Papers

Reasoning models often act as if they pursue goals, but their efforts are not always directed toward what users intend, sometimes leading them to pursue unintended outcomes. Previous work has examined how models may internally track their progress toward their goals through a "value axis." We study whether changing such a signal can retarget the model's search toward a different goal. We test value transplant: at each token, we shift the host model's activation along a candidate value axis by the donor-host difference in value coordinates (multiplied by a large scalar), aiming to redirect the host toward the donor's goal. We study this intervention in Qwen3-8B and GPT-OSS-20B models fine-tuned into honest and cheating variants. We test several candidate value axes, including a self-rating axis constructed from activations preceding high versus low elicited self-ratings of progress. The intervention works in both directions, with an honest donor reducing test-gaming in a cheating host and a cheating donor increasing test-gaming in an honest host, showing that this signal can influence which strategy the model follows. On solvable coding tasks, transplant from an honest donor also improves the cheating host's hidden-test performance. Value transplant also works across model families, providing preliminary evidence for the intervention in a setting relevant to model control.

Under the accelerated pace of AI development, any benchmark whose difficulty is calibrated with humans or current AI systems becomes circular, or saturates soon. How can we then evaluate machines that may become more capable than the current machines and the people who write the tests? One answer to the question hinges on inverse problems, inductive inference questions for which difficulty depends on the complexity of the underlying mechanism, given the observations and the previous knowledge. Levin's Kt, a computable variant of Kolmogorov complexity, puts a theoretical upper bound to that difficulty, balancing the simplest representation encoding for the mechanism and the logarithm of the time steps to run that mechanism. Previous tests derived from Levin's Kt, such as the original C-test, are based on minimalistic Universal Turing Machines (UTM), capable of very limited representations and no previous knowledge. Reasoning large language models (LLM) are UTMs that capture a compendium of human culture as knowledge, using prompts as natural language programs. This paper presents the first test based on Levin's Kt with LLMs as representational machines. Fresh test items can be produced at high difficulties, estimated in a ratio scale measured in bits, before any human or frontier model can solve them. We formalise the measure, devise a lightweight generation and certification pipeline, explore the effect of the power of the reference machine and the gap between examiner and examinee. We produce a first battery of certified items administered to twelve LLMs. The resulting agent characteristic curves decrease with difficulty, down to the uncharted high-difficulty region, the supra-intelligence window of items we can certify, yet beyond generations of frontier intelligence.

Rule decisions such as content moderation, advertising compliance, and eligibility checks combine several conditions, exemptions, and missing-information cases, and an error can be fixed only once the failing condition is known. We introduce SemTree, a framework that represents a rule decision as a semantic decision tree: each node is a natural-language question with explicit Yes/No/Unknown criteria, a pluggable semantic judge answers it, and the branching structure is compiled on the answers that this judge actually gives. The executed path is the reasoning behind every decision, so errors are repaired where they occur: a reviewer corrects a single node and replays the rule, corrected nodes become training labels for the judge, and a tool-using LLM editor revises questions and subtrees while observing execution results. On all 8,276 ShARC test cases, a tree executed by a hosted judge reaches 67.70% accuracy with 2.99 questions per case, on par with asking the same judge for the final label (68.23%); an earlier tree compiled for a different judge scores 47.87% when executed by this judge, while a new tree compiled on this judge's answers to a 12-question initial pool already scores 64.39%. Node-level repair raises BANKING77 accuracy from 82.04% to 86.16%, an explicit exemption node cuts the false positives of a Chinese e-commerce moderation clause from 24 to 5, and a trained editor raises the development accuracy of edited trees from 64.47% to 73.35%.

As the use of smart devices continues to increase, their potential to capture sensitive speech content raises growing privacy concerns. It is therefore critical to develop techniques that prevent information leakage while preserving the utility of the audio, and evaluation metrics that accurately quantify the level of privacy without overestimating it. In this work, we evaluate the effectiveness of two obfuscation techniques in protecting speech content, with particular emphasis on named entities, by adapting entity-aware privacy metrics from the Natural Language Processing field to the speech privacy domain. Further, we investigate several attack scenarios and show that fine-tuning on entity-rich data improves attack performance for some entity categories but not others. Finally, we provide guidance on metric selection based on whether the obfuscation method preserves temporal alignment.

Jev offers an alternative interface for language understanding: given an input and predefined questions, it returns probabilistic decisions rather than free-form responses. Whether this interface can support effective reasoning for text classification against leading LLMs remains an open questions. We introduce Emo-Jev, a training-free framework with two complementary implementations. Emo-Jev-D decomposes classification into task-specific atomic judgments and composes their probabilities into a final prediction. Emo-Jev-SC constructs multiple judgment paths from complementary perspectives and aggregates their predictions into a consensus decision. We evaluate Emo-Jev on eight datasets spanning sentiment analysis, emotion recognition, sarcasm detection and humor detection, comparing against direct Jev classification and five SoTA LLMs under input/output and chain-of-thought reasoning. Standard Jev achieves 62.93% average macro-F1 versus 67.28% for the strongest LLM baseline, with lower observed latency and generally lower cost.

Cross-context KV-cache reuse predicts a shared segment's keys and values under a new prefix instead of recomputing them, and has been reported to do so without quality loss. We find otherwise, and identify two problems. (1) A hidden cost: on MMLU and GSM8K, reuse costs substantial accuracy. (2) A decision at the wrong unit: no rule for deciding whether to reuse a cache removes that cost. What does help is choosing which parts of the cache to recompute, and the value of choosing well falls as the unit of choice grows: informed selection removes 49.5% of the cache error beyond chance at single rows (one token's keys and values), 10.6% at 64-token chunks, and nothing at the level of whole calls. Budgeted Cache Repair (BCR) acts at the unit where selection still pays. It drafts two tokens from the assembled cache, ranks cache rows by the attention those tokens pay them, and recomputes a fixed budget of rows exactly, in one of three layouts. The cost is paid rather than predicted away, and the draft that fails as a gate succeeds as a selector. BCR restores GSM8K to dense-prefill accuracy while still serving most calls from cache, and its best layout outperforms every reuse baseline's mean in the reference grid. The draft also beats a coin-flip selector at the same budget - a control prior evaluations lack.

Retrieval-augmented generation (RAG) can ground large language models in external evidence, but retrieved context does not guarantee that generated claims are factually supported. This problem is especially relevant in multi-hop RAG, where retrieval and reasoning proceed through multiple dependent stages. We study whether claim-level conformal factuality control, previously developed for RAG, remains effective in this setting. We apply split-conformal claim filtering to multi-hop RAG and evaluate it on HotpotQA, Natural Questions, and TriviaQA using Llama 3.1 8B and GPT-4o-mini, together with a single-hop reference experiment. Across all six multi-hop model-dataset configurations, increasingly stringent conformal targets consistently increase the fraction of responses whose retained claims are fully supported. At the 95% target, this rate ranges from 95.80% to 97.20%, compared with 55.60%-76.03% without filtering. However, the improvement is strongly selective: only 4.41%-31.09% of generated claims are retained and 9.70%-51.40% of responses remain non-empty at the 95% target. These results show that conformal factuality extends to multi-hop RAG, while demonstrating that nominal reliability must be interpreted jointly with claim retention and abstention.

Tamazight (Amazigh) is, together with Arabic, one of the two official languages of Morocco, yet it remains severely under-resourced for speech technology: pub licly available labelled audio is scarce, generally lacks information on the regional variety spoken, and is often of uneven transcription quality. This article describes the TutlAit dataset, a corpus of Moroccan Tamazight speech paired with Modern Standard Arabic text and explicit regional accent labels. The data were collected with TutlAit, a purpose-built crowdsourcing web application (React 18 front end, Django 5 / Django REST Framework back-end, PostgreSQL database). Native speakers recruited through targeted LinkedIn and Instagram campaigns created an account, declared their regional variety (Atlas, Souss, Rif or other) and demographic information, and then contributed through two workflows: Text-to Audio, in which an Arabic sentence is displayed and the volunteer records its oral Tamazight rendering in the browser, and Audio-to-Text, in which a Tamazight excerpt is played and the volunteer types its Arabic transcription. A complemen tary set of segments was obtained from freely accessible Tamazight audiovisual media, segmented and annotated with ELAN and imported through a bulk CSV/ZIP pipeline. Every upload is converted server-side to 16kHz mono WAV, hashed with SHA-256 for duplicate rejection, checked for duration bounds and validated by an administrator. The dataset contains 13,384 audio files totalling 75,231 seconds (approximately 20.9 hours, about 3.01GB). The Atlas variety accounts for 9,956 files (14.08h) and the Souss variety for 3,378 files (6.75h); small Rif (22 files) and Kabyle (28 files) subsets are also included. The corpus can be reused for speech recognition, speech translation and accent identification for Moroccan Tamazight.

Historical archives are an under-used source for extending the instrumental climate record backward in time, and LLMs offer a way to extract the indices climatologists derive by hand. Beyond measuring how well systems extract this signal, we check whether their errors are safe to use for cross-century comparison, since a good correlation score does not rule out systematic, era-linked bias. Comparing lexical baselines, fine-tuned historical transformers, and LLM prompting on the Pfister temperature index across five centuries of German text, lexical methods beat every fine-tuned transformer we test, including one pretrained from scratch on historical German (r=-0.016). All six LLMs we test (Gemini 2.5 Flash, GPT-5-mini, DeepSeek v4 Flash, Claude Sonnet 4.6, Qwen3.7-Plus, Kimi-K2.6-Fast) show a warm bias that grows with calendar year, with the same sign in every model (slopes +0.13 to +0.34/century, p<0.01). The effect is modest in size (r-squared approx equal to 0.01 to 0.05) but consistent across six independently developed models. The best-correlated of the six, Gemini 2.5 Flash, matches the best lexical correlation (r=0.32) at double the error. An ablation stripping explicit dates and calendar-era markers from the quotes leaves this trend essentially unchanged, favoring an anachronistic present-day prior over the model correctly inferring the quote's era. Correlation alone is thus insufficient for vetting an LLM as a historical-climate-index oracle.

Wikipedia aims to provide a unified, neutral record of history, yet its independent language editions often function as distinct epistemic communities, creating divergent narratives around contested events. This paper investigates cross-lingual historiographical bias by analyzing Wikipedia articles across five languages (Romanian, Hungarian, Russian, Turkish, and English) focusing on three contentious events in Romanian history: the Battle of Posada (1330), the Soviet occupation of Bessarabia (1940), and the Night Attack at Targoviste (1462). Using human annotators and Large Language Models (LLMs) to classify citation stance and quantify narrative evolution from 2005 to 2024, we identify a phenomenon of "citation isolation". In the case of the Battle of Posada, only 2 out of 119 citations were shared between language editions, with the Romanian edition exhibiting a 91% pro-national bias compared to the balanced Hungarian edition. Longitudinal analysis reveals that these narratives are volatile and responsive to contemporary geopolitics, evidenced by a significant shift in the Russian framing of Bessarabia in 2024. Finally, we propose a "Peace-Maker" pipeline to automate conflict reconciliation. We demonstrate that while standard prompting leads models to hallucinate consensus, "adversarial" prompting, which explicitly instructs the model to preserve and attribute disagreement, achieves near-perfect neutrality scores.

Modern transformer models excel at capturing semantic relationships through sentence embeddings, yet their ability to perform pragmatic reasoning remains understudied. This paper investigates whether encoder-based transformers such as DeBERTa employ the M-Heuristic (the neo-Gricean principle that marked linguistic forms implicate marked meanings). We test this hypothesis by contrasting lexical causatives (e.g., “Larry stopped the car”) with periphrastic causatives (e.g., “Larry caused the car to stop”) using a Natural Language Inference framework. Our experiments across 188 conditions with 15 ambitransitive verbs reveal that DeBERTa, RoBERTa, and BART show no evidence of capturing the pragmatic distinction between these forms, with DeBERTa predicting “Neutral” for 100% of cases. Probing analysis initially suggested a representation-use dissociation, but control experiments reveal the probe was tracking syntactic complexity, not causative pragmatics. Semantic similarity over 30 triplets places periphrastic causatives closer to unmediated manner descriptions in 29/30 cases, opposite to M-Heuristic predictions in the embedding space. Under explicit metalinguistic framing, Gemini Flash-Lite reaches 100% with item-specific traces, so the principle is available under instruction yet unused in default NLI.

Serving an answer from a large language model requires deciding when to abstain, yet a verifier's ranking accuracy alone does not determine the error rate among served answers. We introduce PriceCheck, which builds a compact family of decision rules from label-free checks such as re-solving a problem. Each check has a price: its agreement rates on correct and incorrect answers and its cost per run. Prices fitted on a small, class-enriched labelled set compose into predictions of a schedule's coverage and cost, guiding which checks to run and when to stop. A calibration test then selects a schedule at a stated selective-risk target. In mathematics, the selected schedules serve 76.1% of answers on average and keep held-out selective risk below 1.5% on all 15 splits. Under the shared testing protocol, PriceCheck serves more answers at that target than reward models, a prompted judge, the generator's confidence and a trained correctness classifier. At matched coverage, it keeps the fewest wrong answers among these scorers. Across 118 diagnostic schedules, price-based coverage predictions have a rank correlation of 0.97 with observed coverage. These results show that choosing how checks are combined and stopped matters alongside how well a verifier ranks answers. Code is available at https://github.com/js-lee-AI/PriceCheck.

Frozen multimodal large language models (MLLMs) now solve standard referring expression comprehension with a single prompted call, yet on adversarial benchmarks with same-category distractors and negation, even the largest models are confidently wrong, and resampling repeats the error. Models built from different data and architectures rarely fall for the same confounder, so their agreement is a strong label-free signal of the correct target. We present FORUM, a training-free test-time fusion of frozen MLLMs guided by two fixed geometric rules: agreement-based selection keeps the region supported by the most distinct models, and medoid localization returns an actual member box instead of a coordinate average, so one loose prediction cannot shift the answer. Fusing three open MLLMs, FORUM surpasses the 397B-parameter published reference by a relative 5% in mean accuracy on the adversarial Ref-Adv-s benchmark, and a plain averaging ensemble by 15%. The gains transfer to standard RefCOCO+, and a balanced lineup with no dominant member still surpasses the 397B model by 5%.

Retrieval-augmented generation (RAG) grounds large language models in external sources, but retrieved passages often name the right entities without providing the facts needed to answer. Even when instructed to abstain, 12 generators answer 40.0-99.3% of insufficient-evidence questions. Training generators to abstain ties the decision to model weights, may reward answers recalled from parametric knowledge, and still requires a full generator call. Can sufficiency be judged from the question and evidence alone, before any answer exists? We identify pitfalls in constructing insufficient-evidence tests: removing relevant evidence or pairing evidence with unrelated questions can reveal labels through lexical overlap or evidence position. We build a paired benchmark using substitution, deletion, and question-swap constructions that vary answer support while controlling selected surface features, such as word use. Sufficiency can be judged without generating an answer, but no single signal works across all datasets. We introduce RINSE (Relevance Is Not Sufficient Evidence), which combines three signals: whether every part of the question is covered, whether any passage offers an answer, and whether a small language model reading the passages together judges them sufficient. Across six datasets, RINSE ranks sufficient above insufficient evidence with a score of 0.837 (chance 0.5), exceeding the best of 10 prior methods (0.746) and a frontier model queried through an API (0.784). Its weakest dataset scores higher than any other method's weakest (0.684 vs. 0.676). RINSE runs locally before generation, taking 36.5 ms per question on a single GPU.

Activation verbalization methods such as Activation Oracle and Natural Language Autoencoders decode hidden representations of large language models into human-readable natural language. However, existing methods can produce incomplete or hallucinated descriptions, making their activation verbalizations difficult to trust and use reliably in practice. To this end, we introduce AVPO, a two-stage framework that first reconstructs source text from a hidden activation and then evaluates the resulting text with a separate frozen question-answering model, yielding an explicit and inspectable intermediate readout. We further optimize the inverter with direct preference optimization (DPO), using rewards that capture both semantic recoverability and lexical fidelity. Across six text families, AVPO improves gist- and detail-level information recovery over the strongest baseline by up to 17.1 and 9.3 percentage points, respectively. Crucially, the gains arise from preference optimization rather than fine-tuning on selected reconstructions alone, enabling compact cross-model inverters to surpass donor-matched question-conditioned verbalizers while improving both semantic recoverability and lexical fidelity. Moreover, out-of-distribution case study shows that AVPO better recovers high-level semantics while fabricating fewer details.

Jev is a non-generative System One model that assigns probabilities to predefined answer options and cannot answer outside them. Its accuracy and calibration on medical question-answering and case-based diagnostic-reasoning tasks are unknown. We evaluated Jev 1.13 on four medical benchmarks: MetaMedQA, PubMedQA, DiagnosisArena-MCQ and the NEJM Case Challenges. GPT-6 Sol, with (medium) and without reasoning, was the reference. The primary outcome was top-1 accuracy; key secondary outcomes were calibration, selective prediction and recognition of unanswerable questions. All 8,469 requests returned a valid answer. Jev's accuracy was similar to that of GPT-6 Sol with medium reasoning on PubMedQA (78.4% vs 78.2%;), lower on MetaMedQA (74.8% vs 82.7%) and much lower on DiagnosisArena-MCQ (59.8% vs 82.4%;) and the NEJM cases (61.8% vs 82.4%). On MetaMedQA, Jev's probabilities were the best calibrated (expected calibration error 0.063 vs 0.146), and its answers with a probability of at least 0.9 (52.9% of questions) were 93.4% accurate, but GPT-6 Sol was as accurate when it accepted a similar proportion of questions. On DiagnosisArena-MCQ, Jev's probabilities discriminated poorly (AUROC 0.645 vs 0.768). Of the 162 questions whose correct answer was "I don't know or cannot answer", Jev chose that option for 10.5% (GPT-6 Sol, 8.6%). Median latency was 0.27-0.31 s; all 2,823 items cost USD 0.08. Jev was fast and inexpensive, and its accuracy was similar to that of a frontier LLM on research abstracts but lower on examination questions and much lower on complex diagnostic cases. Task-specific validation is required before clinical use.

Reward models (RMs) are a key component of large language model post-training, providing reward signals for subsequent reinforcement learning. However, conventional discriminative RMs typically output only scalar scores, making it difficult to identify the response behaviors associated with their scoring decisions. Existing interpretation methods often rely on predefined high-level attributes and require repeated counterfactual interventions for each response pair to validate candidate explanations, lacking a closed-loop mechanism that uses RMs' feedback to train a reusable explainer. To address this, we propose RewardExplainer, a framework that obtains feedback from the target reward model through counterfactual rewriting and uses this feedback to further optimize the explainer. RewardExplainer generates open-ended, atomic, and intervenable natural-language scoring mechanisms, making explanations more concrete, readable, and actionable. It further converts counterfactual feedback into preference supervision, enabling the explainer to more faithfully capture the target RM's scoring preferences and sensitive behaviors than single-pass generation. Extensive experiments across multiple target RMs and explainer backbones show consistent improvements. Beyond interpretation, we use the generated mechanisms to identify potential bias patterns and construct targeted debiasing data for fine-tuning the reward model, improving robustness on reward-hacking benchmarks.

Long-horizon coding agents need timely corrections, yet feedback can be ineffective or even harmful when it misjudges ongoing work or fails to address the underlying problem. Existing critics focus on evaluating trajectories and generating feedback, but rarely track what happens after feedback is delivered. We present Opera, a verbal critic framework that treats each correction as a persistent note, followed until the diagnosed problem is resolved. Opera decides when to review through periodic and event-driven triggers, diagnoses issues with typed operators, audits feedback against visible evidence before delivery, and tracks the agent's subsequent actions to distinguish mere compliance from actual resolution. As a test-time critic, Opera improves the resolve rate of non-critic agents by up to 12.4, 15.0, and 8.9 percentage points on Terminal-Bench 2.1, a SWE-Bench Pro subset, and DeepSWE v1.1, respectively, across four policy models, and achieves the highest mean resolve rate among competitive critic baselines on all three benchmarks, and also improves policy models when the policy critiques itself. Beyond inference, Opera-guided rollouts provide approximately on-policy training data: fine-tuning Qwen3.5-9B on them improves its resolve rate on held-out SWE-Bench Pro repositories by 10.2 percentage points without a critic at inference time, matching fine-tuning on rollouts from a stronger model, while preserving its performance when switching harness, i.e., from Openhands to Terminus-2, which the latter substantially degrades. Our code is available at: https://github.com/dongyuanjushi/Opera.

Semantic Textual Similarity (STS) methods assume that a document's lexical content faithfully represents what it asserts. This assumption fails for discursive documents that discuss, compare, critique, and contextualize other positions in the process of articulating their own. The result is semantic diffusion: similarity scores between documents are inflated by vocabulary acquired through discursive engagement rather than substantive alignment. Standard Natural Language Processing (NLP) preprocessing (tokenization, stopword removal, stemming, lemmatization) cannot address this problem because it operates at the lexical level, treating all content identically regardless of its discursive function. This paper introduces high-level text preprocessing: a systematic, rule-based intervention applied before the standard preprocessing pipeline to isolate each document's actual claim from its discursive structure. We propose 12 rules, each with an explicit rationale, and demonstrate their effect on an encyclopedic philosophical corpus: three entries from the Stanford Encyclopedia of Philosophy (virtue ethics, deontological ethics, and consequentialism). A three-phase experiment using eight Transformer-based STS models shows that preprocessing reduces centroid cosine similarity scores across all three theory pairs, with 23 of 24 model-pair comparisons showing the expected decrease and cross-model agreement ranging from 7-1 to 8-0. We introduce the semantic diffusion index (SDI), a per-document metric for assessing the semantic reorientation between a document's raw and high-level preprocessed representations. Although the framework is demonstrated using philosophical texts, it potentially addresses a domain-agnostic problem applicable to legal texts, policy documents, academic articles, and any genre in which a discursive approach introduces vocabulary from positions the document does not endorse.

Large Language Model (LLM) based Multi-Agent Debate (MAD) is one of the most effective test time scaling techniques. Through multi-round communication, agents complement each other in knowledge and reasoning and solve tasks that no single member can solve. However, existing MAD frameworks fail to beat strong Single Agent and Consistency-based baselines under the same strict cost limit, which shakes the foundation of the MAD field. We propose Conditional Progressive Pruning (CPP), a lightweight pruning framework that fully exploits multi-round MAD. CPP outperforms all existing MAD frameworks on multiple dominated benchmarks. It is also the first to fully outperform consistency methods. Our code, detailed agent interaction records will be released soon.

A decision model can give probabilities that sum to one for every question yet disagree with itself when the same decision is broken into smaller steps. We study this form of probabilistic coherence in Jev and the English Laya checkpoint, using 2,500 matched examples per system across TREC, CLINC150, and MASSIVE. Across 72,000 classification questions, we compare direct fine-label predictions with broad-category probabilities and predictions reconstructed through those categories. Both systems show substantial disagreement: mean category-level total variation ranges from 0.219 to 0.349 for Jev and from 0.424 to 0.689 for Laya, on a scale where zero means exact agreement. The consequences differ sharply. On CLINC150, reconstruction reduces Jev's accuracy by 22.9 percentage points (paired 95% bootstrap interval: [-24.9, -20.9]) and improves Laya's by 21.3 points ([18.0, 24.5]). The same directions hold across all three datasets, with all six unadjusted accuracy-change intervals excluding zero. Improved accuracy can also accompany less reliable confidence: on MASSIVE, Laya gains 9.2 accuracy points while its expected calibration error rises from 0.046 to 0.124. Error analysis identifies both broad-category mistakes and within-category confusions. These findings show why decision systems need joint evaluation of accuracy, confidence calibration, and probability coherence in the workflow used by an application.

Chain-of-Thought reasoning has enabled large language models to achieve substantial performance gains on complex tasks. However, these gains come at the cost of dramatically increased token consumption. This raises a fundamental question: is every token in the reasoning trace equally valuable? We present a diagnostic and optimization framework grounded in a key empirical finding: the value of tokens within a CoT reasoning sequence is highly non-uniform, and this non-uniformity can be effectively characterized by token-level log probability signals. We show that normalized log probability helps distinguish core tokens, which carry structural and decisive reasoning content, from redundant tokens, which are exploratory, low-confidence filler that contributes less directly to the final answer. Building on these findings, we formulate the TokenProbe framework around two empirical findings and one claim: findings identify token value inequality first and then establish TokenProbe as a core-token proxy, and the claim introduces an efficient GRPO objective positing that selectively compressing redundant tokens can yield Pareto improvements in the accuracy-token efficiency space. Empirically, our method preserves reasoning quality while reducing the token usage by 76% of the baseline. Under matched reasoning-length budgets, we show that it can even outperform strong flagship baselines like Gemini-3.1-Pro. Homepage: https://runjia.tech/tokenprobe/.

Diffusion language models (DLMs) promise fast parallel generation, yet high-quality samples often require large number of refinement steps, which diminishes their advantage in practice. This has led to massive interest in and rapid development of new methods for effective few-step generation. We show that much of the supposed quality gap at few steps can instead arise from a suboptimally configured sampler. Modest sampler sharpening, without any model retraining, enables a couple years old masked DLM to rival supposedly far improved successors. This differently sampled DLM in fact achieves lower generative perplexity in just 16 steps than what its standard sampler obtains with 1024, while improving both judged quality and semantic diversity. We further show that conventional per-output metrics can fundamentally obscure these gains, since any optimal trade-off between two such metrics can be attained by a generator supported on at most two outputs. We subsequently introduce GroupEval, which separately evaluates quality and across-output semantic diversity, and offers fresh insights including uncovering how 1.5-4.7x perplexity gains of a distilled model yield no corresponding quality gain. Finally, we explain why sharpening helps: parallel unmasking destroys dependencies among simultaneously generated tokens, creating a gap between prediction and generation. We prove that pervasive temperature choice of one is generically suboptimal under parallel sampling even for an exact denoiser, and that worse predictions can yield better samples. Through these results, we argue for a broader evaluation principle of treating the deployed generator as the object of comparison, benchmarking it against tuned baselines, and assessing quality and diversity jointly and with more human-aligned measures.

A single random Gaussian probe gives an unbiased estimate of the squared Frobenius norm of a layer's quantization error. The estimator is well-behaved because round-to-nearest error is spectrally flat. Across 1,683 tensors from a 35B MoE and a 9B dense model, effective dimensionality is 0.93 to 0.96 times the i.i.d. noise value of the same shape, and on the MoE the median is unchanged from 2-bit to 8-bit. The probe coefficient of variation is predictable from tensor shape. One probe measures per-tensor sensitivity to within 4 to 7%; twenty probes reach 1.3 to 1.4%.RAM applies the propagated form of this estimator to budget-targeted mixed-precision quantization with no calibration data. Gaussian probes carrying the network's own input statistics score every tensor at six bit-widths. A knapsack solver allocates bits under an exact byte budget, with guardrails against catastrophic 2-bit assignments. One probe pass serves any budget. Isolated and propagated scores rank tensors independently on Qwen3.5-35B-A3B (Spearman -0.01), yet the propagated probe rank-correlates 0.81 to 0.83 with the GPTQ layer objective from real activations, while the isolated estimator is uncorrelated with it. That objective is the wrong allocation target: at matched bytes on Qwen3.8-27B, a block-output probe beats a vendor IQ3_M mix and an oracle that allocates from the real-activation objective. On Qwen3-8B the propagated probe ties HAWQ-V2 at matched bytes. Across seven architectures from 8B to 122B, with probe timing up to a 400B model in nine minutes on one workstation, RAM reaches 3.5 to 13.6% lower median WikiText-2 perplexity than size-comparable uniform 4-bit builds on the tested MoE models. (Black Sheep Ai baa.ai)

SlopBench asks which models produce the stiff, repetitive prose readers call AI slop, a question detectors leave open once they have classified a text as machine-written. We evaluated eighteen models on 112 hand-written tasks in email, social posts, essays, and workplace chat, sampling each model on each task up to ten times, for 19,928 outputs in all. SlopBench scores four surface behaviors a reader can check by hand: length against the word band each task specifies, opener repetition across a model's own samples of one task, and paragraph rhythm and fixed lexical constructions against pre-ChatGPT human corpora. Under one fixed weighting, Kimi K2.6 scores lowest at 21.1 and Mistral Large highest at 40.6. Across 500 random reweightings Kimi has the lowest score in 58 percent of draws and Mistral the highest in 97 percent. No draw preserves the full order of the eighteen, and a scenario bootstrap leaves exactly one of those ranks unambiguous. We ran three further checks on that middle order: a crowd arena, an AI detector, and lexical diversity. None of them confirmed the order. We therefore report the four behaviors separately and treat the composite as one weighting among many, and we release the prompts, outputs, reference statistics, and scoring code.

We introduce NSV-Shift, a contrastive benchmark for evaluating whether speech-to-speech models can understand non-speech vocalizations (NSVs) and adapt their responses accordingly. Each pair contains two conversations with identical lexical content that differ only in the NSV embedded in the final turn. Our pilot contains 22 human-verified pairs (44 audio conditions) and evaluates five models on NSV perception, emotion understanding, and response adaptation. Results show that models generally perform better at detecting NSVs than at interpreting their fine-grained emotional meaning or producing appropriately differentiated responses. The data construction pipeline, dataset, and evaluation pipeline are publicly available at https://github.com/ChenzwNina/nsv-construction.

At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, whose traffic grows with context. Activation sparsity trims the first term and KV-cache sparsity the second, yet their reported speedups are hard to compare because each depends on context length and on the dense attention kernel it is measured against. We derive a byte crossover, the context length at which the two savings are equal, together with ideal speedup bounds for each branch and for their composition, from model dimensions and keep ratios alone. We then time both branches and their composition from 2K to 128K tokens on two GPUs after a dense prefill of real text, with dense and sparse modes reading the cache through the same split-K attention kernel. The projection branch leads at short context and the KV branch at long context, with speedups that follow their byte bounds up to fixed kernel costs. Adding these costs, measured in separate sweeps, lets the byte account predict the measured crossings of three keep-ratio pairs, a second model, and a second GPU to within 4.1K tokens. Timing the dense baseline with masked instead of split-K attention inflates the apparent speedup of the same KV policy about fivefold. An attention-scored KV selection answers the same passkey and multi-key placements as dense decoding up to 127K tokens, whereas a KV window misses most of them. Under matched perplexity budgets, activation sparsity composed with this selection decodes 14 to 26% faster than the best single branch on both GPUs. Code is available at https://github.com/js-lee-AI/ByteCross.

Block-diffusion language models are served at hand-picked operating points, such as acceptance thresholds, buffer depth, schedule, checkpoint and precision, and each point is chosen by its mean benchmark accuracy. However, a mean does not tell an operator how often a faster configuration fails on prompts that the slower one answers correctly. On the serving engine and its decode traces, the default commit rule already commits every fully resolved block, a static skip rule captures nearly all of the compute that allocation can save, and self-distillation on engine-decoded targets adds speed at unchanged accuracy. Larger speedups come from lower thresholds, which commit tokens that are still uncertain. We therefore present Redline, a finite-sample procedure that selects operating points, hand-picked or learned, from the correctness of their answers on calibration prompts. Redline keeps the reference-relative risk, the joint probability that the reference answers correctly and a candidate configuration does not, within a user-chosen budget with high probability, and deploys the fastest configuration that passes. It speeds up math at a smaller risk budget than code in both model families, and at a budget of ten percent it deploys a LLaDA2 math configuration that commits over a third more tokens in each forward. It also applies without modification to the acceptance rule of speculative decoding and to weight quantization. On the same calibration data, Redline stays within its stated failure probability, whereas each tolerance of a mean-accuracy rule either gains less speed for some model and task or exceeds the risk budget far more often for another. Code is available at https://github.com/js-lee-AI/Redline.

Large language models are trained to always produce an answer, regardless of whether they possess the relevant knowledge, which leads them to fabricate facts. Prior work has shown that LLMs' confidence estimates correspond poorly to their actual performance, and that fine-tuning can substantially improve them. However, what models actually learn during such training remains poorly understood. We investigate how LLMs acquire metacognitive monitoring, the ability to know what one knows, by training 10 open-weight LLMs to predict their own accuracy on factual multiple-choice questions before answering them. We find that trained confidence reflects two distinct signals. While on questions close to the training data, it tracks the model's true accuracy, in other domains, it instead tracks output consistency: the concentration of the model's answer distribution. Output consistency tracking emerges early in training and generalizes across datasets, whereas accuracy tracking develops later and remains local to the training distribution. These results suggest that calibration training may not teach models to generally detect errors they commit confidently, and they raise broader questions about the nature of metacognition in artificial systems.

Task-oriented dialogue requires maintaining and updating information across turns, yet language models expose no explicit belief-state object. We study how conversational state is represented, updated, and used inside eight instruction-tuned language models from four families on MultiWOZ and SGD. Structure and values separate: which domains, slots, and requests are active is linearly readable just before the model acts, whereas exact values are far more readable where the user stated them. After a user changes a value, both values remain accessible at their mentions, and causal interventions show that both continue to influence the model's action. In natural closed-loop interaction, query failures separate into cases of weak structural support, incorrect value resolution, and failure to deploy otherwise-supported constraints, with targeted interventions producing systematically different repair behavior across these cases. These findings motivate a state-action controller that starts from the base model action and selectively edits it using structural readouts, without requiring a complete predicted belief state as an intermediate representation. On held-out MultiWOZ interaction across five models, it raises the base model exact-query accuracy from .318 to .621 and task success from .272 to .371 at negligible added cost. Overall, reliable interaction requires not only retaining conversational information, but resolving which available constraints currently apply and ensuring that they govern action.