncertainty score and a binary error label. For each candidate uncertainty threshold, CIC estimates the acceptance-conditioned error rate and constructs a high-probability upper confidence bound using either Hoeffding-style or Clopper-Pearson confidence intervals. It then selects the largest threshold whose upper bound is below a user-specified risk level $\alpha$, thereby maximizing the answering rate subject to a finite-sample reliability constraint. Under exchangeability, CIC guarantees with probability at least $1-\delta$ that the selected threshold, if non-null, controls the error rate among accepted answers at level $\alpha$. We evaluate CIC on both closed-ended and open-ended QA benchmarks across seven LLMs and multiple uncertainty estimators. Experimental results show that CIC consistently achieves valid risk control while retaining strong answering efficiency, providing a practical and statistically grounded mechanism for deploying LLMs in reliability-sensitive QA workflows.

The dominant practice in language model evaluation is to report a single accuracy number per model and declare the higher one better, without testing whether the gap could plausibly be sampling noise. On benchmarks of a few thousand items, and under temperature sampling where a model can differ from itself run to run by more than the reported gap between models, this practice routinely overstates confidence in headline claims. The statistical machinery to fix this – confidence intervals, paired significance tests, power analysis, clustered standard errors, multiple-comparison correction – is well established, but no standard, pip-installable tool packages it in the shape an evaluation actually takes: a per-item results table. We present evalci, a pure-Python library (numpy/scipy/pandas only) that turns a per-item results table into a publication-ready claim – e.g., "Model A beats Model B, $\Delta=3.1$ pts, 95% CI [1.2, 5.0], paired permutation $p=0.002$, $n=1{,}319$" – in one function call, with adapters for lm-evaluation-harness and HELM output. Every routine is validated against an independent reference (statsmodels, or brute-force exact enumeration) rather than only against itself. As a case study, we re-analyze a public comparison of nine language models' MMLU accuracy and find that 3 of the 8 adjacent leaderboard-rank gaps are not statistically significant after correcting for the 36 pairwise comparisons the ranking implies. evalci is available at https://pypi.org/project/evalci/ (source: https://github.com/Shreyaskc/evalci, DOI: https://doi.org/10.5281/zenodo.21201815)

Diffusion large language models (dLLMs) generate text by iteratively denoising a masked sequence, offering a parallel alternative to autoregressive models, but eliciting strong reasoning through post-training remains difficult: supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards and is hard to apply without tractable sequence likelihoods. On-policy self-distillation (OPSD) offers a promising alternative, using one model as both student and teacher to provide dense, token-level, on-policy supervision, but its effectiveness hinges on giving the teacher privileged information (PI) - typically an instance-specific ground-truth reference unavailable at inference - so the student ends up distilling a weak PI-free consensus policy that yields little improvement on dLLM reasoning. We introduce dOPSD, which instead derives the teacher's privilege directly from the student's own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher's advantage emerges from the model's own decoding process; on Dream and LLaDA, dOPSD improves both in-domain math reasoning and out-of-domain code generation, outperforming supervised and on-policy baselines.

Recent advances in multimodal foundation models and agent systems have driven GUI agents from single-platform task execution toward cross-platform interaction. However, unified multi-platform GUI learning remains challenging: high-quality cross-platform trajectories remain scarce, while platforms share transferable capabilities but differ in action semantics and interaction conventions. Naively mixing supervision or merging specialized models can blur native behaviors and produce imbalanced performance. To address these challenges, we construct Uni-GUI, a high-quality dataset containing nearly 10K executable cross-platform interaction trajectories collected through a unified desktop-mobile harness. Building on Uni-GUI, we propose UI-MOPD, the first framework to introduce multi-teacher on-policy distillation (MOPD) into unified multi-platform GUI agent training. UI-MOPD trains a shared student on its own rollouts and dynamically routes each rollout to the corresponding platform-specialized teacher. At student-visited states, teacher guidance serves as a platform-conditioned behavioral anchor, enabling the integration of complementary desktop and mobile expertise without averaging their distinct interaction conventions. On OSWorld and MobileWorld, UI-MOPD achieves task success rates of 38.2% and 12.0%, respectively, outperforming parameter-matched integration strategies while preserving general GUI grounding. These results demonstrate that multi-teacher on-policy distillation provides an effective approach to building unified cross-platform GUI agents. Project page: https://elispectre.github.io/UI-MOPD/.

We present the AI Wizards submission to EXIST 2026 for multimodal sexism identification in memes. The task is composed of three, increasingly harder subtasks. We model them hierarchically as conditional soft-label prediction over empirical annotator distributions. Our system maps fixed Gemini Embedding 2 vision-language representations through a lightweight Gated MLP trained with KL divergence and homoscedastic uncertainty weighting. Our submissions ranked first on Task 2.3 and fourth on Tasks 2.1 and 2.2 on the official Soft-Soft leaderboards. The code is available at https://github.com/NLP-AI-Wizards/EXIST-2026

Long-term memory remains a structural weakness of AI agents. The dominant approach, retrieval-augmented generation (RAG), relies on embedding-based similarity search, which is opaque by construction, difficult to audit, and bounded by the theoretical limits of vector representations. We present the Memory-Orchestrated Semantic System (MOSS), an agentic memory architecture in which the agent drives retrieval over a structured relational database. MOSS is model-agnostic, storage-agnostic, and API-agnostic: it runs on any relational engine, connects to any LLM provider (or to deterministic non-LLM processes), and deploys on any infrastructure, local or cloud. Its retrieval execution is symbolic and reproducible (once a query is formulated, no LLM participates in the retrieval loop) and every step of the system, from indexing to answer formulation, is logged and inspectable, making MOSS auditable by construction. Rather than imposing an external ontology, MOSS derives its conceptual vocabulary from the corpus itself. We report on a longitudinal deployment unique in the agentic-memory literature: a year of continuous production over an individual scholar's working corpus–a conversational corpus reaching back to October 2024 (some 44 million tokens, retroactively indexed) comprising 110,183 segments, alongside 163,494 catalogued documents, 569 inductively derived concepts, 322,662 concept annotations, and eleven metadata graphs totaling approximately five million relations–across four successive infrastructure generations. While the present case is that of a single researcher, the architecture is in no way specific to one person: it serves a team, an institution, or any entity that accumulates knowledge over time. We argue that auditable, sovereign, structurally unbounded memory is a precondition for AI agents intended to accompany a person or an organization over years rather than sessions.

Online social media posts provide scalable signals for early depression screening, and recent studies mainly improve pre-classification evidence through risk-post selection, symptom grounding, and clinically informed feature construction. However, these screening-stage designs often leave final decisions to a single detector, overlooking how users heterogeneously express depressive risk after screening. A monolithic classifier must average across heterogeneous users, which may dilute localized evidence and cause misclassification, especially for non-self-disclosing users. To address this issue, we propose WPG-MoE, a weak-prior-guided dense mixture-of-experts framework built on a shared large language model (LLM) backbone. WPG-MoE derives user-level weak semantic priors to softly route users to experts matched to different evidence layouts. We formulate this process as learning using privileged information (LUPI): rich LLM-extracted structured evidence guides training-time routing, while inference retains only Patient Health Questionnaire-9 (PHQ-9) template screening and the deployable backbone. Experiments on Chinese and English datasets show that WPG-MoE outperforms strong baselines with interpretable routing behavior.

This state-of-the-art report provides an overview of controllable 3D human avatar creation. We describe current 3D avatar systems, which typically consist of three stages: (i) learning priors of human appearance and motion, (ii) creating a personalized avatar, and (iii) animating the avatar. To limit the scope, we focus on the prior learning and avatar creation stages. We define current avatar representations and introduce a taxonomy that categorizes existing work along multiple axes, including body regions and employed priors. We review methods for full-body and head avatars, as well as layered representations that decompose the body into components such as hands, hair, and garments. Finally, we outline common underlying principles, reference key literature for newcomers, and discuss open challenges and future research directions.

A companion paper showed that a transformer's feed-forward layer can be rebuilt from explicit fuzzy set operations - intersection, set-difference, and a self-forgetting sequence quantifier - so its hidden units read as named logical operators at no cost to language-model quality. That left the other half of the transformer opaque. Here we carry the same idea into attention and join the two into one model. The mechanism is minimal: a head's value is passed through a sigmoid, so each value channel becomes a readable detector of whether a feature holds at a token. This adds no parameters and leaves the standard head otherwise untouched. A Boolean variant goes further, restructuring the value into an explicit within-token intersection and negation-capable set-difference. In both designs the output projection is left free, not tied to the vocabulary, which is the load-bearing decision: bounding what a head detects while leaving what it writes unconstrained yields selective detectors, whereas constraining the write does not. A bounded value is shaped into a readable detector by two selectivity pressures - one for sparse firing, one for decisive firing at the rails - and which a design wants is not universal. Across five specialized-attention designs at 125M parameters, 44 to 62 percent of value channels become crisp, contextually selective detectors, and their legibility rises with depth rather than crystallizing only on punctuation. Language-model quality is at parity with a conventional baseline. Finally, we couple the Boolean attention to the legible feed-forward layer and train an end-to-end legible-by-construction language model at benchmark parity: its feed-forward units are named set and quantifier operations throughout, and we can take a token it generates and read the named units that compose to produce it.

We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16. To our knowledge, HiFA4 is the first Ascend-HIF4-targeted design of this kind evaluated on standard NLP benchmarks. HiFA4 combines two mechanisms. Smooth-QK applies a calibration-static per-channel equivalent rescaling to Q and K after RoPE, transferring quantization difficulty from K to Q without per-tile online reduction at inference. P-Reordering accumulates the softmax normalizer from the same quantized attention weights P_hat used in the PV GEMM, rather than from a higher-precision reconstruction. We show that this inconsistent formulation introduces a coherent output-scaling error, and validate the effect on a Qwen3-8B Layer-0 MMLU trace, where all 3.6M measured attention tiles exhibit net probability-mass loss with median epsilon_bar = -0.064. P-Reordering also allows the normalizer to be fused into the PV Cube GEMM. Across five LLMs, HiFA4 reduces quantization-induced decision drift. On Qwen3-8B, it recovers 37.5% of the accuracy gap introduced by direct HIF4 quantization, narrows the sample-weighted accuracy loss from 1.12 pp to 0.70 pp, reduces BF16-inconsistent MMLU predictions from 16.3% to 8.2%, and cuts MMLU accuracy regressions by 57% (1071 to 465). On Gemma2-9B, mild smoothing keeps HiFA4 within 0.7 pp of BF16 while reducing MMLU regressions by 27%. On LLaMA3.1-8B, Mistral-7B, and Phi-4B, where Smooth-QK is disabled, P-Reordering with the adopted Q-Mean auxiliary still reduces full-set MMLU regressions by 41-52%. A preliminary instruction-scheduling analysis projects a 35.4% critical-path latency reduction relative to BF16 by fusing the softmax normalizer into the PV Cube GEMM; on-hardware validation is left to future work.

Building AI Scientist agents with Large Language Models (LLMs) has recently attracted growing attention. Since scientific discovery fundamentally relies on uncovering causal relationships from observations, the capability of causal thinking, i.e., distinguishing causation from correlation and recognizing hidden biases, is essential to LLM agents. Although a number of benchmarks exist for AI Scientists, none explicitly incorporate challenges from selection bias, measurement error, and hidden confounders that widely exist in real-world scientific discovery. To this end, we present CausalGame, a benchmark that evaluates the causal thinking capabilities of LLM agents through interactive games. CausalGame asks LLM agents to actively design experimental protocols, collect observation data, and derive a final solution with an explanation report. To emulate realistic scientific discovery challenges, we design 14 scenarios that incorporate selection bias, measurement error, and hidden confounders. Across 30 LLM agents, none demonstrates reliable causal thinking: the best model reaches only 68.0% survival against analytical optima of 78-85%, and merely 5-7% of sessions receive credits on the causal-reasoning rubrics. CausalGame provides a scalable and controlled testbed for evaluating the causal thinking of AI Scientist agents.

Semantic caching reduces the latency and cost of retrieval-augmented generation (RAG) by serving cached answers to semantically similar queries, but most existing methods do not model the time-varying freshness of open-web evidence. We present FreshCache, a three-tier semantic cache that treats cache reuse as a risk-constrained temporal inference problem: before approving a cache hit, FreshCache estimates the probability that the cached result is stale using a fitted exponential decay model enhanced by a learned MLP, and approves reuse only when that probability falls below a per-tier error budget across answers (epsilon = 0.10), URL lists (epsilon = 0.20), and page content (epsilon = 0.35). This allows the system to degrade gracefully as entries age rather than forcing a binary choice between a stale hit and a full pipeline execution. We introduce FreshCache-Bench, a benchmark of 8,072 base queries across five freshness classes with ground truth staleness labels drawn from real web snapshots at 1, 12, 24 hours, and 7 days after a baseline crawl, expanded to 31,201 queries via paraphrase generation. At the 24-hour evaluation window, FreshCache_MLP achieves 97% search API savings at 0.1% hash-based stale error, and an LLM-judge evaluation on 396 confirmed change pairs shows that only 34.3% of detected content changes actually affect answer correctness, placing true answer-affecting stale error at approximately 0.034%. The rule-based FreshCache achieves 98% search savings at 3.3% stale error under a temporal holdout calibration, outperforming SemanticTTL (14.9% stale, 72% saved), vCache (7.2% stale, 47% saved), and SCALM (5.2% stale, 96% saved). Ablations show the temporal risk gate accounts for an 11.6 point reduction in stale error over similarity-only reuse, and the learned MLP reduces stale error a further 3.2 points over the rule-based model.

Large language model agents operate in partially observable, long-horizon settings where obtaining supervision remains a major bottleneck. We address this by utilizing a source of supervision overlooked in existing post-training methods: unintended yet successful goals embedded within agent rollouts. Specifically, we introduce Hindsight Supervised Learning (HSL), where an auxiliary LLM reviews each completed trajectory and relabels it with all of the natural-language goals the agent actually achieved. HSL then pairs the trajectory with its relabeled goals and uses these pairs for additional fine-tuning. To mitigate suboptimality in the relabeled data, we propose two learning techniques for HSL, irrelevant-action masking and sample reweighting. Our experiments show that HSL is flexible and compatible with existing post-training pipelines. It improves both SFT and DPO, with larger gains on long-horizon tasks with more diverse goal spaces. Moreover, HSL is sample-efficient: on ALFWorld, it surpasses baselines trained on the full dataset while using only one quarter of the ground-truth demonstrations.

Unlike natural-language specifications, executable formal specifications provide machine-checkable constraints for verifying, debugging, and repairing code. However, writing such specifications is labor-intensive, and existing LLM-based methods mainly infer whole-program pre/postconditions, missing the intermediate semantic commitments that programmers rely on when reasoning about an algorithm. Our study further shows that prompting current CodeLLMs often produces executable assertions that are syntactically invalid, trivial, or too weak to reject behavior-changing faults. In this paper, we study executable checkpoint specification generation, where assertions are inserted at meaningful internal program points to describe expected intermediate states. We introduce SpecCoder, a verification-guided CodeLLM training framework that learns from validated reference programs, behavior-changing mutants, and multi-turn specification-refinement traces. SpecCoder selects specifications that hold on correct executions while rejecting faulty executions, turning specifications from passive annotations into executable evidence. To evaluate this setting, we introduce HumanExec, a benchmark built from recent Codeforces competitive programming problems with test suites, reference solutions, and human buggy submissions, supporting three tasks: specification generation, program correctness checking, and program repair. Experiments on HumanExec show that SpecCoder substantially improves checkpoint-specification quality over base CodeLLMs. Across Qwen2.5-Coder models, SpecCoder improves inline-specification correctness by up to 55.8%, completeness by up to 358.1%, and executable assertion validity by up to 26.6%. These gains further translate to downstream correctness reasoning and repair, showing that executable checkpoints provide fine-grained evidence for reliable verification.

Retrieval-augmented generation (RAG) assistants summarize records in clinical and legal work, where one unsupported sentence can mislead a reader. The contrast between an output's likelihood with and without its source is an established faithfulness score for whole summaries and answers, but it has not been measured as a detector of the individual unsupported sentence in multi-passage RAG answers, against trained verifiers, or for its cost. We implement it as a training-free detector that re-scores a fixed answer under the full context, no context, and each chunk removed, and returns the chunk whose removal lowers a sentence's likelihood most as a candidate supporting passage. We evaluate it on RAGTruth, TofuEval, and RAGBench with six scorers and against five verifiers, up to a large language model (LLM) judge, on identical inputs under a source-level split. Scoring per sentence ranks unsupported sentences better than the answer-level form of the same signal on all three benchmarks, by 0.033 to 0.071 in the area under the receiver operating characteristic curve (AUC). On RAGTruth the training-free score reaches an AUC of 0.717 to 0.745 across scorers and 0.773 with a classifier, above entailment and attribution baselines and level with per-chunk fact-checkers, at about one forty-seventh of the LLM judge's compute on a 1.5B scorer, while a full-context fact-checker and the judge are more accurate and are not improved by it. The signal is weakest on short-answer question answering, where the scorer can answer from memory.

This work establishes that trigger-word data poisoning of vision language action models is practical, while at the same time the open-source robotics ecosystem holds trust assumptions about community contributions. A few poisoned samples can silently embed a backdoor that disables a robot on command. We evaluate this threat against smolVLA on a real-world pick-and-place task, training on three poison ratios and evaluating across different prompts on the LeRobot platform. Three poisoned episodes in 320 clean episodes suffice for a complete denial of service. Success rate drops to 0.0 plus minus 0.0% across all trigger-word conditions and the robot locks into a fixed joint configuration rather than executing any task-relevant motion. Clean-prompt behaviour holds at approx. 50% success rate across all poison ratios, confirming the attack is stealthy under normal operation. A single poisoned episode already reduces success rate to 6.7 plus minus 6.7%. The robot still moves, but no longer completes the task. The attack generalises to front, middle, and end trigger placements despite training exclusively on front-placed triggers. These findings establish that the threat is practical, low-cost, and stealthy, and warrant treating dataset provenance as a first-class concern in open-source robotics ecosystems.

Autoregressive (AR) text-to-speech (TTS) models generate discrete speech tokens sequentially, which makes inference slow and can degrade robustness, since local errors propagate to later positions and can escalate into hallucination. This limitation stems from their left-to-right AR commitment: each token must be determined before future speech-token context is available. However, such ordering is not an inherent requirement for TTS, since the model receives the full input text before synthesis. In this paper, we introduce DELTA-TTS, a lightweight LoRA-based adaptation framework that converts a pretrained AR TTS model into a discrete diffusion language model (dLLM) for confidence-ordered speech-token decoding. To better capture the local structure of speech, DELTA-TTS incorporates a convolution module that injects local acoustic context, together with a 1/t-weighted training objective and a time-shifted inference schedule that together defer low-confidence positions to later steps. Trained on only 585 hours of LibriTTS, DELTA-TTS achieves a 1.75% WER on Seed-TTS test-en, outperforming its AR backbone while generating tokens 3.3 times faster. Further analysis shows that DELTA-TTS produces sharper text-speech alignment, increases overall decoding confidence, and mitigates the hallucinations observed in AR generation. Audio samples, inference code, and model weights are available at https://slpr-lab.github.io/DELTA-TTS.

Multimodal LLMs struggle to systematically model the temporal evolution of visual scenes in videos or multi-image sequences. Such inputs require models to predict or simulate multiple levels of dynamic constituents, such as actions taken in the visual sequence, and the associated changes to the visual environment that result. To address this challenge, we propose a dynamic schema-guided world model, DynaVieW, optimized for visual dynamic prediction and simulation. DynaVieW achieves an in-depth understanding of visual dynamics by learning interleaved state-transition sequences, where states cover broad visual scenes from video keyframes, and transitions capture comprehensive dynamic constituents within a hierarchical schema. DynaVieW jointly models transition prediction and state simulation under a mixture-of-experts architecture, with a cross-expert selective attention and a schema token re-weighted loss, to ensure effective and robust learning. DynaVieW's understanding of visual dynamics boosts its downstream performance in visual narrative creation and world simulation, showing improved consistency, controllability, and instruction-following.

Word surprisal is a well-established computational predictor of human neural responses during language comprehension, but it remains less clear whether local semantic fit explains neural response variation beyond lexical expectation during naturalistic reading. Using the Dublin EEG-based Reading Experiment Corpus (DERCo), this study examined whether contextual semantic relevance predicts word-locked EEG activity in the N400 and P600 windows. Contextual semantic relevance was computed as an attention-aware measure of how strongly a target word is semantically connected to its recent discourse context, and it was compared with GPT-based word surprisal. Across 22 participants and 32 EEG channels, we tested both predictors using regression-based ERP analyses and generalized additive mixed models while controlling for lexical variables and repeated observations. Both predictors were reliably associated with EEG responses, but they showed partly different temporal and scalp-level patterns. Surprisal captured expectancy-related variation, whereas contextual semantic relevance showed robust effects across N400- and P600-window mean voltages, with particularly strong explanatory support in the P600 window. Model comparisons indicated that contextual semantic relevance contributed explanatory value beyond lexical controls and surprisal. These findings suggest that naturalistic reading depends on both lexical expectation and local semantic integration, and that contextual semantic relevance offers an interpretable computational link between discourse semantic fit and ERP dynamics.

Portuguese remains underrepresented in text embedding evaluation, despite being one of the most widely spoken languages in the world. As a result, embedding models are often selected based on English or multilingual metrics, while their effectiveness in Portuguese remains unclear. We present MTEB-PT, a Portuguese benchmark constructed from a subset of MMTEB, comprising 14 existing datasets across Semantic Textual Similarity (STS), classification, retrieval, and reranking. We use this benchmark to evaluate 17 open- and closed-source embedding models under a unified protocol. Our results show that Portuguese performance is strongly task-dependent: multilingual rankings do not reliably predict Portuguese-specific performance across task families, no single model dominates all settings, and models with stronger long-context capacity are particularly advantageous on longer-input tasks such as retrieval and reranking. The benchmark also shows that language-specific fine-tuning still improves model performance in Portuguese, especially on task types that match the adaptation data most closely. To examine this effect, we fine-tune three representative backbone models with Portuguese contrastive supervision and Matryoshka Representation Learning (MRL). These benchmark-informed baselines yield their strongest gains on STS, consistent with the predominantly symmetric supervision used during training, while also improving retrieval and remaining competitive under dimensional truncation. We release the MTEB-PT benchmark, the fine-tuned models, and the training and evaluation code.