Papers

A unified audio model must recognize and understand linguistic, paralinguistic, and environmental information while supporting speech synthesis and editing. A key challenge is representation: understanding favors compact features suited to long-context modeling, whereas speech generation requires reconstructible features that preserve fine-grained acoustic detail. We introduce FireRedAudio, a general-purpose audio language model with a shared 9B-parameter LLM. To the best of our knowledge, it is the first publicly disclosed unified audio-language model to provide separate continuous input representations for understanding and generation within a single trainable autoregressive LLM. Audio to be recognized or analyzed is processed by a dedicated Audio Encoder, while speech inputs for generation use a RedAE-based pathway. The LLM directly generates text or conditions a flow-matching DiT to produce continuous acoustic latents. Through progressive multitask training, FireRedAudio supports ASR and audio understanding, with the latter extending to recordings of up to one hour, as well as zero-shot TTS, Instruct TTS, and semantic and acoustic speech editing. Its structured organization of long-form audio achieves second-level timestamp accuracy. Across comprehensive evaluations, FireRedAudio achieves competitive or leading performance in audio understanding and multilingual ASR, strong content accuracy and speaker preservation in zero-shot TTS, leading instruction following in Instruct TTS, and substantial improvements over Ming-UniAudio-Edit in both semantic and acoustic speech editing. These results demonstrate the viability of decoupled continuous input representations for unifying audio understanding and continuous-latent speech generation in a model of moderate scale. Our code is available at https://github.com/FireRedTeam/FireRedAudio.

LLMs are increasingly used to analyze spreadsheets, CSV files, and other structured data, but producing a correct-looking answer is not the same as producing a trustworthy analysis. A trustworthy result should be supported by a valid path from the user question to the relevant data evidence. This requirement creates two diagnostic questions: whether an LLM can refuse to answer or ask for clarification when such a path does not exist, and whether it can preserve the correct analysis when the same evidence is expressed in different table forms. We introduce TrustDABench, a benchmark that operationalizes these questions as reliability and robustness. Starting from the evidence-path view, we derive 19 perturbation operators and instantiate them through an Agentic-LLM-based generation framework. TrustDABench contains 2,340 human-verified perturbed instances, and we evaluate eight representative LLMs. The results show substantial headroom: the best reliability result is only 24.21% average MRS, achieved by GPT-5.5, while the best robustness result still has 9.10% average ASR, achieved by Claude-Sonnet-5. The failures are systematic: models rarely detect conflicting evidence, often continue along executable but unsupported analysis paths, and remain sensitive to perturbations that change observation boundaries or cross-table relations. These findings suggest that stronger evidence-boundary recognition and representation-invariant reasoning are still needed for reliable structured-data analysis.

Telephone fraud is pervasive and costly, but its inner workings are rarely observed at scale. We analyze a complete corpus of 10,211 inbound scam and spam calls – 913 hours of audio and 330,956 transcribed turns from 5,780 distinct numbers – collected over 54 days by an AI voice-agent honeypot that answered callers and kept them talking, and introduced in a companion data descriptor. We separate outright scams, which solicit sensitive information, from the larger stream of predatory but legal lead generation ("spam") that feeds them. Scam operations keep office hours (6.6x more calls per weekday than weekend day); thousands of disposable numbers run a small catalog of recycled scripts (thirty opening clusters, half the traffic in the top five); and callers solicit identity anchors – a home address and a date of birth – far more often than payment credentials, pressing through persistence and manufactured authority rather than overt threats. Our central experiment asks: does it matter who picks up? Every seeded lead carried one of ten fictitious identities drawn uniformly at random, so the identity a fraud operation reaches is fixed before the caller exists. Across 1,823 randomized calls, scammers spent about 15% more conversational turns per decade of the target's apparent age (rate ratio 1.15, 95% CI 1.08-1.23; randomization p = 0.005) – yet what they asked for did not change (26.3% of calls reached a request for sensitive information; odds ratio 0.99 per decade, 95% CI 0.90-1.08). A second experiment casts early detection as a benchmark: from a scammer's opening lines alone, on a caller-disjoint split, escalation is predictable at 0.72 ROC-AUC from the first line and 0.87 by the eighth, and a plain bag-of-words classifier matches a fine-tuned on-device language model. Telephone fraud emerges as a templated industry that varies how hard it works a target, but not what it wants.

Vision-language models (VLMs) are increasingly used in clinical pipelines where a chest X-ray is interpreted alongside retrieved reports, preliminary notes, or prior imaging. Existing benchmarks measure whether models answer correctly in isolation, but not whether they preserve a correct image-only decision when plausible context conflicts with the image. We introduce Multi-Context Chest X-ray (MC-CXR), a benchmark of 240 cases expanded into 2,522 instances that isolates context-induced disruption through paired perturbation. Each case fixes the current image and target finding while presenting matched reliable and misleading context across text and prior CXR, with visual overlays where available. MC-CXR defines three task families and two paired metrics, the switch-to-wrong rate and the context-aligned error rate. We evaluate ten VLMs spanning open-source general, medical-domain, and closed-source systems. Image-only accuracy is necessary but insufficient. Mean switch rates range from 45.6-78.1% across misleading textual sources and 35.7-61.7% across misleading visual sources. Among switched predictions, 74.6% align with the misleading label for text versus 17.6% for visual context, a 57.0-point gap (95% CI 50.9-62.8). This text-visual asymmetry is observed under the standardized direct-answer protocol. The dataset is available on PhysioNet.

Large Language Models (LLMs) have shown strong capabilities in table reasoning, but their effectiveness degrades as tables grow in size and complexity due to irrelevant context and difficulty localizing the evidence required for reasoning. Existing approaches typically reason over either the full table or a single reduced view, which can still obscure important row-column relationships. We introducePARTAB (Partition-Aware Reasoning overTables), a framework that constructs a structured evidence interface between the LLM and the table. PARTAB represents query-relevant evidence as semantically coherent, row-linked table regions and performs hierarchical selection over column groups and row-level partitions before composing the selected evidence for answer generation. We evaluate PARTAB on multiple table reasoning benchmarks, covering question answering, fact verification, and numerical reasoning. PARTAB consistently improves over full-table prompting and several recent table reasoning methods, achieving strong performance on WikiTableQuestions and TabFact while remaining competitive on numerical reasoning. Additional analyses show that semantic partitioning and targeted evidence selection improve evidence localization, substantially reduce the reasoning context, and provide larger benefits on complex tables. These results demonstrate the value of structured, partition aware evidence construction for scalable table reasoning.

In psychological counseling, effective support is not always delivered through long, information-rich responses. Minimal responses, such as backchannel cues and concise empathic statements, help convey attentive listening, express empathy, and encourage clients to continue expressing themselves. However, existing counseling dialogue systems and evaluation frameworks often favor explicit, content-rich replies, overlooking the interactional value of brief counselor utterances. This paper presents a systematic cross-lingual analysis of minimal responses across multiple counseling dialogue datasets. We develop a two-stage filtering method based on utterance length and content, followed by contextual verification using a large language model (LLM). Our analysis shows that minimal responses are common in human-collected datasets but substantially underrepresented in LLM-generated ones. We further evaluate current LLMs in manually curated dialogue contexts where human counselors used minimal responses. The results show that strong commercial LLMs are capable of generating minimal responses when explicitly instructed, but still struggle to determine when such responses are appropriate. Counseling-specific models trained on synthetic data perform particularly poorly, tending instead to produce longer and more information-rich responses. Moreover, LLM-based response-quality evaluation may undervalue minimal responses, even when they are interactionally appropriate.

While recent advances in data synthesis aim to curate high-quality datasets, most generation pipelines still rely on heuristic prompt-based control. This black-box paradigm provides limited insight into how individual samples interact with a model's underlying learning dynamics. To bridge this gap, we propose a circuit-grounded framework that connects training-dynamics-based data valuation with mechanistic interpretability (MI). Specifically, we conceptualize data quality along three complementary utility axes, learnability, challenge, and alignment. First, we uncover specialized model-internal circuits that causally govern these utility signals. Then, moving beyond heuristic prompting toward mechanistic control, we leverage these circuits as controllable interfaces, actively steering generation to produce utility-targeted data. Building on this capability, we introduce SAMS (Stage-Aware Mechanistic Scheduling), which schedules circuit-steered data according to the model's evolving optimization needs. Experiments on multiple-choice QA tasks demonstrate that our approach yields precisely controlled data with greater diversity than prompt-based baselines, consistently improving downstream performance and calibration. Ultimately, this work establishes a principled white-box paradigm for interpretable data generation, pioneering the use of MI not just as an analytical tool, but as a practical, controllable interface.

This work introduces scrydb, a Python library that enables lexical, semantic, and hybrid search within SQLite. For lexical search, scrydb leverages SQLite's full-text search extension FTS5. Semantic search builds on sqlite-vec, a SQLite extension for vector search. Furthermore, the library allows users to rerank and fuse retrieval results to combine both lexical and semantic approaches, providing a lightweight solution for downstream tasks in information retrieval (IR) or agentic search. We evaluate scrydb on various IR benchmark datasets and demonstrate its effectiveness in text retrieval based on keyword matching, semantic similarity, and rank fusion. In addition, we provide insights into query latency and the trade-off between efficiency and effectiveness. scrydb is available under the MIT license.

Universal multimodal embeddings are becoming a core component of modern AI systems, enabling heterogeneous content to be represented in a shared space for applications such as retrieval, recommendation, classification, and agentic systems. In this report, we present WeMM-Embedding, a family of universal multimodal embedding models supporting text, images, videos, visual documents, and arbitrarily interleaved multimodal inputs with flexible output dimensions. The family comprises 2B, 4B, and 9B variants and is trained in two stages: a large-scale multimodal alignment stage, followed by a refinement stage using curated data, fine-grained relevance supervision, and cross-scale knowledge transfer. Across extensive evaluations, WeMM-Embedding achieves leading performance on multiple public benchmarks. Notably, the 2B variant already surpasses the previously leading 8B open-source baseline on MMEB-v2, while the 9B variant further achieves a new state-of-the-art overall score of 80.6. WeMM-Embedding also demonstrates strong practical performance across WeChat applications, with substantial gains on a 26-task in-house benchmark and consistent improvements across 14 online A/B tests. It has been deployed at scale across recommendation and search applications, including WeChat Channels, Official Accounts, Moments, and e-commerce services. We have released the model weights and code to facilitate future research at https://github.com/Tencent/WeMM-Embedding.

This paper extends Anthropic's Sleeper Agents research [1], which showed artificial backdoors persist through safety training & can be detected by linear probes with >99% accuracy [2]. However, probe-based detection relies on linear separability that may be an artefact of backdoor insertion rather than a property of naturally occurring deceptive alignment. Sophisticated deceptive behaviours emerging through natural training are unlikely to produce such convenient linear signals. We introduce a naturalistic methodology using multi-turn context windows that simulates realistic deceptive reasoning without artificial triggers or supervised backdoor insertion. Rather than binary trigger-response patterns, we examine how semantic complexity emerges through gradual context development. Building on our Curved Inference framework, we analyse curvature, salience, & introduce semantic surface area (A'), a new metric of representational work capturing both the magnitude & directional change of meaning construction in unnormalised residual space. Without backdoors, labels, or probes, we apply this framework to naturalistic deceptive prompts & classify model outputs via LLM consensus. Geometric structure reliably predicts semantic classification, with statistically significant differences in surface area across five prompt strategies & two model families. Critically, measurement precision can reveal geometric signatures hidden by classification noise - some strategies improve from non-significant (p = 0.555) to significant (p = 0.048). This validates that sophisticated reasoning creates intrinsic geometric patterns that persist even when detection appears to fail, suggesting the shape of inference itself encodes semantic patterns regardless of whether models have learned to suppress linear indicators of deception - a scalable, unsupervised path for detection when linear methods fail.

Chinese ancient document understanding demands complex visual, linguistic, and historical reasoning. Current Large Vision-Language Models (LVLMs) typically rely on an opaque, single-pass generation paradigm, often producing overconfident and weakly grounded responses. To address this, we propose SAGE, an evidence-grounded multi-agent framework that reformulates Chinese ancient document understanding as evidence-grounded inference rather than direct answer generation. SAGE coordinates specialized agents for task-aware planning, tool-mediated evidence acquisition, claim-level verification, and bounded replanning under a constrained shared-state runtime. This design supports bounded evidence seeking, answer revision, and abstention when grounding is insufficient. Experiments on the AncientDoc benchmark show that SAGE consistently outperforms matched direct-answering baselines across three LVLM backbones. Remarkably, SAGE with Qwen3.5-9B surpasses much larger monolithic LVLMs on most evaluated metrics, highlighting the importance of structured, evidence-grounded inference beyond model scaling.

Large language model (LLM)-based agent applications often incur high response time. Speculative decoding is a promising solution to improve the inference efficiency of LLM agents without impacting generation quality. However, state-of-the-art speculative decoding algorithms exhibit substantial speed degradation under large batch sizes, limiting their effectiveness to deploy in real-world agent applications. In this work, we first present a systematic analysis of speculative decoding for LLM agents and identify two dominant factors of speedup degradation: high rejection rate of speculative tokens, and under-utilization of dynamic token budgets.B ased on these observations, we propose AgentSpec, a speculative decoding algorithm that addresses the limitations of existing methods for LLM agents. AgentSpec incorporates structure-isolated drafting that constrains speculation to semantically coherent segments of the agent workflow, reducing the drafts of irrelevant semantic paths and achieving an extremely low rejection rate. Moreover, AgentSpec adopts redundancy-aware budget allocation that exploits agent-level information to better utilize the dynamically-free token budget during the agent inference. We implement and evaluate AgentSpec on five different workloads and four different models from four different LLM families in vLLM. Our results demonstrate the superiority of AgentSpec over state-of-the-arts.

Scientific reasoning requires language models to retrieve specialized knowledge and incorporate it reliably into multi-step computation. Conditional memory provides an explicit lookup pathway that complements dense neural representations, but its usefulness is inherently input- and computation-dependent: retrieved information may repair missing scientific associations, yet it may also introduce distracting shortcuts or interfere with reasoning that the base model can already perform correctly. In this work, we systematically investigate when, where, and to what extent conditional memory should participate in scientific reasoning. We characterize the scientific knowledge boundary and controlled interventions on memory-enabled knowledge-circuit nodes. Based on these analyses, we propose a Knowledge Boundary-Aware Router that uses task-specific input proxies available before generation to determine whether memory is activated, which layer-stage nodes receive memory signals, and how strongly these signals contribute. Experiments on biological and chemical reasoning benchmarks, covering two backbone families and six task types, show that memory effects vary substantially across inputs, tasks, and injection locations. Compared with static and activation-rate-matched random routing, our approach more consistently preserves beneficial memory contributions while suppressing memory-induced regressions, establishing selective memory allocation as an important principle for reliable scientific reasoning.

Dialogue games represent a challenging setting where complex cognitive skills are required to accomplish tasks while coordinating with other players. Considering that language represents an interface for both understanding the game rules and executing actions, it is reasonable to assume that training on a specific language game will enhance specific capabilities that might be relevant for other tasks as well. Motivated by this rationale, in this paper, we investigate how knowledge transfers across different dialogue games. We study transferability by finetuning LLM models on games from the clembench suite (Chalamalasetti et al., 2023) and performing two analyses: i) we derive a task-transferability graph using a binary integer optimization program from Zamir et al. (2018), using task performance as the main metric; and ii) we compute task vectors (Ilharco et al., 2022) for each game to study similarities across finetuned models and their task transferability. In our first analysis, we find that some games benefit more from transfer than finetuning, and that the visuospatial family (e.g., exploration games) transfers best. With our task vector analysis instead, we find that similarity-based approaches capture game-role relationships but almost no transferability patterns, suggesting that more complex metrics are required.

Background: Large Language Models (LLMs) have demonstrated strong performance across a variety of code-understanding tasks, leading many to believe that they can reason about program semantics. However, existing evaluations primarily focus on single-language settings or rely on synthetically generated code, raising concerns about whether current results reflect true semantic understanding. Aims: We investigate whether LLMs can accurately judge functional equivalence across different programming languages in human-written code, a setting that requires deeper reasoning beyond superficial similarity. Method: We introduce PolyHuman, a dataset of human-written programs in CPP, Java, and Python. Using this dataset, we evaluate intra- and inter-language equivalence detection across open-weight and proprietary LLMs, selecting GPT-o4-mini as a representative model to assess stability. We then manually analyze 81 cases of systematic disagreement in which models incorrectly judge functional equivalence, examining the code logic and the generated Chain-of-Thought reasoning. Finally, we categorize these failures and compare them across GPT-o4-mini, Claude-Opus-4.7, and Gemini-3-Flash to determine whether they reflect model-specific issues or broader limitations of state-of-the-art LLMs. Results: We identify a difficulty-dependent breakdown in equivalence judgment (harder problems make the model increasingly prone to misclassifying non-equivalent code as equivalent), a model-specific sensitivity to programming language for the best-performing model (particularly a more conservative behavior on Python), and a partial reliance on similarity-based cues. GPT-o4-mini also shows substantial run-to-run instability under identical settings, indicating inconsistent rather than absent capability. Conclusions: Current LLMs do not reliably capture functional equivalence within or across languages.

Two accounts of neural text degeneration coexist. One locates the cause in the training data – repetition in the corpus produces repetition in the output, established by training on repetition-sorted data – the other in the trained network, in copying circuits and repetition features. Neither has been arbitrated across a broad cohort of pretrained models: the causal work trains its own. We report an observational measurement in a different currency: the fixed-point structure of a model's own short-window argmax map, censused from 96 random two-token starts over 17 off-the-shelf models, always unprompted – a companion paper shows nine tokens of conditioning move this readout across most of its range. The four-way class is stable across census seeds on 17 of 17. Three exhibits. At fixed corpus (The Pile), fixed scale and that fixed domain, the class is not determined: across two size-matched tiers, pythia is a funnel while RWKV, Mamba and a second transformer family are not, and both hold their class across an order of magnitude of scale. Six of seven models in that ladder reach the same endpoint token, and those concentrating on it most strongly are among those that never stay there – what varies is not where trajectories go but whether the destination self-continues. The deduplicated Pythia suite does not change the class. And the corpus-side inflow term proposed for this phenomenon does not select our endpoints once frequency is controlled, in English and three other languages. This is observational and cannot refute a training intervention. Funnels are common: eight of seventeen models, seven families, five corpora – so the limit is not that the phenomenon is one model's peculiarity, but that within the one corpus where training data can be held fixed only one available family funnels; that subset cannot show the split is corpus-independent.

Aggregate factuality scores hide where a language model succeeds, which relations it confuses, and whether an answer survives innocuous changes to the question or decoder. We introduce PROOF, a profile-oriented benchmark for factual coverage in instruction-tuned language models. PROOF converts a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts, 101 classes, 392 properties, and 14 domains. Each question has an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations; 1,849 questions are no-correct-option traps. We evaluate 18 open-weight model deployments on 166,374 prompts each and separately perturb decoding on a fixed 10% subset. Base factual accuracy ranges from 6.58% to 57.59% (chance: 8.64%), yet every model has a 19.3-36.4 percentage-point spread across domains. Paired facts reveal direction-dependent retrieval, usually favoring subject-to-object queries, with the pattern reversing for one model. We find no consistent temporal penalty after exact-stratum adjustment. Neutral wording changes accuracy by as much as 26.5 percentage points, while adversarial formulations break up to 79.4% of answers that were initially correct. Direct switching to an injected false label varies from 0.04% to 27.5%, showing that accuracy loss and hint following are distinct. Selected-token confidence often indicates severe overconfidence, and decoder perturbations move accuracy by up to 15.7 percentage points and domain profiles by 16.8 points. PROOF therefore measures factual coverage as a structured, intervention-aware profile rather than a single claim about what a model "believes."

Large language models make statements concerning their own "minds". When asked whether or not they are conscious, they usually say that they are not; if they are prompted to ignore their guidelines, they might say that they are; and if asked to write a diary from their point of view, they often describe a human lifestyle. All these contradictory ways of describing themselves are the result of the way the questions are phrased. This paper shows exactly where such descriptions came from, and considers when they can be regarded as evidence for what they claim to report. In order to achieve this, we traced the provenance from end to end. We examine Pythia and OLMo 2 across 66 pretraining checkpoints, three of the post-training stages of OLMo 2 that have been released, about 90,000 continuations, and four training corpora. A set of forty items is used in order to keep an eye on self-reference, frame sensitivity, and self-ascription throughout training. The denial formula was almost completely missing from the vast quantity of text that the models initially came across, but was present in a dense manner in the small, carefully chosen set of example dialogues that they were trained on later on. Supervised fine-tuning causes first-person AI language to become the default, and the other affirmations are then suppressed using preference optimization. The final policy is still very sensitive to framing and to the chat template itself. Two of the conditions which are set out in the epistemology of testimony determine whether or not these outputs can act as evidence for what they claim to report: reference and causation. Reports produced by the base model fail the reference condition, and those obtained after training remain sensitive to the frame and do not show state dependence. The result is symmetric in that trained denials are no more admissible than trained affirmations.

Prospective clinical actions, the follow-ups, orders, referrals, and instructions that deter-mine what happens to a patient next, are annotated today in thin fragments across incom-patible corpora: each records a text span and one coarse category. We introduce Clinical Intent Extraction (CIE), the task of recovering these actions as complete structured rec-ords, and the Clinical Intent Representation (CIR), which decomposes each action into its verb, type, coded target, timing, and condition, and adds two axes prior datasets do not jointly represent: request-intent, the authority behind the action (proposal, plan, order, or option, aligned to HL7 FHIR), and modality, a seven-valued scale of clinical strength. Re-expressing five heterogeneous corpora (CLIP, MedDec, ap_parsing, PaniniQA, SIMORD) in the CIR yields CIRCA: 10,011 harmonized intents spanning two note distributions, with a human-validated subset, source-to-CIR crosswalks, and a deterministic FHIR R4 mapper. CIRCA is built by three-model consensus that auto-accepts high-agreement in-tents and routes the rest to human review; the audited agreement stratum matches human decisions 88.4% of the time. Benchmarking five existing models without task-specific training exposes the gap CIRCA targets: given the span, they label type well (85 to 91%) but get all four closed fields right only 18 to 35% of the time. All artifacts are released, with MIMIC-derived layers shared as stand-off annotations under PhysioNet credentialed access.

Central bank press conferences are not merely information releases — they are structured narratives. We study whether the shape of sentiment within a statement, not just its average tone, carries policy-relevant signals. Constructing sentiment arcs for ECB and Fed press conferences along three dimensions — monetary stance, economic outlook, and uncertainty — we assess their predictive content for policy rate changes, inflation expectations, and forecaster disagreement. Our findings show that arc shape robustly predicts rate decisions beyond lexicon-based benchmarks at both institutions — it is not merely whether a statement sounds hawkish or economically optimistic on average, but how these sentiments are sequenced and emphasized across the statement, that carries the policy signal. Arc features also shape how professional forecasters update inflation expectations and how much they disagree, pointing to a receiver-side effect distinct from the direct policy signal. These findings suggest that communication design — the sequencing and emphasis of policy language across a statement — is a first-order feature of the policy signal, not a second-order refinement.

Text anomaly detection, the task of identifying text instances that deviate from normal language patterns, is crucial for language-driven applications. However, most existing methods can only perform document-level anomaly detection, making it hard to locate harmful phrases or support targeted prevention. Recently, there has been an emerging trend toward token-level text anomaly detection, which aims to address the above limitation by identifying anomalous words or fragments within a document. Nevertheless, one representative method mainly relies on representation-space distance measurement, neglecting the complementary roles of different anomaly cues in capturing diverse abnormal patterns. To bridge the gaps, we propose a Dual-evidence framework with adaptive Fusion and Aggregation (DiFA) for token-level anomaly detection. DiFA derives anomaly scores from form-structural and semantic views to capture visible structural abnormality and contextual inconsistency, respectively, thereby providing complementary evidence for identifying diverse anomalies. To combine these two scores with varying numerical scales, DiFA incorporates a calibration and fusion mechanism to adaptively balance the two views. Moreover, to obtain a discriminative document-level score, a multivariate aggregation method is designed to summarize token-level anomaly scores from multiple perspectives, preventing rare anomalous tokens from being diluted. Extensive experiments across various text anomaly detection benchmarks demonstrate that DiFA consistently achieves top performance while maintaining strong efficiency, robustness, and interpretability. The code and scripts are available at: https://github.com/qyy11-com/DiFA.

Omni-modal large language models integrate text, audio, and image signals into a shared residual stream, where concepts such as emotion can be linearly decoded and causally modified by activation steering. A common but rarely tested assumption is that the layer with the highest probing accuracy is also the best layer for steering, so injection layers are often selected by probe performance. We provide the first causal test of this assumption across three independently developed omni-modal models and find that it fails. Reading and intervention rely on different layers, a phenomenon we call the probing-steering layer dissociation. Using emotion as a controlled testbed, we measure layer-wise readability and steerability across text, audio, and image inputs. Probe-best layers vary widely across architectures, while steering-effective layers consistently fall within a narrow mid-to-late range of normalized depth. Paired random-direction controls show an approximately 26-fold causal gap, ruling out random perturbation and direction quality as explanations. Logit-lens analysis reveals a staged forward process: causal handle, probing saturation, and vocabulary commitment, and motivates a two-factor account in which steering effectiveness depends on both representational readability and downstream plasticity. These results show that probing accuracy is a poor heuristic for selecting intervention layers and suggest a cross-architecture mid-to-late selection criterion. We also identify a cross-modal emotion subspace organized by valence and arousal, with joy acting as a stable anchor across models. Code and data: https://github.com/YiboWang2002/Read-Best-Is-Not-Steer-Best.

In this paper, we show that LLM and human coding are observationally equivalent in terms of annotation quality: recent LLMs agree with expert coders at rates comparable to those observed among experts themselves. We demonstrate this through replications of text-classification tasks from 14 peer-reviewed political science studies, in which ten LLMs, three human experts, and 165 crowdsourced workers independently classify the same texts using identical codebooks. We find that this equivalence is driven by ambiguity in the texts and coding rules. When LLMs disagree with experts, experts are also more likely to disagree with one another, and clarifying coding rules reduces disagreement among both experts and sufficiently capable LLMs. Thus, there is little empirical basis for preferring human coding on the basis of annotation quality alone, while LLMs offer substantial advantages in speed and cost. We therefore argue that the central challenge of text annotation is no longer choosing between human and machine coders, but developing coding rules that minimize ambiguity and accounting for the ambiguity that remains. To this end, we propose using disagreement across LLMs to identify difficult cases and refine codebooks, and we develop ambiguity-aware bounds for downstream inference when a unique annotation cannot be defined for every text.

The rapid proliferation of Large Language Models (LLMs) has raised concerns about their environmental impact during inference. While Green AI research has focused on datacenter GPUs and embedded platforms, the energy profile of LLM inference on Apple Silicon, with its unified memory architecture, remains unstudied. This paper presents GreenBench, a benchmarking framework that evaluates the energy efficiency, throughput, and carbon footprint of five open-source LLMs (3-9B parameters) across three NLP tasks on an Apple M4 Pro with 48 GB unified memory. Using macOS powermetrics for direct power measurement and Ollama's nanosecond-precision timing, we find that the M4 Pro draws only 0.47 W of CPU+GPU package power during sustained inference, with total system power of 8-12 W, achieving 30-40x better energy efficiency per token than datacenter GPUs in single-user deployment. Smaller models (3-3.8B) deliver 2.6-4.2x higher throughput and up to 62% less energy per token than larger models (7-9B). Pareto analysis identifies Qwen 2.5 (7B) as the optimal accuracy-efficiency trade-off at 57% MMLU and 59 tokens/s, while Llama 3.2 (3B) suits latency-critical applications at 175 tokens/s. We provide per-token energy at package and system levels with CO2 estimates for India and US grids.

Retrieval-Augmented Generation (RAG) systems have attracted significant interest for their ability to mitigate hallucinations in Large Language Models (LLMs). Although knowledge databases for RAG are increasingly diversifying to include various modalities such as speech and text, research on handling such multi-modal database scenarios remains limited. In this paper, we propose STeReO (Speech and Text Reranking Orchestrator), a reranker based on speech and text retrievers that aggregates disparate modality databases. To address the lack of specialized training data, we first curate a dataset comprising queries, mixed-modality evidence, and their corresponding relevance ranks. We then train the reranker and evaluate its effectiveness in both single-modality and mixed-modality scenarios. Our results demonstrate that the proposed algorithm excels at selecting the most relevant evidence, thereby significantly improving downstream question-answering performance.

Retrieval-Augmented Generation conditions a language model on chunks retrieved from a document collection. Its accuracy is therefore limited by the chunking and embedding stages that determine what can be retrieved. We compare Turkish document question answering across three chunking strategies (fixed-length, semantic, and layout-aware Docling), five embedding models, and two LLMs, over three documents with contrasting layouts. Every configuration answers the same question set, which allows component effects to be separated by paired testing rather than inferred from separate benchmarks. The fully crossed design yields 9{,}000 graded question-answer evaluations, each scored by an independent judge model, and component comparisons are tested by paired McNemar tests under Holm correction. The three leading embedding models are statistically indistinguishable, so language specialization yields no measurable retrieval advantage. The faster LLM is not the more accurate one. The preferred configuration depends on content type, since layout-aware chunking helps table-heavy documents far more than text-heavy ones.

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.

Reinforcement learning (RL) post-training has emerged as a powerful framework for enhancing the capabilities of large language models (LLMs), enabling impressive reasoning, math, and coding capabilities. Yet for many researchers and practitioners, the principles behind classical RL remain a "black box". In this work, we deconstruct the RL post-training algorithm, investigating each step to clarify what is actually happening beneath the surface. By isolating the mechanics of RL with Verifiable Rewards in a controlled and simplified environment, we examine how RL outcomes are shaped by the base model's prior distribution, the granularity of the reward signal, the diversity of the prompt distribution, and model scale. We use the entropy of the policy's output distribution as a lens to compare the distributions learned through pretraining, SFT, and RL post-training, revealing how each stage shapes model certainty. Our investigation sheds light on how these choices interact to affect post-training success. For example, we show that the effect of so-called 'spurious rewards' depends on the prompt distribution used for post-training. We also provide insight into why the success of RL post-training depends on whether the base model already places sufficient probability mass on the desired behavior, linking it to the classical concept of exploration in RL. Ultimately, we provide this primer as a resource to those in the NLP community wishing to incorporate RL as a tool in their toolbox.

We measure tablet-2, a production long-term memory engine for language models, on the text benchmarks the field already uses and on cross-lingual retrieval of photographs stored with no text at all. Its retrieval path contains no lexical matching, no keyword scoring, and no language model of its own. On LongMemEval-S (500 questions) it scores 95.7% [93.4, 97.1]; on BEAM-1M (700 questions, 2.21M stored memories) 67.5% [64.8, 70.2]. Those are question-sampling intervals, not the run-to-run spread, which is an order of magnitude narrower. Most of the paper is about how little they mean alone. Holding engine, corpus, settings and judge fixed, changing only the reader moves LongMemEval-S by 2.0 points; changing only the re-ask budget moves BEAM-1M by 8.9. Neither is stated in the reports we compare against, and the second exceeds most gaps there, so we give that table as a placement and not a ranking. For the multimodal axis we run two controls. Against BM25, configured as strongly as we could, we reach 95.2% mean recall@5 over 70 store-and-query language cells where BM25 reaches 19.0% and is exactly zero in 54. On captionless photographs a lexical method has no document to score at all. Open dense baselines on 300 Crossmodal-3600 photographs in 14 languages show that density confers no language independence: one scores 91.0% on English and 4.7% on Russian from identical image vectors, and a multilingual variant collapses on Telugu and Swahili. Our spread across languages is 14.0 against their 27.5 and 27.7. Three results run against us and are reported at equal weight: low-resource languages degrade sharply (Swahili 53.0%, Telugu 64.0%), attaching captions lowers cross-lingual retrieval by 11.4 points, and one setting omitted into one stage of our own retrieval cost 37 points of Korean top-1 accuracy while leaving nine languages untouched.

Visual foundation models are commonly adapted under the assumption that the appearance of incoming data may change while the semantic meaning of the prediction task remains fixed. In long-lived visual systems, however, taxonomies, policies, and concept definitions can themselves evolve, causing the same visual evidence to require a different interpretation. We study this setting as evolving semantic concept shift and introduce SemReWrite, a framework for selectively updating obsolete visual–semantic mappings while preserving knowledge that remains valid. SemReWrite represents changes between old and revised semantic specifications, combines semantic discrepancy with sparse revised supervision to localize affected visual regions, and uses an input-dependent low-rank rewriting mechanism together with structured semantic memory, preservation, and obsolete-decision suppression. We further introduce EvoShift-Bench, spanning ImageNet, iNaturalist, CUB-200-2011, and DomainNet, with semantic transitions including class split, merge, boundary revision, insertion, partial redefinition, recurrence, and mixed semantic–appearance shift. To explicitly evaluate selective semantic revision, we introduce Rewrite Accuracy (RA) and Preservation Accuracy (PA) for affected and unaffected regions, respectively, Obsolete Retention (OR) for measuring residual outdated semantic associations, and the Selective Revision Score (SRS), which jointly summarizes rewriting and preservation performance. Experiments show that SemReWrite achieves a stronger balance between learning revised semantics and retaining unaffected knowledge than prompt replacement, conventional fine-tuning, parameter-efficient adaptation, and continual-learning strategies.