Pipelines that let a large language model (LLM) make discrete decisions need to know which decisions they may act on unattended, and a record that explains each one. We study a small, fully specified reconciliation contract: commit the target model's greedy answer (the anchor), draw one or two greedy checks that reach the answer by a different route (an alternate prompt, or a model from another family), and let a pure function return one of five outcomes, of which only Agree is acted on. On 48,000 model-item evaluations over 6,000 questions, with every method given access to verbalized confidence as a tie-breaker, the main finding concerns where calls come from: pooling another family's samples raises semantic entropy's AUROC by 1.7 to 4.0 points at every budget, and a same-model temperature check is the weakest design. As complete pipelines, the point estimates of the three-call contract and pooled semantic entropy over 3 or 5 samples differ by less than one point, with intervals containing zero. As detectors of the anchor's errors, the contract is ahead of pooled entropy with 3 and 5 samples under both tie-breaker conventions; with 10 samples the point estimates differ by less than one point and the sign depends on the convention. Under the strict Agree gate it automates 70% of items and still acts on 16% of errors. What the contract adds is a fixed committed answer, an explicit outcome and a decision replayable from its record; generation itself was not reproducible on the hosted API.
Preference training is the standard recipe for reducing object hallucination in vision-language models, and each new method reports a lower CHAIR score. We ask what that reduction consists of. Every trained system is compared with its base model's caption truncated to the same object recall, which removes the part of the gain that comes from saying less; the remainder is compared with what a deliberately weak, frozen, eleven-feature post-hoc probe fitted on 500 annotated images can reproduce. All criteria were specified in the analysis code and frozen before the outputs they judge existed. For eight public methods on LLaVA-1.5-7B, the probe reproduces at least two thirds of what training removes beyond saying less for 7 of 8 methods on COCO, 8 of 8 on AMBER and 6 of 8 on Object HalBench, with RLAIF-V the consistent exception; a later public method released for both LLaVA-1.5 and Qwen2.5-VL (SENTINEL) meets the same bar on both. On our own DPO checkpoints the accounting holds on Qwen2.5-VL but fails on InternVL2.5 on COCO (2 of 4 training designs). At matched recall, the base model plus the probe beats truncation for 8 of 9 LLaVA-1.5 systems on COCO and 9 of 9 on AMBER, on the two newer backbones, and when realized as actual sentence deletion. The same measure prices training signals: DPO on the model's own samples labelled by the probe or by an object-level CLIP score recovers 0.70 and 0.87 of the beyond-brevity gain of ground-truth-labelled DPO on Qwen2.5-VL (the probe 0.73 on InternVL2.5), both beat length-matched random labels, and CLIP labels buy more than probe labels on COCO. Human judgements show that the sentences the probe deletes are wrong more often than those truncation would delete; a human check of the CHAIR evaluator's precision falls below our pre-set bar, which we report as a limitation.
Speculative decoding has emerged as a critical technique for accelerating Large Language Model (LLM) inference by decoupling generation into a drafting phase and a parallel verification phase. A crucial challenge is therefore designing a drafter that is both lightweight and accurate enough to approximate the predictions of the target model. We present a highly optimized Multi-Token Prediction (MTP) architecture that extends existing backbone LLMs to efficiently produce high-quality drafts. Built around a single shared drafting head, our approach features three core architectural optimizations: target activation injection to anchor semantic planning, zero-copy KV-cache cross-attention to bypass prefill compute and reduce memory overhead, and an efficient clustered vocabulary projection to reduce the cost of full-vocabulary decoding. Testing these improvements with the Gemma 4 model family, our G-MTP shows out-of-the-box speedups for both large server-side setups (up to 3.1$\times$) and smaller models targeting memory-constrained on-device environments (up to 2.3$\times$ on H100 and up to 2.2$\times$ on a smartphone GPU). In addition, we present parameter-efficient fine-tuning recipes for G-MTP that keep the drafter aligned with the backbone LLM under domain shift, requiring up to 1.83$\times$ fewer target forward passes than tuning only the backbone.
Prosodic prominence can largely be predicted from the words of a sentence, but part of it depends on how the speaker reads it. We study how much of prominence language models recover from text, whether they can report it when asked, and what the speech signal adds. On read English speech labelled for word-level prominence, we compare frozen and fine-tuned text models, prompted large language models (LLMs) and speech models on the same words. All text models end within about one point of each other, at best 68.8% on the 3-way task, and largely make the same errors. With the prompts we tested, LLMs do no better than a word-majority baseline, even with 500 examples. Linear probes find more in their representations. Speech models reach 84.7%. We argue that text determines only part of prominence, and that the rest has to be heard. LLMs encode what the text determines, but prompting does not bring it out: they represent more about where the emphasis falls than they tell you.
Routing states in mixture-of-experts (MoE) models can distinguish correct from erroneous reasoning steps. If changing routing also changes a model's step-level verification judgment, it is tempting to conclude that the judgment follows what routing encodes. We test this inference with RouteVerify, a controlled intervention study that replaces selected internal activations from one solution trace with activations from another, while keeping the evaluated input unchanged. Across nine MoE checkpoints, we compare three intervention interfaces: router logits, MoE outputs, and post-MoE residual states. We vary the layers and token positions at which activations are replaced. We measure changes in a verification margin that quantifies the model's preference for INCORRECT over CORRECT. We also evaluate the same input with activations from correct versus erroneous steps to test whether the resulting judgment depends on the source step's correctness. We report four linked results. First, probes on router logits separate correct from erroneous steps in eight checkpoints (AUROC 0.80-0.88). Second, replacing router logits across all routed layers changes the judgment in all nine checkpoints, and by more than a norm-matched random replacement. Third, this change rarely depends on whether the source step is correct. At one layer, routing from erroneous steps and from correct steps produces indistinguishable margins in all nine checkpoints. Under all-layer routing, their difference is resolved and positive in only one. Fourth, the same difference is positive for local residual replacements in eight of nine checkpoints, so the measure does detect such dependence where it exists. Correctness is therefore readable from routing, and routing replacements move the judgment, but the movement rarely follows correctness.
Diffusion language models offer an architectural route to low-budget generation by updating multiple tokens in parallel, which can reduce the number of sequential decoding steps relative to token-by-token autoregressive decoding. However, realizing this advantage at low budgets is challenging, as generation quality often declines when the number of denoising and refinement steps is reduced. Editable diffusion models address this quality loss through post-draft refinement, but useful intermediate results may still be lost before final submission. We introduce CaRR (Commitment-aware Refinement and Recovery), a training-free framework which improves submission quality from a fixed editable checkpoint through updated-context refill and trajectory-based recovery within its existing forward schedule. Token Commitment Remasking (TCR) remasks selected native pure replacements and refills them on the next scheduled edit forward using the updated context. Submission Commitment Recovery (SCR) uses the resulting refinement trajectory to recover supported candidates when the final response does not satisfy task-specific output requirements. Both modules operate within the native mask/edit schedule without additional backbone forwards. Experiments on code generation and mathematical reasoning show consistent gains in average accuracy over native ME-DLM across four inference budgets, with the largest gains under tight budgets. At one-eighth of the full scheduled forward budget, CaRR improves the six-metric average accuracy by 7.44 pp, including gains of 11.37 and 10.05 on MBPP and MBPP+.
We evaluate five Qwen2.5-Instruct checkpoints (0.5B to 32B) under four prompting conditions using three fixed tasks per information condition and five generations per task. Two AI raters scored 600 trajectories. Under the coordinated instruction, both raters recorded strict declarations of insufficiency in 11/15 cases at 7B, 0/15 at 14B, and 5/15 at 32B. The runner sent a standard follow-up when a request detector fired. The three checkpoints received at least one follow-up in 15, 12, and 15 cases, respectively. On information-sufficient tasks at 3B through 32B, scored format markers appeared in 60/60 protocol trajectories and 0/60 baseline trajectories. After adjudication, 1/60 in each condition met the diagnostic-execution criterion (H3). Across all checkpoints and prompts, 7/300 information-sufficient trajectories met H3 after adjudication, all on the same task. The prespecified upward transition did not occur in this run. The results separate visible structure, diagnostic execution, and persistence after a request for evidence.
Rare language model behaviors can be difficult to study because obtaining them through ordinary sampling can require a very large number of samples. This motivates the behavior posterior sampling problem: given a language model and a behavior of interest, is there an algorithm that efficiently samples from the model's distribution over transcripts that exhibit this behavior? We introduce PosteriorBench, a benchmark for comparing algorithms by their compute cost and by how close their samples are to the true posterior in forward and reverse KL divergence. PosteriorBench contains 21 tasks on Qwen3.5-0.8B with behavior probabilities from $10^{-5}$ to $10^{-1}$, each with exact posterior samples from a corpus of one billion labeled responses. We find that no method we evaluate is both efficient and faithful: simple methods like prompting are cheap but plateau at a high KL divergence from the posterior, fine-tuning on posterior samples is often cost-prohibitive, and several training methods from prior work do not perform better. Building a posterior sampler that can cheaply approximate the true posterior remains an open problem, and we hope PosteriorBench can measure progress towards this objective. Code and data are available at https://github.com/TransluceAI/PosteriorBench.
In recent years, GPUs have incorporated increasingly specialized support for low-precision computation, from Hopper's FP8 to Blackwell's FP4 Tensor Cores. Precision formats have evolved alongside this hardware, from MXFP4 and NVFP4 on Blackwell to E2M1 FP4 with direct UE5M3 block scaling on Vera Rubin, which we term as $\boldsymbol{VRFP4}$. These advances help serve larger language models at lower cost, but four-bit activation quantization can substantially degrade model quality. We introduce VR-SparseQuant, which uses Vera Rubin's elementwise 2:4 sparsity to recover accuracy while retaining native FP4 execution. By allowing two sparse streams to overlap, we use the same FP4 value budget as dense quantization to refine activations with larger errors. Offline scale-ratio calibration avoids runtime search, while a shared-weight GEMM accumulates both streams efficiently. On Nemotron-3-Nano-30B-A3B, VR-SparseQuant reduces WikiText-2 perplexity from $\boldsymbol{6.85}$ to $\boldsymbol{5.57}$ versus dense RTN W4A4. Including activation preparation, our native Rubin kernels for VRFP4 achieve up to $\boldsymbol{3.1\times}$ GEMM and $\boldsymbol{2.7\times}$ MoE speedup over BF16. To the best of our knowledge, this is the first study of adaptive activation quantization using Vera Rubin's elementwise 2:4 structured sparse FP4 instructions.
Long-term memory is essential for personalized agentic systems that interact with users over long time periods, yet current architectures only capture what happened while leaving the reasoning of why it happened to be derived later by underlying LLMs.%, and silo knowledge within individual users. These events can, however, convey not only what a user experienced but also how that experience shaped their behavioral patterns. We introduce STARK, a dual-graph memory architecture which treats these events not just as nodes but also as edges between behavioral insights. STARK maintains an Episodic Graph which records interactions as temporally, causally, and semantically linked episodes; and an Insight Graph which represents LLM-extracted behavioral signals scored by how many events support them, connected by the very events that led to them. At inference, behavioral knowledge is retrieved directly from insight nodes grounded by event-as-edges, eliminating repeated episodic traversal. The event-as-edge structure also naturally enables projecting the knowledge graph into a relational embedding space where cross-user transfer becomes possible under shared contextual circumstances, allowing the system to infer missing insights for cold-start users. Evaluations show that STARK improves by 1.48 judge score on LongMemEval information extraction and 0.93 on LoCoMo open domain questions over state of the art. On the public LongLaMP abstract-generation benchmark, cross-user transfer improves open-ended judge scores by 0.7 points and ROUGE-L by 3 at 5% graph sparsity.
Large language models (LLMs) can associate places with geographic descriptions, but recovering an observed association does not establish an ability to generalize across geographic space. We distinguish geospatial memory, the recovery of location--description associations observed during training, from spatial generalization, the prediction of descriptions for locations in unobserved regions. We investigate whether continuous location representations improve this generalization and encourage geographic knowledge to become spatially organized within an LLM. We frame this organization as a latent spatial field: a learned mapping from locations to descriptions connected through spatial dependencies. Building on a continuous spatial encoder, we project location representations into the model's embedding space and supply them as spatial tokens. We evaluate the resulting models on two newly curated geo-text benchmarks for Los Angeles and London using two complementary assessments. Geospatial captioning tests both memory and spatial generalization by generating descriptions at observed locations and at locations beyond the spatial coverage of the training observations. A separate geospatial grounding model tests whether generated location--description pairs provide useful additional supervision for mapping text to geographic space. Across multiple LLM families, continuous location representations improve captioning at unobserved locations relative to textual coordinates, while their generated descriptions yield modest but consistent improvements in downstream grounding. Layer-wise probing and representation clustering provide additional evidence consistent with spatially organized geographic representations. Together, these findings show the value of evaluating spatial generalization separately from memory and of incorporating continuous spatial representations into language models.
Facebook reactions record how an audience received a post, a signal absent from the post's text. We treat the six differentiated reactions on 74,241 Arabic food-insecurity posts from public Facebook pages (2020-2024) as compositions, apply a centered log-ratio transformation, and cluster them into four audience profiles: Solidarity (Love-dominant), Dismissal (Haha-dominant), Grief (Sad-dominant), and a flat Ambivalent profile. Dismissal peaks in 2022 (22.3%), the year of the wheat price shock, and grief in 2024 (14.0%), during the Gaza food emergency; a page-clustered GEE confirms both shifts (odds ratios 2.18 and 1.34), and they survive minimum-reaction filtering, alternative zero replacements and clustering methods, and bootstrap resampling. We then test whether the profiles can be predicted without reactions. A zero-shot Arabic LLM recovers Grief for 83.8% of posts but Dismissal for only 3.4%, and in a matched ablation, it shows the reaction percentages do not help. A simple classifier trained on the reaction-derived labels and tested on unseen pages recovers dismissal for 60.0% (macro-F1 0.56 against 0.43), mainly from mentions of officials and institutions. Derisive audience responses are thus learnable from text, yet a zero-shot LLM misses them. We release the corpus, LLM, and human annotation layers, and an evaluation protocol as FoodReact-AR.
Multimodal sarcasm detection requires textual, visual, and contextual evidence, particularly when multiple images provide complementary cues. Multimodal large language models (MLLMs) can generate auxiliary information from a post's text and images, but their outputs may mix useful background knowledge with redundant visual descriptions. We propose SEAF, a dual-branch framework for selective semantic enrichment and visual evidence fusion. SEAF assigns generated fields to distinct roles: entity background enriches semantic input, while image descriptions supply alignment statistics. The semantic branch combines the original post text, background from a frozen MLLM, and optical character recognition (OCR) text. The aligned visual branch combines global and local visual features with statistics of cross-modal and inter-image similarity. Text-Anchored Smoothed Aggregation (TASA) weights images by their similarity to the original post text while preserving contributions from images with low literal similarity. Summing the branch classification scores integrates contextual semantics with visual evidence. Experiments on the multi-image MMSD3.0 benchmark show that SEAF outperforms all compared methods.
People rarely specify an exact color when asking someone to revise a design: they might say \textit{a little darker} or \textit{more blue} after seeing a proposal. A language-model assistant must then infer not only which way to change the color, but how far. We study this distinction with \textsc{ColorRef}, a color reference game in which a model guesses a color from a description and a teacher, knowing a recorded target hidden from the guesser, supplies verbal corrections. On 1{,}000 crowd-sourced description--target pairs, three rounds of deterministic oracle feedback reduce mean target distance from 42.33 to 22.15 $\Delta E_{76}$. However, those corrections provide additional information, and improvement alone does not establish well-calibrated feedback following. A controlled Qwen3-14B diagnostic reveals a more specific failure: following a requested direction and ordering graded phrases do not guarantee a suitable step size; strong modifiers can induce excessive changes. We therefore measure each receiver's typical response to \textit{a little}, \textit{somewhat}, and \textit{much} on separate colors and select wording matched to its observed steps. On 128 held-out starting colors, calibration reduces one-step mean error from 12.36 to 6.86 for Qwen3-14B and from 11.06 to 5.89 for Gemma 4 12B against a fixed graded-wording rule. The benefit is concentrated on large corrections, with many unchanged cases and no consistent increase in threshold success. Our findings show that following the direction of feedback is not the same as making an appropriately sized revision.
Machine translation with large language models requires both high translation quality and low generation latency. Accelerating translation while preserving quality is therefore essential for practical deployment. Diffusion large language models (dLLMs) can generate tokens at arbitrary target positions through iterative denoising, offering a new paradigm for parallel translation beyond sequential autoregressive decoding. However, existing methods still incur redundant computation: fixed target generation lengths waste computation on unused positions, while confidence-only commitment delays medium-confidence tokens whose predictions have largely stabilized. To address these issues, we propose TRAC, a training-free adaptive decoding method for diffusion machine translation. TRAC adapts the target generation length using source length and translation direction, with a conservative fallback to the maximum length for empty or unusually short outputs. It preserves high-confidence commitment while identifying temporally stable medium-confidence tokens for early commitment using cross-step top-1 persistence and approximate Jensen-Shannon divergence over Top-$K$ distributions. On bidirectional Chinese-English translation across FLORES-200, WMT23, and NTREX-128, TRAC reduces model forward calls by 24.9% to 36.8% and achieves sentence-level generation speedups of 1.91 to 2.19 times over Fast-dLLM, with only a 0.02-point decrease in macro-average COMET. These results demonstrate that adaptive generation lengths and stability-aware token commitment improve the efficiency of diffusion machine translation.
Text-based generative recommenders can go beyond ranking existing items by directly generating item titles or other textual representations. In the news domain, this raises a societal concern that models may repeatedly surface already popular articles and thereby narrow the diversity of information to which users are exposed. We investigate training-data memorization as one mechanism behind this concentration. Using approximately 20 million browsing sessions from a Japanese news service, we continually trained an instruction-tuned Llama 3 8B model to generate the next article title in a session. The generated-title distribution was more concentrated and less diverse than the distribution of titles observed in the browsing logs. Frequently repeated target titles exhibited greater verbatim and approximate memorization, and more strongly memorized titles were generated in response to a wider range of prompts. These findings support a pathway in which popularity in behavioral logs becomes repeated training text, is preferentially memorized, and is subsequently reproduced by a generative recommender. An exploratory target-level deduplication condition reduced both memorization and output concentration, while substantially reducing the training-set size and recommendation accuracy. Our results suggest that popularity bias in generative recommendation may arise not only from user preferences or ranking objectives, but also from the model's tendency to reproduce frequently encountered strings, with important implications for information diversity on news platforms.
Recent benchmark evaluations increasingly rely on Large Language Models (LLMs) as automatic judges. This paradigm assumes LLMs evaluate based on the rubric's semantic meaning rather than its numeric scale. We question this assumption by introducing semantic faithfulness: the principle that semantically equivalent rubrics must yield consistent scores. To quantify this, we propose a scale-reversal probe that inverts score anchors without altering evaluation criteria. Furthermore, we introduce floor normalization to calibrate reversal disagreement against intrinsic model variability, enabling fair cross-model comparisons. Experiments across five meta-evaluation benchmarks with seven diverse LLM judges spanning open-weight and proprietary models reveal widespread violations of semantic faithfulness, showing that raw reversal agreement alone is misleading. Finally, we demonstrate that bidirectional anchor averaging reduces dependence on any single anchor direction without compromising alignment with human judgments. Our findings emphasize that trustworthy LLM judges must reliably follow rubric semantics, not just correlate with human ratings.
Retrieval-augmented generation (RAG) is a well-established paradigm for augmenting large language models (LLMs) with information retrieved from external corpora. However, developing effective RAG systems requires substantial engineering effort and iterative experimentation, as effective design choices vary across tasks and domains. With rapid advances in LLM agent capabilities, we ask how well these agents can autonomously build RAG systems and how effectively they can improve them through iteration. In this paper, we propose AutoRAGBench, the first benchmark to evaluate how well LLM agents develop RAG systems from scratch in an open code space. It comprises nine RAG system development tasks. For each task, agents build RAG systems from scratch without starter code or prescribed strategies, under a six-hour budget on a single H800 GPU. They analyze task data, implement RAG pipelines, and iteratively refine their designs based on experimental feedback. We evaluate nine agent configurations and analyze both the effectiveness of the resulting RAG systems and their development trajectories. GPT-6.1 Sol with Codex CLI achieves the highest average score of 56.1. The composite Human-SOTA baseline averages 44.4. These results show that LLM agents can automate effective RAG system development. When starting from a human-designed system on each of seven tasks, GPT-6.1 Sol with Codex CLI improves the score by 16.6 points on average.
Task-specific rubrics are increasingly used to evaluate open-ended language-model outputs, yet constructing comprehensive evaluation criteria from complex source documents remains challenging. We study automatic rubric construction in professional mergers and acquisitions (M&A) due diligence, where evaluation criteria must be inferred from task instructions and supporting documents without access to an expert-authored reference rubric. We introduce a multi-axis generation framework that explores complementary structural and free-form views of the source evidence to construct a broad pool of candidate evaluation criteria. To distinguish failures of criterion discovery from failures of candidate selection, we introduce the coverage-selection gap, which measures the difference between reference-free ranking and a reference-informed greedy diagnostic under a fixed selection budget. Developed on 16 tasks and evaluated on four held-out tasks, our framework achieves 0.876 macro recall against senior-authored reference rubrics under an LLM-based criterion-equivalence protocol. Ablation studies further show that introducing new structural search axes improves recall by 6-10 percentage points, whereas increasing generation volume along existing search directions yields at most 1.5 points. These findings suggest that, once a high-coverage candidate pool has been constructed, effective reference-free selection becomes a substantial bottleneck in document-grounded rubric construction.
Omissions are difficult to verify because missing information leaves no output span to inspect: a summary may be factually supported while still excluding a decision, qualification, quantity, or condition needed to preserve the source's meaning. We introduce COVERAGE-TRACE, a source-side counterfactual framework that evaluates omission as dependence on a designated atomic source fact. It replaces only the target fact with type-matched alternatives, holds the candidate summary fixed, and measures the resulting change in its teacher-forced, length-normalized likelihood. A summary covering the target should respond more strongly than one omitting it, while aggregation across multiple counterfactuals reduces sensitivity to replacement choice. We evaluate COVERAGE-TRACE on 80 matched covered--omitted QMSum pairs spanning several fact types and four open-weight models. Multi-CF achieves 0.764--0.825 AUROC, compared with 0.640 for the strongest evaluated text-only baseline. On Falcon-7B, targeted interventions achieve 0.779 AUROC, whereas shuffling counterfactual assignments reduces performance to 0.533; target deletion is effective while unrelated deletion is not. However, adjustment for meaning-preserving paraphrases reduces AUROC to 0.591--0.624. COVERAGE-TRACE therefore provides a practical, complementary signal of target-specific source--summary dependence, but not a wording-invariant semantic coverage probability. These results establish source-side intervention as a promising approach to omission detection while identifying semantic invariance as the central remaining challenge.