
ALT Hybrid attention models combine full and linear attention to achieve a better quality–efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average score

ALT What makes a small trained map from encoder states to the embedding space of a frozen language model sufficient for competitive speech recognition? The premise of the SLAM-ASR architecture is that a frozen language model possesses linguistic knowledge rich enough to perform the recognition task. This assumption, while reasonable, provides little insight into the mechanisms underlying these models. We therefore seek the concrete internal structures on which they rely. This work shows that SLAM-ASR models rely on induction heads of the frozen language model: the same heads that copy repeated token sequences in text are also important for mapping speech embeddings to transcript tokens. Moreover, we find pairs of heads that behave as a constrained form of differential attention, cancelling attention noise and sharpening attention onto the previously generated word. Finally, we show how induction heads could be used to improve transcription quality under noisy audio.

ALT Continual pre-training (CPT) is widely used to adapt pretrained language models to dialectal and non-standard language varieties, yet it remains unclear when this additional training is necessary, what drives its benefits, and whether aggregate evaluation scores adequately capture its effects. In this paper, we present a controlled study of CPT for Québec French language understanding across nine LLMs spanning different families and scales. Using a budget-matched protocol of 30M tokens per experiment, our results show that the benefit of CPT depends strongly on the starting model's existing competence in the target variety. When CPT is beneficial, the improvements are driven more by corpus composition than by data volume, with literary text accounting for most of the gain. Paired Québec and standard French evaluations further show that CPT primarily improves expression comprehension rather than grammatical competence, and that corpus register determines which regional expressions are a

ALT A central challenge in adapting large language models (LLMs) is capability preservation: an adaptation can achieve its target while degrading general capabilities. We introduce CurvDelta, a scalable framework that incorporates the preservation constraint into the parameterization of the model update. By analyzing the local geometry of the preservation constraint around the pre-adaptation model, CurvDelta constructs a low-curvature space in which parameter updates leave the preservation loss nearly invariant. We make its construction practical for LLMs through Kronecker-factored approximate curvature (K-FAC) and a factorized projection. We also introduce CurvLoRA, a low-rank variant within the same preservation geometry, designed for more efficient adaptation. Both methods apply to preservation-constrained adaptation tasks such as capability learning and selective unlearning. We evaluate our methods on mathematical reasoning, code generation, and knowledge unlearning. For mathematical r

ALT We present Piccolo, a family of multimodal decision models (0.8B to 27B) that coevolves model capabilities and training data via Agentic Data Engineering and recursive self-improvement (RSI). The models predict probabilities over candidate answers for general-purpose text and visual judgments, agent harnesses, data workflows, and interactive games. Piccolo-27B tops the Decision Index Vision Full leaderboard with a score of 75.59, while Piccolo-9B and Piccolo-4B achieve 68.35 and 67.46, respectively. All three rank first on the Private benchmark in their respective model-size categories. Research agents analyze model behavior, construct training data, conduct experiments, and use the results to refine subsequent proposals. Semantic slices and a shared research memory connect these actions across the model family's development. With training recipes comprising hundreds of thousands of examples, data selection and feedback-driven refinement are central to the self-improvement process. Our

ALT We introduce Clover-35B-A3B-Preview (Clover), a mixture-of-experts diffusion language model with 35 billion total parameters and 3 billion activated parameters, supporting thinking-mode generation. Clover adapts Gated Delta Networks to block diffusion by combining recurrent memory of previously generated context with bidirectional softmax attention inside the current block. Historical memory readout competes with current-block tokens in a shared softmax, with learnable block-size calibration and a separate, zero-initialized right-context convolution branch. Clover leads the seven evaluated agent metrics, including 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. On internal agent workloads, it reaches a reported generation speed of 680 tokens/s using one NVIDIA GB300 GPU at BF16 precision and batch size one. We organize the evaluation around model quality, drafting and verification costs, and deployment under constrained memory. We also describe a decision-model variant that

ALT Chain-of-Thought (CoT) prompting improves the performance of Large Language Models by eliciting intermediate reasoning steps. However, CoT is far from perfect: on some tasks, models perform worse than with direct-answer prompting, and on others they generate a few hundred tokens to reach the correct answer. We start by labelling each CoT hidden state as recovering if patching it into the direct prompt yields the correct answer, and failed otherwise. A linear probe shows that the model encodes these two classes differently. We then show that the representation space of the direct prompt holds a direction for each question that leads the model, in a short generation, to the correct answer more often than CoT does. Specifically, we rotate the direct prompt's last hidden state once, before generation, towards a direction built for each question from its own recovering and failed states. Across 25 tasks, these directions raise the direct prompt's mean accuracy from 46% to 85% on LLaMA 3.1 8

ALT This research investigates whether a large language model or an autonomous AI Agent performs better at translating from English to Arabic, which is considered a low-resource language. The authors introduce WikiEn-Ar190, a new dataset of 190 carefully selected Wikipedia articles from 18 different areas, which were created and verified by two native Arabic speakers to ensure that they accurately reflect both meaning and style. Using this dataset, the study evaluates the translation quality of GPT-5-mini both as an independent neural translator and when used within a more complex process that includes self-assessment and improvement. As a standalone model, GPT-5-mini produced very fluent translations, achieving a COMET score of 0.8649 and receiving high human scores for adequacy, fluency, and style (averaging 4.53 out of 5). When used in a system that combined translation with self-assessment, it improved lexical accuracy by 28.76% (BLEU score of 22.83) while maintaining a similar level o

ALT Looped language models deepen reasoning at a compact parameter count by iterating a shared layer stack T times per token, turning recurrence depth directly into sequential decoding latency. Conventional early-exit decoding halts after initial passes without cache completion, truncating cached key-value states of past tokens to shallow depth. We study the opposite strategy, late entry, where new tokens skip initial passes and enter directly into the final K < T passes over a deep context cache. Once lightweight fine-tuning resolves hidden-state norm mismatch at the entry boundary, deep cached context substitutes for skipped token-local passes, allowing late entry to outperform matched early exit across model families and pass budgets. To turn this asymmetry into wall-clock speedups without stalling to build deep caches, candidate blocks verify through only the final K passes on the critical path while previously committed tokens complete passes 1 -> T alongside subsequent steps at zero

ALT Manifold-constrained Hyper-Connections (mHC) extend a Transformer's residual stream into multiple parallel streams with learned cross-stream connections. Prior work trains all streams toward a shared objective. We instead study whether assigning distinct learning objectives to different streams can improve task performance and induce objective-specific specialization. Using complementary lexical (sparse) and semantic (dense) retrieval objectives as a test case, we find that objective-specialized mHC consistently outperforms a non-specialized mHC model across the full sparse–dense fusion range, for both two- and four-stream models and at two model scales. We then examine how specialization changes the model internally using probes of stream representations, residual mixing, objective-specific gradient flow, and read/write gates. The specialized and non-specialized models behave similarly through most of the network but diverge in the final layers: the specialized model develops increasi

ALT Building a small language model requires choosing which published improvements to combine. We compare five changes in the same model, using the same data and training budget. We test selected combinations and repeat each one with a second random seed. Mixture-of-experts layers, Muon optimisation and a Gated DeltaNet hybrid give the largest gains. Extra normalisation also helps, while an attention gate with a matched parameter budget adds little on average. The gains mostly combine, but small interactions between changes appear again in the repeated runs. Repetition also reveals benefits that an initial statistical estimate treated as noise. These findings apply to our small-model training setting; the expert models also have greater total capacity. The practical lesson is to test promising components together and repeat the comparisons before deciding which ones to keep.

ALT Patient history taking is a strategic, multi-turn process that refines diagnostic hypotheses through iterative questioning, mirroring clinical reasoning where each inquiry updates beliefs over candidate diagnoses. Deploying AI in this setting is constrained by privacy, cost, and latency, making Small Language Models (SLMs) a more desirable backbone compared to larger LLMs. However, SLMs capture only the fast, intuitive "System 1" side of this process and lack explicit mechanisms for the deliberative "System 2" reasoning needed to maintain and update diagnostic uncertainty over time. Bayesian Networks (BayesNets) offer a natural complement, providing an interpretable framework for auditable probabilistic belief tracking. Yet, clinical BayesNets are difficult to construct due to their reliance on expert curated structure, dependencies, and parameters. We introduce ProbMedTOD, which automatically synthesizes a clinical BayesNet from a small number of clinical notes in three stages: ontolo

ALT Multimodal Entity Linking (MEL) aims to disambiguate mentions in multimodal contexts by grounding them to specific entities in a knowledge base. A pivotal challenge in MEL is capturing multi-level correspondences: the semantic consistency between mention-entity pairs and the complementary correlations across modalities. However, existing methods often suffer from element dominance due to their reliance on coupled interactions or coarse global aggregations. In response, we propose the Multi-Element Collaborative Interaction (MECI) framework. First, to capture multi-element mention-entity correspondences, we develop a Multi-view Experts Network that leverages a "divide-and-conquer" strategy for decoupled feature learning to mitigate element dominance, supported by a KL-guided routing mechanism that governs expert specialization and collaboration. Furthermore, to model cross-modal complementary correlations, we propose a Hierarchical Multimodal Interaction Module, where a dynamic modality

ALT Recent research suggests that individual neurons can have interpretable roles in AI behaviour. In Qwen3.5-4B, steering along one neuron's output direction increased decisions to conceal information more in teams than alone across five scenarios, even when teammates provided no messages. When all members participated and received steering, collective concealment increased in four scenarios, while inter-agent messaging helped temper concealment under steering. Moreover, removing the neuron's contribution had negligible effects. Collective settings can therefore change how an internal intervention affects AI behaviour.

ALT Judgements of sexism are subjective and may vary across communities, yet sexism detection datasets usually reduce annotator judgements to a single majority label, obscuring disagreement and how models behave on contested cases. We introduce MuDAS, a dataset of YouTube comments in Saudi, Egyptian and Moroccan Arabic, each labelled by three native speakers of its dialect, with the individual annotations preserved. We use it to evaluate Arabic encoders fine-tuned with hard-label, soft-label and perspectivist objectives, GPT-5.6 Terra, and four Arabic-centric large language models. Model errors are more frequent on comments with annotator disagreement. Despite relatively close Macro-F1 scores, GPT-5.6 Terra and the fine-tuned encoders exhibit contrasting error patterns: GPT-5.6 Terra, even with the annotators' definition of sexism, over-classifies gender-related content such as religious advice on modesty, whereas the encoders miss implicit and prescriptive sexism. The Arabic-centric model

ALT Looped Transformers achieve large effective depth with compact models by repeatedly applying a weight-shared block across R recurrent steps, each maintaining a loop-specific key-value (KV) cache. To accelerate sequential loop execution, recent wave-style self-speculative methods like WaveFront Decoding (WFD) and LoopSpec pipeline draft and target tokens across loops in a single step. However, they face two critical limitations. First, they draft only at fixed loops and ignore intermediate prediction changes, wasting computation on stale trajectories. Second, batching tokens from different loops loads up to Rx the KV cache per step, severely amplifying memory traffic in long-context scenarios. To address these challenges, we propose LoopCache, an efficient self-speculative decoding framework for long-context Looped Transformers. Based on the correlation between prediction switches and output confidence, LoopCache introduces confidence-guided dynamic draft updating to refine drafts in fl

ALT Automatic spelling correction must handle both keyboard typographical slips and cognitive or phonetic misspellings where writers approximate a word's pronunciation. Conventional spell checkers rely either on orthographic edit distance and keyboard proximity–which struggle when phonetic errors diverge by multiple characters–or on English-centric phonetic hashes such as Soundex and Metaphone. We present LipSpell, a unified, language-independent phonetic scoring engine paired with swappable language-specific G2P frontends, keyboard maps, and lexicons integrating: (1) a precomputed |P|x|P| phone-distance matrix from universal IPA distinctive features, (2) physical keyboard geometry, and (3) corpus unigram frequency priors, accelerated by a two-stage symmetric-delete and seven-class phonetic-code index. Across six English benchmarks (including 1,000 L2 errors from TOEFL-Spell), LipSpell outperforms Norvig and Hunspell on every dataset and beats Aspell on Norvig1 (78.52% vs. 75.56%), Holbroo

ALT The growing demand for psychological counseling has motivated the development of LLM-based systems that support human counselors, but training such systems requires large-scale, high-quality dialogue data. Authentic counseling transcripts are difficult to collect and share because of privacy concerns, making synthetic dialogues an important alternative. However, existing synthetic datasets, despite covering diverse scenarios and client profiles, often consist of brief exchanges and fail to capture session-level counseling processes. We propose CounselTraj, a dialogue-act trajectory planning framework for expanding brief synthetic dialogues into process-aware, session-level psychological counseling dialogues. CounselTraj learns dialogue-act patterns from role-play counseling dialogues produced by human counselors, plans coordinated trajectories for clients and counselors, and uses them as flexible guidance for long-form dialogue generation. Experiments across three Chinese synthetic cou

ALT Large language models are increasingly used to create stories for children. Since bedtime stories can shape children's perceptions of social roles, this raises concerns about whether LLMs adapt narrative portrayals according to the intended recipient's gender. We conducted a large-scale study to examine gender stereotyping in LLM-generated bedtime stories. Using matched daughter/son prompts and a five-generation design, we analyze 62,873 retained stories from 16 LLMs across 35 country settings and 12 narrator roles. We introduce a multi-level evaluation framework combining context-adjusted lexical rate contrasts, relative stereotype alignment, narrator-recipient interaction analysis, human evaluation, and representation-level interventions to examine both the consistency and interpretability of these differences. Our results reveal systematic differences in narrative framing: son-directed stories emphasize higher agency and lower communion compared to daughter-directed stories. While n

ALT Multi-hop Question Answering (MQA) requires large language models (LLMs) to retrieve and synthesize multiple pieces of evidence into a coherent reasoning path, which is essential for complex tasks. However, prior work on knowledge updating for MQA has largely focused on simple fact edits, overlooking a more challenging and realistic setting—Knowledge Injection—in which the model must simultaneously accommodate both conflicting updates and entirely novel information. Under this setting, a hop-level answer is often wrong simply because the model trusts the wrong knowledge source: it clings to outdated parametric memory when the injected knowledge conflicts with it (K2), and it cannot answer when the required fact is entirely new (K3). To investigate this, we introduce NEXUS, the first benchmark tailored for multi-hop reasoning under Knowledge Injection, which explicitly covers both conflicting (K2) and incremental (K3) knowledge. We further propose MIRAGE, an efficient, training-free fra
