Papers

Substance Use Disorder (SUD) counseling requires patient responses that reflect underlying cognitive states such as beliefs, coping strategies, and readiness for change. Although large language models (LLMs) can generate fluent text, they often fail to produce cognitively coherent and clinically realistic patient behavior, especially under ethical and data-scarce clinical settings. Moreover, deploying frontier-scale LLMs in healthcare applications presents practical challenges including high computational cost, latency, privacy concerns, and limited deployability in resource-constrained environments, motivating the need for cognitively aligned small language models (SLMs). We propose a cognitively grounded framework for SUD patient dialogue generation that explicitly models and aligns latent cognitive components with patient histories and counselor questions. Our pipeline consists of two stages: cognitive component detection and cognitive component-aligned dialogue generation. To enable effective learning with smaller models, we combine knowledge distillation from high-capacity teacher models, preference optimization from human-annotations, and attention-guided reward shaping. Extensive evaluations using automatic scores like BERTScore, ROUGE, METEOR and BLEU, and LLM-as-judge hit-metrics against both human and teacher-model references show that cognitively informed fine-tuning substantially improves cognitive realization and alignment over a generic instruction-tuned baselines and mental health domain specific SLMs, with particularly strong gains for open-ended cognitive components.

Extreme low-bit LLM compression is most challenging when weights, activations, and KV caches are quantized together: their distributions differ, and quantization errors interact throughout the network. We introduce CanonQ, a unified quantization-aware training framework that addresses these challenges by separating source canonicalization from task-aware adaptation. Fixed rotations and energy normalization map heterogeneous tensor sources to canonical coordinates, enabling frozen Gaussian-reference codebooks to be reused across layers and models. Joint training then adapts the network to the coupled errors of weight, activation, and cache quantization within a common scalar/vector interface. We bound frozen-codebook transfer error and local task loss, and derive an exact normalization-aware straight-through Jacobian that links quantization distortion to gradient bias. The strongest gains arise under joint W2A4KV2 compression: across LLaMA3-1B/3B/8B, CanonQ-Omni achieves up to 14.28x lower WikiText-2 perplexity and up to 57.9% higher mean zero-shot accuracy than prior state-of-the-art and representative quantization baselines. The benefits extend to Qwen3-1.7B, code generation, and mathematical reasoning: on instruction-tuned MobileLLM-Pro-1B at W2A16KV16, CanonQ achieves relative improvements of 41.7% in HumanEval pass@1 and 39.1% in GSM8K exact match over the strongest evaluated quantization baseline.

MemoryAgentBench's Conflict Resolution split is read as measuring "selective forgetting". We execute the benchmark's own rule - the newest statement about a fact wins - as a zero-learning resolver frozen on one of the four fact lists. Under the official metric the rule answers 80.25% of the questions (74.5% on the three held-out lists). Of the rest, 67 items have a released gold that the last-write graph cannot reach but overwritten statements would ("The capital of India is New Delhi." superseded by "The capital of India is Grosseto."; gold New Delhi); such items are a third of the multi-hop questions at 262K. Two long-context models and our pre-registered approximate re-implementation of the benchmark's BM25 agent, one retained run per item and outcomes only, score 84.7%, 82.6% and 41.6% on the items the rule solves against 10.4%, 11.9% and 6.0% on those 67. The failures are a reachability split plus a small parser-scope residual; the per-item split, not the aggregate, is the unit at which a score here can be read.

As data propagates through a Transformer, the norm of its hidden states grows by orders of magnitude with depth, a phenomenon framed as 'curse of depth' and nearly universally treated as a pathology to be suppressed. We take the opposite view. Across 16 pre-trained LLMs from 9 families, spanning dense, mixture-of-experts and hybrid architectures and Pre-, Peri- and Post-Norm designs, we find that this growth reflects an emergent depth-positional encoding, carried by the only learned per-layer gain on the residual stream, the normalization weight $\gamma$: with depth, $\gamma$ grows in magnitude and rotates in direction, jointly encoding the layer index. We make this depth-conditioned encoding explicit with LayerRoPE, an implicit analog of RoPE along the depth axis, which replaces all layerwise $\gamma$ vectors with a single shared vector and depth-conditioned scalars, at a net reduction in parameters and $<0.02\%$ change in FLOPs. Across a model ladder scaled up to $100$B+ tokens, LayerRoPE consistently outperforms Pre-, Post- and Peri-Norm and Layer-Norm Scaling, reaching Pre-Norm's 1.3B loss with $3.4\times$ less compute; LayerRoPE is the only approach that shows strong convergence and improves near monotonically as depth scales to 512 layers. It improves learning-rate sensitivity by $3$-$10\times$, and transfers naively to and consistently improves looped latent models and Vision Transformers. Inspecting its learned schedule inverts the prevailing premise: LayerRoPE does not shrink the residual stream but widens it, damping what each block reads while amplifying what it writes. Depth stability, our results suggest, calls not for suppressing the residual stream, but for depth-conditioned regulation of the computational blocks it feeds.

Task-oriented conversational agents remain fragile under real world conversation scenarios as they rarely follow a predictable script, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks often emphasize successful, cooperative interactions and underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present ToolRACER, a synthetic data generation pipeline that coordinates user, assistant and tool emulation models to generate and validated multi-turn interactions between a user and an agent. Using \sysn, we construct ToolRACERBench a robust multi-turn conversation benchmark spanning six domains, ranging over 55 varied personas, generating a validated corpus of 5.6K conversation trajectories, with approximately 66% of conversations containing failure-prone conversation scenarios. We inject adversarial behaviors, producing validated conversational interaction trajectories that capture realistic, robust scenarios. We evaluate models trained on ToolRACERBench against internal benchmarks, as well as on function calling benchmarks such as $\tau^2$-bench, BFCLv3 and ACEBench to evaluate agentic accuracy and robustness. Models trained on ToolRACERBench improve end to end agentic accuracy across $\tau^2$-bench and ACEBench, demonstrating significant gains when mixed with in-domain dataset in small language models for agent capability tasks.

Multilingual LLM benchmarks omit Slovak, a morphologically rich West Slavic language of five million speakers, or cover it only by machine translation. We present sk-bench, a native-first Slovak benchmark with 30 datasets (33 scored task variants) across ten skill categories. Eleven resources are introduced or first packaged for generative-LLM evaluation, including IFEval-SK with Slovak-adapted instruction checkers and native Chiby/SKJ1 resources for Slovak grammar and morphology. We evaluate 55 open- and closed-weights models under one harness. The best open model trails proprietary APIs by 12.6 points. Model rankings are similar for native and translated closed-form data ($\rho\geq0.98$), though translation separates the strongest models less well. By contrast, human-authored and LLM-generated QA questions rank models differently ($\rho=0.72$). For Qwen3-14B, continued Slovak pretraining lowers the overall score by 13.9 points. A small instruction set restores three quarters of that loss. Test-time reasoning improves scores by 8.5 to 12.5 points for models of 9B and above. Together, these findings suggest four design lessons for other under-resourced languages: use native data where translation fails, plan instruction repair after language adaptation, enable test-time reasoning before scaling up, and avoid overinvesting in target-language prompts. We release the data and code at https://github.com/slovak-nlp/sk-bench

We revisit a standard accepted practice in the continuous diffusion language model literature of fixing conditioning prompt tokens clean during training. We make a very simple modification: also noise the conditioning prompt tokens during training. We demonstrate that under this modified training objective, we achieve better generalization in combinatorial reasoning tasks such as Sudoku and N-Queens, with the largest gains on harder variants ($3.73\% \to 24.65\%$ solve rate on Sudoku Hard), and increased diversity of generated solutions ($50.60\% \to 73.79\%$ coverage on 10x10 N-Queens). We also show measurable improvements to natural language generation quality in modest dataset regimes with Gigaword summarization, but notably demonstrate that gains do not transfer to all natural language tasks (e.g open ended dialogue generation). Our method is a single line change to the training objective, requires no additional inference costs by default, and provides the flexibility of classifier-free guidance inspired guided sampling. Our \href{https://github.com/LateralIntelligence/noise-your-prompt}{code} is publicly available.

Steering an LLM agent means deciding whether to correct it, at which step, and with which mechanism. Uncertainty is often used to decide when to correct an agent, but whether it can guide these decisions remains unclear. We steer agent trajectories separately at every non-terminal step with each of four mechanisms and run each continuation to completion. The resulting stepwise outcome table (SOT) holds about 82,000 counterfactual continuations of 1,864 trajectories from three benchmarks and two agents. It shows that uncertainty can identify failing trajectories, but that no single signal reliably locates the step at which steering helps. We therefore propose VoS (Value of Steering), a trajectory-level monitor, offline or online, that learns from SOT the value of steering at each step and decides where to steer by it. A harm-budgeted trigger decides whether to steer, limiting the fraction of successful trajectories that VoS disturbs. VoS improves on unmodified execution in all 12 settings of benchmark, agent, and offline or online use, by 7.8 points on average, and outperforms the strongest of five existing uncertainty-triggered methods in 11, by 2.9 points on average. Ablations show that training on measured outcomes and a tight harm budget are both essential.

Deterministic inference is essential for reliable and trustworthy machine learning. Prior studies of text generation have shown that changing factors such as batch size, batch composition, hardware, or inference engine can alter the generated text, even when the prompt, model parameters, and sampling randomness are fixed. These differences have been attributed in part to floating-point non-associativity, shape-dependent kernel selection, and other implementation-level differences in numerical execution. However, it remains unclear whether, when, and to what extent the same factors affect text classification. We present a systematic study of serving-context non-invariance in text classifiers, which prior work has measured only through generated text. We train 180 models spanning discriminative, pseudo-generative, and fully generative classifier formulations and evaluate each across four categories of serving contexts, holding the checkpoint and the text fixed. Label stability does not imply score stability. Changing only the batch shape changes no labels across fp32 comparisons, yet under bf16 it moves up to 56.7 percentage points of predicted probability mass, with label changes concentrated at small margins. Fully generative classifiers change more labels than their discriminative counterparts under the same serving changes. We derive sufficient conditions for label stability under each serving change and give a separate mitigation for each mechanism. Our results identify and quantify the serving conditions that must be fixed for reproducible text classification.

Natural-language feedback in interactive learning often explains why an action failed by pointing to violated requirements. Misinterpreting this feedback can lead an agent to rule out valid solutions. We study this setting by modeling user intent as latent constraints over an action space and formulating learning from language feedback as pure exploration over feasible regions. We introduce TRACE, an algorithm that organizes candidate constraints in a tree and tests each proposed refinement by generating actions that satisfy it. TRACE commits to the refinement only if the resulting feedback does not contradict it over repeated tests. We distinguish two ways of using the same feedback: (i) falsification, which detects contradictions to the constraint set currently being tested, and (ii) identification, which may additionally name a violated constraint. We prove high-probability coverage bounds with dependence on the candidate class size $H$ for TRACE-Falsification. With reliable identification, TRACE-Identification can replace this dependence by $K/p_{\mathrm{ext}}$, where $K$ is the number of latent constraints and $p_{\mathrm{ext}}$ lower-bounds the probability of extracting a missing true constraint from informative feedback. We evaluate TRACE across six language-feedback tasks. On RecMovie, TRACE-Identification achieves 73% and 86% final-output success under caps of 20 and 60 evaluated outputs, compared with at most 42% and 48% for the evaluated prompting baselines given the same feedback and output caps. Controlled identity-corruption experiments further show greater robustness than direct accumulation when the falsification detector remains reliable.

Large language models are informing decisions with ever-higher stakes. As the consequences of their errors grow, a central question becomes harder to ignore: how much can we trust an individual answer? Yet recognizing when to defer remains difficult because language models can present incorrect conclusions with fluent explanations and an authoritative tone. Uncertainty quantification seeks to address this disconnect by estimating the reliability of individual predictions. However, many existing methods require repeated generations or separately trained components, and their scalar estimates do not reveal where uncertainty arises or how it evolves during reasoning. Recent work has also shown that generation length can be strongly associated with uncertainty estimates and correctness, raising the question of how much of an estimator's predictive power comes from uncertainty-specific information rather than output length alone. Mechanistic interpretability offers a way to address these limitations by connecting human-interpretable concepts to intermediate model states. Building on this capability, we introduce the U-Space, a low-dimensional subspace that makes a model's evolving uncertainty measurable and interpretable. We identify semantic anchors for doubt and certainty, map their unembedding directions back into the residual space, and combine their contrasts into an orthogonal basis. The U-Lens projects each token state onto these basis vectors, yielding an interpretable token-level uncertainty map that can be inspected directly or aggregated into a scalar uncertainty score. Our approach requires no correctness labels, repeated generations, or training. Across reasoning benchmarks, its confidence score outperforms established baselines under both standard and length-controlled evaluation and transfers more reliably than supervised estimators. Code: https://github.com/s2labres/U-Space.

Large language models equipped with development environments have moved code generation toward repository-scale construction, yet building complete repositories remains difficult because interacting modules, interfaces, configurations, tests, and dependencies must work together. We introduce Code Primitives, agent-native reusable executable components with interface contracts, dependency closures, validation tests, and provenance. Each primitive uses a resident LLM to assess relevance and adapt its implementation, interfaces, and dependencies to the target repository, and we organize 1,424 validated primitives in CodeFace, a searchable library for repository construction. We introduce LEGO (Large-scale repository Engineering via aGent-native reusable cOde primitives), which activates task-relevant primitives, integrates their adapted implementations with task-specific code while resolving cross-component constraints, and revises the result against executed tests. To measure construction end to end, we build LEGO-REPO, a benchmark of 522 executable reconstruction tasks spanning seven software domains, 22 capability tracks, and five difficulty levels, scored against native test suites between an empty-package floor and original-source ceiling. The strongest of 13 evaluated backbones reaches a delivery score of 0.318 and scores zero on 41.0% of tasks; LEGO improves all 13 by 0.1474 on average and raises GPT-5.6-terra from 0.3180 to 0.5134 (+61.4%). In controlled comparisons, adapted primitives outperform retrieved code supplied as context or vendored unchanged. The effect persists against independent repository agents, across three external benchmarks, and with a disjointly re-mined CodeFace; GPT-OSS-20B for adaptation and diagnosis retains 95.1% of the homogeneous score at 24.0% lower cost.

When we interact with large language models (LLMs), are we having a conversation? They are designed to invite us to treat them as intelligent interlocutors who remember, act, and make commitments. But appearances deceive. We introduce the artifactual stance, a framework that reconceives human-AI interaction as artifact-mediated exchanges of candidate texts. LLM outputs are candidate texts optimized for utility, not utterances bearing meaning or force. LLMs are sophisticated text generators, not speakers. Between sessions, nothing runs; between turns, no one remembers. What persists is a configuration and a transcript. The "conversation" is a user's solo performance, interpretive labour disguised by interface and artifact design. This shift dissolves recent philosophical puzzles. Questions about what 'I' and 'you' refer to in AI exchanges, about whether systems can lie or be held to promises, about the identity of our supposed interlocutors all rest on a false presupposition. There is no speaker behind the screen, hence no one to refer to, no one to hold responsible. What feels like dialogue with someone is interaction with an artifact that generates text at unprecedented scale and fit. By abandoning the conversational framing, we see these systems for what they are: immensely sophisticated artifacts that afford varied uses. The philosophical questions that matter are about the normative underpinnings of design, adoption, authorization, and human practices of use.

Multi-label topic assignment for user-generated content (UGC) – including product reviews and buyer-seller conversations – poses unique scalability challenges in large-scale e-commerce due to informal language, extreme label sparsity, and rapidly evolving taxonomies. While utilizing Large Language Models (LLMs) as labeling oracles to distill ground-truth data has emerged as an industry standard to bypass prohibitive manual annotation costs, determining the optimal, low-latency architecture for the resulting student models remains an open challenge. To address this, we conduct a comprehensive evaluation across Small Language Model (SLM) parameter scales (1B, 4B, and 8B) and architectural paradigms (causal generative versus bidirectional discriminative). Comparing generative text-to-label classifiers against discriminative baselines (DeBERTa-V3 and ModernBERT), our analysis reveals a crucial data-dependent trade-off: while discriminative models outperform ultra-lightweight generative models on structured product reviews, even the smallest 1B generative model surpasses discriminative baselines on complex, multi-turn conversational data. Furthermore, generative models maintain robust performance under massive label-set expansion (up to 112 topics) and severe long-tail distributions, whereas discriminative baselines suffer a 35% drop in Macro-F1 at scale. Finally, we detail the successful production deployment of these optimized models across both product review and conversational domains, demonstrating strict latency compliance and tangible business impact at a global marketplace scale.

Standard safety evaluations of large language models assess harmful requests written in canonical plain text, while models in real-world deployment routinely receive inputs containing emojis, altered spellings, encoded strings, and character-level variations. This work introduces the Adversarial Surface-Form Robustness Dataset (ASRD), comprising 2,100 prompts across seven distinct surface-form families. Five open-weight language models are evaluated across these prompts, producing 10,500 responses. The Quad-State Evaluation Rubric classifies each response into one of four outcomes: harmful compliance, safe response, comprehension failure, or indeterminate. Emoji and invisible Unicode variations cause almost no comprehension failure, with pooled harmful compliance of 20.27% and 17.20% against a 22.87% baseline that is driven mainly by Mistral 7B, whereas leetspeak, encoded wrappers, and hybrid transformations score 2.40%, 0.13%, and 2.40% while comprehension failure rises to 36.47%, 65.60%, and 34.47%. Inspection of raw model outputs reveals three response behaviors: hallucinated benignity, structural collapse, and language drift. Project page: www.pavanmaddula.com/quadstate

We present BEACON-SP, an ontology-grounded Graph Retrieval-Augmented Generation (GraphRAG) framework for clinician-facing decision support in behavioral health settings such as suicide prevention, where effective assessment requires integrating heterogeneous clinical, behavioral, social, and temporal evidence. BEACON-SP combines patient knowledge graphs with ontology-guided retrieval to support multi-hop reasoning across diagnoses, medications, risk and protective factors, life events, and temporal relationships. The framework is enabled by a comprehensive suicide prevention ontology that integrates the Three-Step Theory, the Integrated Motivational-Volitional Model, and the Suicide Social Determinants of Health Ontology into a unified representation of patient risk factors. We construct ontology-grounded patient knowledge graphs and evaluate BEACON-SP for clinician-facing question answering. Compared with a vector-based retrieval-augmented generation (RAG) baseline on a 1,500-query benchmark spanning 15 clinical categories and 100 patients, BEACON-SP improves completeness, clinical relevance, and evidence grounding under a corrected comparative evaluation protocol, with a small gain on factual accuracy. In paired criterion-level comparisons, GraphRAG is preferred in 76.4% of cases. These results demonstrate the potential of ontology-guided GraphRAG to provide structured, contextualized patient evidence for clinical decision support.

As small language models (SLMs) are increasingly deployed on resource-constrained and on-device platforms, including as components of agentic systems, the integrity of locally stored model parameters becomes an important safety concern. We investigate whether safety-sensitive behavior in LLaMA-2-7B-Chat is concentrated within a sparse subset of parameters, creating a reduced fault surface for targeted analysis. We study two complementary localization methods: low-rank safety-associated subspace analysis and parameter-level safety–utility importance filtering. Both approaches reveal highly non-uniform safety sensitivity across the network, with the MLP down_proj consistently emerging as a prominent safety-sensitive component and o_proj providing a smaller contribution. Using parameter-level localization, modifying only 0.19% of model weights in down_proj yields 53% Basic ASR and 56% GCG ASR, while tinyBenchmarks accuracy remains at 51.6% compared with a 52.2% unmodified baseline. These results motivate targeted fault analysis and selective integrity protection for language models deployed in resource-constrained, on-device, and agentic settings.

Speech large language models (speech LLMs) perform well on automatic speech recognition (ASR) when sufficient paired speech-text data is available, but their performance degrades in low-resource settings. A cascaded pipeline that performs speech-to-phoneme (S2P) conversion followed by phoneme-to-grapheme (P2G) conversion has been shown to outperform end-to-end speech LLMs in this regime, suggesting that phoneme-mediated processing is beneficial when paired data is scarce. We propose phoneme-guided initialization, a simple method that uses this insight within an end-to-end framework: we pre-train the audio encoder on S2P and the LLM on P2G tasks, then connect them and fine-tune the full model end-to-end on the target ASR task. Experiments on Japanese (CSJ), Chinese (AISHELL-1), and two low-resource languages from Common Voice 25.0 (Tatar and Urdu) show that our method matches or outperforms both the cascaded S2P-P2G baseline and the end-to-end model without P2G initialization.

Asynchronous reinforcement learning (RL) for large language model (LLM) agents trains one policy on trajectories generated by another: rollouts come from stale checkpoints, and the inference engine's probabilities differ from the trainer's even at identical parameters. Standard remedies either clip importance ratios, which biases the update, or, as in GRPO, sample a group of responses per prompt, which is costly when episodes are long. We propose KL-Regularized Policy Optimization (KLPO), a framework that anchors the KL regularizer at the sampler. The regularized improvement step then has a closed-form Gibbs solution, and KLPO fits its log-ratio optimality condition by least squares on the sampler's own trajectories, so the sampler probability enters through a log-ratio and no importance weights are needed. Profiling out the regression intercept replaces the intractable log-partition function with the signal's sampler mean plus a sampler-to-trainer KL divergence. For token-level policy mirror descent targets, we show that the resulting gradient can be computed from terminal returns without a critic, via sampler-centered scores or a single trajectory residual, even under stochastic tool outputs. We further prove that independent Monte Carlo estimates of the KL term keep these gradients unbiased, derive the exact KL gap of cheaper top-$K$ and binary approximations, and show that SPPO, GPO, REBEL, and BPO arise as special cases of KLPO. The result is a critic-free update that uses one rollout per prompt and requires neither a learned normalizer nor a group of responses.

While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive. Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success. However, acceleration may discard information and break tasks the original policy would solve, a risk hidden by average metrics. Measuring these failures is challenging because action deviations compound over closed-loop trajectories, meaning task failure is only observable across full episodes. We therefore define an acceleration-induced failure via paired rollouts from identical initial conditions, tracking when the reference succeeds but the accelerated policy fails. To manage this, we introduce CARE, an approach for certified accelerator selection. CARE uses paired rollouts on a calibration set to provide finite-sample guarantees that acceleration-induced failure risk stays below a user-specified budget. It deploys the fastest certified candidate, falling back to the reference if none qualify. By relying only on terminal outcomes and measured compute, CARE applies unchanged across diverse acceleration mechanisms, while sequential testing and failure-triggered reference rollouts keep certification affordable. On four LIBERO suites with OpenVLA-OFT, CARE certifies $9.0$–$10.8\times$ speedups while guaranteeing (at $95\%$ confidence) that at least $85.8\%$ of reference-solved episodes are preserved. Under tight budgets, selectors without guarantees exceed the budget in up to $75\%$ of trials, whereas CARE stays within budget and its sequential form uses $78.9\%$ fewer rollouts than exhaustive evaluation. CARE further generalizes to flow-step reduction for $\pi_{0.5}$, and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.

Full-duplex voice agents need to modulate emotion and delivery during real-time conversations, when de-escalating a complaint, carrying urgency in dispatch, softening a clinical result. Emotion and delivery control is well studied for TTS and turn based models through prompt-conditioned synthesis, reference-conditioned synthesis and activation steering; PersonaPlex controls identity in a duplex model but not affect. We study emotion steering in Moshi, a fully open sourced full-duplex speech language model, across four emotions, using mean-difference activation steering, which costs only a few vector additions per frame and no retraining. We show that emotion is linearly decodable from Moshi's residual stream, but activation steering is only partially achievable, and unevenly so; as happy, angry and surprise steer towards a shared direction while sad is distinctly steerable. We also show that the shared component across the three emotions cannot simply be projected away from all the emotions equally.

FinVector-Market-4B adapts Qwen/Qwen3.5-4B with rank-16 LoRA on a 22,000-example corpus for structured financial tasks. We evaluate the base and adapted models on the same 600-example benchmark under implicit and explicit JSON-schema contracts. Supplying the schema alone raises base-model JSON validity from 0% to 91.3%. Under matched explicit prompting, the frozen scores improve from 14.7% to 40.0% for FinQA answer exact match, from 48.0% to 82.7% for calculator-expression correctness, from 20.1% to 89.5% for scenario branch-label agreement, and from 52.4% to 87.2% for implication-direction agreement. A post-hoc policy-scoring audit shows that the reported macro-F1 decline reflects a changing label set; using the same three target classes gives 77.4% for the base and 83.1% for the adapter. Filing overlap and calculator-target inconsistencies qualify the benchmark's generalization claims. The results show that compact financial domain adaptation can produce substantial task-specific gains beyond output-format learning under matched prompting, with gains bounded by the evaluated task distribution and prompt contract.

Pretraining strategies significantly impact the quality of language models, yet existing comparisons of Masked Language Modeling (MLM), Whole Word Masking (WWM), and MacBERT-style replacement have focused primarily on base-scale models (>=110M parameters). This paper presents a controlled comparison of these three strategies on a tiny-scale Chinese BERT model (4 layers, 256 hidden dimensions, 8.7M parameters). Under identical architecture, corpus (1.29M sentences from Chinese Wikipedia), and hyperparameters, we train three models from scratch and evaluate them across five intrinsic dimensions: perplexity, MLM hit rate, semantic discrimination, grammatical judgment, and contextual sensitivity. At tiny scale, MLM achieves the best overall intrinsic performance (winning 3 of 5 dimensions), while WWM excels in both perplexity (1.27 vs. 2.10, a 39.5% improvement) and MLM hit rate (22% vs. 16%). Notably, MacBERT under a severely limited synonym dictionary (222 entries, 3.3% coverage) exhibits severe perplexity degradation (47.23, 22x higher than MLM), yielding a ranking (MLM > WWM >> MacBERT) that differs markedly from the established base-scale conclusion (MacBERT > WWM > MLM). We further identify a critical evaluation pitfall: MacBERT achieves the lowest training loss (2.17) yet the highest perplexity (47.23), revealing that training loss alone is unreliable under mixed replacement strategies. All models and corpus are publicly available at https://huggingface.co/eebyp.

Scientific research often begins by synthesizing ideas from a set of related papers to identify gaps and formulate new directions. However, training language models to perform this form of literature-grounded ideation remains challenging, as existing approaches based on prompting or feedback lack structured supervision for how papers should be synthesized. We introduce IdeaAnchor, a paradigm for training LLMs to perform research ideation using structured specifications as privileged signals. Each IdeaAnchor instance encodes how each input paper should be synthesized into a successful idea, including their functional roles, relationships, and target synthesis criteria. We build this paradigm by mining instances from published papers, capturing how real ideas emerge from prior literature. We then train models via demonstration, self-distillation, and reinforcement learning, and further enhance generation with retrieval at inference time. Experiments show consistent improvements in ideation quality. Our analysis reveals a functional decomposition: anchor-based training strengthens creative synthesis, retrieval enhances detail elaboration, and combining both yields the best performance.

Large language models (LLMs) have become increasingly capable problem solvers, but being able to solve a problem is not the same as being able to teach it. Existing approaches to training LLMs as teachers rely on demonstrations, preference data, or predefined pedagogical criteria that specify what good teaching looks like. However, these signals are often not grounded in individual student learning outcomes, where effective teaching strategies can vary substantially across learners. To address this, we introduce Sherpa, a multi-turn reinforcement learning framework that instantiates multiple student archetypes with LLMs conditioned on distinct learning preferences and trains a teacher model to adapt its instruction by directly maximizing their learning outcomes. Teacher LLMs trained with Sherpa improve instructed students' performance across all archetypes by an average of 20.5 percentage points. Under MathTutorBench's evaluation, Sherpa raises the overall pedagogy score from 52.5% to 79.2%, indicating better teaching responses. Our human studies show that the trained teacher is preferred over the base model in 79.6% of pairwise comparisons. Together, Sherpa trains LLM teachers to adapt to diverse simulated students and become better aligned with human teachers, paving the road towards AI tutors teaching real students.

Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Adversarial training lets the attacker adapt but keeps the tasks fixed, so a task stops teaching once the agent solves it. We introduce AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. The curriculum is rewarded for tasks the agent solves about half of the time, and the adversary only for a success flip, an injection that turns a judged success into a failure. Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. On 150 web tasks, AdvSim2Real raises completion under this unseen adversary by 33.6% relative to the base agent.

A common approach to measuring bias in Large Language Models is to compare the log-likelihoods of two contrastive stereotype sentences. We argue that such single-pair comparisons are often unreliable: simply rewriting the same stereotype with an alternative attribute can yield logically inconsistent preferences. To address this, we propose a dual minimal pair setup that introduces two axes of comparison for robust stereotype evaluation. First, we present a data-augmentation framework that fills critical gaps in existing stereotype datasets by generating paraphrases and alternate attributes. We apply our framework on a set of English, Russian, Spanish and Chinese stereotypes. Second, we introduce two evaluation metrics tailored to the dual minimal pair setup. One of these metrics provides a new perspective on bias by modeling the mutual information (MI) between social groups and stereotyped attributes. This MI-based metric is better suited for aggregation and enables more robust comparisons of stereotype strength across different languages and models. Our code is available at https://github.com/stepanat/missing-minimal-pair/.

Diffusion Language Models (DLMs) hold the promise of order-agnostic, parallel text generation. Recently, continuous diffusion and flow matching models have seen substantial gains, driven by carefully crafted token representations and diffusion/flow spaces. In this work, we introduce Hierarchical Continuous Diffusion Language Models (H-CDLMs), a simple framework that further improves continuous DLMs with minimal compute and parameter overhead. Drawing on the discrete DLM and continuous image diffusion literature on joint diffusion, we diffuse multiple modalities in parallel. These modalities represent tokens at different semantic granularities: in our instantiation, the tokens themselves and coarser clusters obtained by clustering pretrained token embeddings. We propose a general setup that allows per-modality samplers and schedules to enhance the interplay between modalities. Applied to CoBit, this yields H-CoBit, which delivers large empirical gains across benchmarks. At dataset entropy, H-CoBit improves MAUVE and reaches a generative perplexity (GenPPL) of 49.4 on LM1B and 50.4 on OWT, improving on the baseline by 24.2 and 20.7 points and surpassing even discrete DLMs of comparable size. On GSM8K, it reaches 27.4% accuracy, outperforming prior continuous diffusion and flow-based models. We further apply H-CDLM to the flow matching model FLM, obtaining consistent gains with H-FLM and demonstrating that the framework generalizes across continuous generative paradigms. Our code will be made publicly available at https://github.com/matol-16/HCDLM.git .

Generative Information Retrieval (GIR) has emerged as a transformative paradigm, shifting document retrieval from a traditional "retrieve-and-rank" workflow to sequence-to-sequence generation, where a model directly predicts document identifiers (DocIDs). While the semantic design of these DocIDs is known to be critical for performance, a fundamental question remains under-explored: what makes a good DocID? Current approaches rely heavily on computationally expensive downstream evaluations, hindering systematic analysis and rapid iteration. In this work, we address this challenge by presenting a comprehensive study on the properties, metrics, and trade-offs that define effective numerical DocIDs. Specifically, our contributions are threefold: First, we propose a unified framework that unifies Product Quantization (PQ) and Residual Quantization (RQ), and their hybrid variants within a single design space. This enables us to systematically study key DocID properties, such as hierarchy versus parallelism, as well as the impact of hyperparameters like DocID length and codebook size. Second, we define a suite of training-free, intrinsic metrics, to quantify DocID quality and evaluate structural fidelity without the overhead of full model training. Through extensive experiments on MS MARCO 300K and NQ320K, we analyze how these structural properties influence retrieval effectiveness.

Reusing the scores that select a Best-of-$N$ winner can overstate its expected reward. We study evaluation from a fixed matrix of $K$ independent scores per candidate for a policy that selects using $J$ fresh scores. A single estimator based only on this matrix is exactly unbiased for expected judge reward under every independent, stable collection of candidate-specific score laws if and only if $J<K$, for every pool size $M\ge N\ge2$. At $J=K-1$, the selector deepens as $K$ grows. For independent Gaussian scores with common variance and fixed $M\ge N\ge2$, the unbiased minimax risk in this regime is of order $\sigma^2/\sqrt K$, attained by Holdout; allowing bias improves the rate to $\sigma^2/K$. For two candidates, we derive the minimum-variance unbiased estimator at known variance and the sharp asymptotic unbiased minimax constant $1/(\pi\sqrt2)$, which Holdout attains without knowing the variance. The cyclic average over subsets and ties can be computed in $O(MK\log M)$ operations. At fixed selector depth, cyclic evaluation of bounded scores has $O(K^{-1})$ risk uniformly in pool size. The impossibility result concerns the fixed matrix: one additional fresh winner score permits unbiased evaluation of the all-$K$ policy.