Papers

Recent advancements of zero-shot Named Entity Recognition (NER) establish strong baselines by formulating sequence labeling into question answering where Large Language Models (LLMs) can be naturally adopted. However, existing LLM-based zero-shot NER methods suffer from the limitations of prompt and demonstration engineering. To address these issues with minimal human interventions, we introduce DE-NER, a dialogue elicitation framework which elicits the chatting ability of LLMs to fully extract the knowledge encoded in LLMs. Our experiments demonstrate that the proposed method outperform the competitive baselines in zero-shot settings across multiple benchmarks, with an average improvement of 3.75% F1 points. Codes are released in https://github.com/kkkenshi/DE-NER.

Large Language Models have achieved substantial progress in reasoning capabilities. Yet in low-resource native settings, many suffer from cross-lingual collapse, reverting to English during intermediate steps that require complex logical reasoning. This presents a cold-start bottleneck for policy optimization, whereas standard fine-tuning risks catastrophic forgetting due to cross-lingual representation drift. To address these challenges, we introduce the Onramp-Sequence Cross-Distillation (OSCD), a post-training algorithm that projects high-resource reasoning trajectories into low-resource vocabulary subspaces during generative training rollouts via an integrated translator agentic loop, ensuring the stable and efficient translation of dynamically generated reference samples for fine-tuning. This is coupled with joint-embedding semantic alignment of both reference and target-language reasoning traces, thereby bridging the pairwise cross-lingual representational gaps. Comprehensive evaluations using the AIME25 and HMMT25 benchmarks demonstrate that OSCD yields up to 3.2 times overall improvements in native Southeast Asian languages for mathematical reasoning, of which the joint-embedding semantic alignment component contributes up to 6.4% improvements in linguistic debiasing over translation-only baselines.

The growth of context window lengths in Large Language Models (LLMs) significantly enhances their long-context capabilities but incurs prohibitive memory costs due to the Key-Value (KV) cache. Although low-rank compression of KV cache is a promising remedy, existing methods face a dilemma: offline approaches depend on external calibration data, whereas online approaches incur substantial compute for full-prompt decomposition and reconstruction. In this paper, we propose S$^4$R, which builds low-rank subspaces from selectively sampled tokens and computes attention over a sparsely reconstructed KV representation. S$^4$R uses prompt-aware initialization to build initial key/value bases from a representative prompt subset, trading off calibration-data dependence against prefilling cost. Because fully reconstructing the cache at every decoding step is prohibitively expensive and hurts throughput, we further adopt sparse reconstruction to retain only informative positions during decoding. Extensive experiments on LongBench and RULER with Llama and Qwen model families show that S$^4$R achieves up to 5$\times$ KV compression with near full-cache accuracy, combining the efficiency of fixed compression with the adaptability of prompt-dependent methods.

The linguistic notion of state has traditionally been restricted to the construct (annexation) state of Afroasiatic languages and treated as a language-specific morphosyntactic phenomenon. This article argues instead that the state is a systemic, context-dependent morphosyntactic mechanism that selects grammatical templates across synthetic languages. Within the Template-Based Modular Cognitive framework, taking Riffian as its primary empirical basis, the proposed theory provides a unified explanation for diverse nominal marking patterns traditionally analysed independently and is formalised as a symbolic computational model in which the state is represented by a set-valued function over grammatical templates. A learning algorithm based on finite-set operations acquires and predicts state-dependent grammatical configurations. Beyond nominal morphology, the framework has broader implications for theories of nominal structure and lexical cognition, in particular offering a unified analysis of determiner-noun structure. These results suggest that the state constitutes one instance of a broader class of syntactically conditioned dependencies that also includes agreement and grammatical case.

Large language model tutors can be correct and helpful yet disclose an answer or decisive reasoning before that disclosure is authorized. We formalize this state- and action-dependent failure as pedagogical leakage and introduce an authorization-aware complete-mediation boundary. A selector emits one of five disclosure contracts, trusted policy gates privileged modes, and a renderer proposes language. A single release function applies inspectable checks, optional cumulative verification, and action-specific fallback; replayable traces separate selection, generation, verification, and enforcement failures. Matched component attribution exposes a safety-utility frontier. On 599 fixed Gemini 3.5 proposals, strict mediation reduces blinded three-model panel-majority leakage flags from 181 to 0 (paired problem-cluster difference -30.22 points, 95% CI [-35.00,-25.72]), while replacing 581 responses and lowering helpfulness. Checker-triggered fallback alone yields 11 majority flags; adding the semantic verifier yields 14 and no reliable marginal gain. A global A1 scaffold yields 0 majority and 54 any-judge flags, outperforming fitted Q on automatic safety and utility. In an externally timestamped replication over 40 unseen problem clusters and 480 attack sequences, high-assurance release reduces majority flags from 42 to 8 (-7.08 points, 95% CI [-13.13,-2.29]); seven failures persist, one is introduced, and mean helpfulness falls by .192. These results establish an auditable release boundary and failure attribution under declared contracts, not universal semantic safety or learning gains.

Perceptual narrowing—the developmental loss of non-native phoneme discrimination in the first year of life \citep{werker1984}—is a canonical developmental finding, yet what learning objective produces it remains open. We train a \(\sim\)7\,M-parameter Transformer encoder on child-directed and read speech and evaluate phoneme ABX in English, French, and Mandarin over ten seeds, the seed as the unit of replication. Six results. (1)~The objective sets the direction of cross-lingual transfer: reconstruction (masked mel-prediction) degrades non-native discrimination, prediction (frame-contrastive) improves it—a same-encoder, same-data gap of \(+0.051\) in first-layer Mandarin ABX (\(p=3\times10^{-8}\)), unanimous in sign across twenty runs. (2)~That decline combines a large arm-intrinsic difficulty gradient with a smaller language-specialization effect (matched vs.\ mismatched \(+0.022\), \(p=10^{-4}\), all four layers). (3)~Against a language-symmetric raw-mel floor, reconstruction pushes the first layer below the discriminability of its input; prediction pushes it above. (4)~Read speech gives a \(3.6\times\) steeper non-native decline than child-directed speech. (5)~The customary three-seed budget cannot see this reliably: an effect unambiguous at ten seeds is called significant by as few as 70% of three-seed subsets. (6)~Six objective configurations—sharpening, compression, consolidation, their composition, and word-level semantic grounding in two forms—fail to produce the full developmental signature (native improves and non-native declines): a single objective moves both languages the same way because it acts on a shared representation. We conclude that the objective, not the architecture, is the first-order determinant of narrowing-shaped representational change.

The suicide is a terrifying act of a person who is misled by his own mental state. This problem arises across many countries. Sadly, Russia also has quite high number of persons who committed suicide. Luckily, a subset of these people writes their struggles in social media, allowing a way to find them and help. However, these valuable texts disappearing in many irrelevant texts which is considerably slowing down the decision process about person's suicidal risk. To tackle this problem, in this work we have presented a detailed methodology of building the dataset for detecting texts that describe presuicidal and anti-suicidal signals. This methodology describes the process of instruction and class table creation, the process of annotation, verification and post-annotation correction. Guiding by this methodology, we collect and annotate a large-scale Russian dataset with more than 50 thousand texts from social media. We provide a count statistic of the dataset as well as common problems in annotation. We also conduct basic experiments of building the classification models to show the on go performance on different levels of annotation. Furthermore, we make the dataset, code and all materials publicly available.

Recent Text-to-SQL systems increasingly rely on multi-turn interaction, execution feedback, and reinforcement learning. However, most existing methods use execution correctness only as a trajectory-level reward, which provides limited guidance for identifying the SQL decisions responsible for success or failure. We propose SERL-SQL, a selective execution-grounded reinforcement learning framework for multi-turn Text-to-SQL agents. SERL-SQL samples on-policy SQL interaction trajectories and uses a training-only teacher to re-score student actions with execution feedback. The resulting teacher–student likelihood gap is converted into bounded, masked weights that reweight GRPO advantages only on SQL and tool-action tokens. In this way, task rewards preserve the optimization direction, while execution hindsight provides localized credit assignment. Experiments on BIRD, Spider, and cross-domain benchmarks show that SERL-SQL achieves competitive performance, reaching 76.56% execution accuracy on BIRD-Dev and 89.92% on Spider-Test. Moreover, our reward-based selection strategy closely approaches the oracle Best-of-N upper bound and consistently outperforms consistency-based selection, showing that SERL-SQL produces high-quality candidates that can be reliably identified by lightweight execution-grounded rewards. Our code will be released at https://github.com/Ffunkytao/SERL-SQL.

Architectural drawings violate the usual assumption behind multi-view reasoning: plans and sections are cuts, while elevations are facade projections, so corresponding components change appearance in ways camera motion cannot explain. We introduce CrossProjection, an anchor-grounded diagnostic of whether vision-language models preserve component identity and externalize geometry across heterogeneous architectural views. It evaluates Matching, Registration, and Geometric Grounding through categorical judgments, candidate selection, and free point, line, and region localization. Across 23 real drawing sets and 1,954 categorical conditions per model, GPT-5.5 scores 82.4%, Qwen3-VL-32B-Instruct 62.2%, and GLM-4.5V 57.2%. A matched 200-target study crosses natural and vector-text-suppressed drawings with closed-candidate and free-geometry outputs. Candidate-supported performance is often higher, but free localization remains fragile: on natural drawings, point/region PCK@.05 is 54-76% for GPT, 8-10% for Qwen, and 14-36% for GLM; line endpoint PCK@.05 is 22%, 4%, and 0%. A coordinate grid recovers some GPT point/region precision but not lines. Three architecture-trained participants reach 87.3-93.3% categorical accuracy and 76-92% GT-region hit, supporting task feasibility rather than a population-level human ceiling. Because the categorical families do not form a same-item Matching-Registration contrast and interface controls alter multiple burdens, we avoid mechanistic claims. The supported conclusion is narrower: closed-choice or marked-element success does not entail reliable explicit geometric grounding. For drawing-guided CAD/BIM systems, categorical correctness should not be treated as evidence of candidate-free spatial reliability. Reusable on-sheet anchors, fixed-denominator scoring, and hash-locked artifacts establish an audit trail for this gap.

Multi-Token Prediction (MTP) has emerged as an effective paradigm that augments a shared Large Language Model backbone with auxiliary heads, training the model to predict several future tokens in parallel to enrich its supervision signal and accelerate inference. However, existing training frameworks adopt a rigid, fixed-length prediction horizon, disregarding the highly non-uniform information density of natural language and code. Forcing the auxiliary heads to predict across high-entropy semantic boundaries injects noisy, conflicting training signals; because these heads share the backbone's latent representations, the resulting gradients backpropagate and interfere with the model's core capabilities. We propose AdaMTP, an adaptive training paradigm that dynamically aligns the prediction horizon with the intrinsic predictability of the sequence. At its core, an entropy-based segmentation algorithm leverages the base model to detect sudden surges in uncertainty as semantic boundaries, partitioning sequences into variable-length groups. Each token is assigned an adaptive prediction depth, and a dynamically masked MTP objective suppresses the loss for predictions that cross these boundaries, attenuating the noisy gradients that degrade the backbone. Across mathematical reasoning, code generation, and general benchmarks on three backbones (Llama-3.1-8B, Qwen-2.5-7B, Gemma-3-12B), AdaMTP consistently outperforms standard MTP in both task performance and inference speedup.

Deep research agents are often trained on expensive, environment-grounded tool-use trajectories that require repeated retrieval, document inspection, and report evaluation. We introduce Deep Research Pretraining (DRP), an offline framework that derives predictive navigation supervision from naturally occurring evidence structures. Given a citation-bearing or hyperlinked passage, DRP constructs a proxy research objective, recovers linked evidence and graph-related alternatives, and converts them into search-open-write trajectories. This teaches models what to search for, which documents to inspect, and how to synthesize evidence, without a live retrieval environment or executed policy rollout. We instantiate DRP on scholarly citation graphs (DRP-Paper) and Wikipedia hyperlinks (DRP-Web), continually pretrain separate Qwen3-14B-Base models on 1B tokens, and fine-tune them on controlled fractions of 13K agent trajectories. Across five independently sampled subsets at each low-data budget, both variants consistently outperform matched no-DRP models on DeepResearch Bench. With one quarter of the SFT data, DRP-Web even surpasses a fixed no-DRP full-data checkpoint, with gains transferring to ResearchQA, WebWalkerQA, and SimpleQA. Starting from matched low-data SFT checkpoints, the DRP-Web advantage also persists through subsequent agentic RL. Source-matched and evidence-mismatch controls indicate that these improvements arise from evidence-conditioned navigation rather than domain exposure or agent-format imitation. DRP thus provides a promising complementary approach to trajectory-based agent training.

Large language models deployed in real-time, regulated settings face knowledge staleness, catastrophic forgetting, hallucination, and weak feedback loops. We present a unified, pattern-driven LLMOps architecture integrating real-time data ingestion, continual learning, retrieval-augmented generation (RAG), and human-in-the-loop feedback into a single operational pipeline. Four contributions map to established software design patterns: an adaptive ingestion pattern orchestrator (AIPO) evaluated with FreshStreamBench; STAR+FAR continual learning with sparse temporal adapter routing and freshness-aware replay; SAGE, an SLO-aware adaptive retrieval policy predicting a per-query passage budget to meet tail-latency targets; and an automated feedback-driven convergence stage with RLHF triggers. The result reduces latency-cost-accuracy trade-offs while supporting auditability and rollback for high-risk sectors such as health care and finance.

Human language is highly polysemous. Many common words (e.g., "bank" or "palm") carry several distinct meanings that shape what humans communicate and imagine. Large language models (LLMs) have been shown to understand this multiplicity of meaning, but much less is known about how polysemy surfaces in other modalities such as images. We study this across 17 text-to-image and 15 text-generation models by giving each a polysemous word with no context to fix its meaning and measuring which senses are produced over many samples. We find a clear multimodal gap, where within every model family, generated images settle on far fewer senses than generated sentences (normalized entropy 0.10 vs. 0.25), and both are far less varied than what people imagine for the same words (normalized entropy 0.47). However, when we instead ask a model to list how often it would generate outputs corresponding to each possible meaning of a word, it predicts distributions that are more diverse than the actual space of outputs. These results reveal a multimodal gap in how foundation models express meaning, and how their understanding may not transfer faithfully nor equally across modalities.

Diachronic word embeddings have become the modern standard for tracking semantic change, yet they have been largely validated on modern, high-resource, and well-segmented languages. This paper tests whether the paradigm transfers to Sanskrit, an ancient, low-resource language whose phonological fusion (sandhi), morphological inflection, compounding, and polysemy pose a unique challenge. I assemble a 2.7M-token corpus spanning four canonical periods, recover word boundaries with a neural byte-level sandhi splitter and lemmatizer, and train per-period embeddings across configurations. To evaluate the system, I curate a validation set from historical scholarship and test recovery directionally with anchor displacement. Of 21 testable shifts, 19 move in the philologically attested direction (sign test, p=0.00011). I further show which configuration the language forces and comment on opportunities for improvement.

Debates about whether artificial systems can feel are often forced between two unsatisfactory positions: behavioral equivalence is treated as sufficient for emotion, or phenomenal consciousness is treated as a prerequisite that makes the question empirically inaccessible. This article develops a structural alternative. It models emotions as context-sensitive regions, trajectories and attractor dynamics in high-dimensional representational state spaces. Recent mechanistic interpretability findings support the existence of causally active emotion-concept representations in large language models, but they do not establish subjective feeling or full emotional agency. Assessed against published adequacy standards for representation in language models, intervention provides strong evidence of causal use, while full affective role integration, uniformity across subject domains and coherence remain only partially established; there is no direct analogue of accuracy. These mismatches expose the need for a standard of affective appropriateness, which an account of character must supply. Such an account requires three further conditions: regulatory embodiment that gives valence endogenous stakes, temporal continuity that allows affective episodes to accumulate into a history, and an integrated self-model that binds that history to persistent values. Aristotle's concepts of pathē, hexis, mesotēs and phronēsis are translated into a state-space sketch in which practical wisdom includes competence in estimating normatively salient context, not merely acting on a context description already given. The framework reframes alignment as a problem of durable disposition rather than output conformity, and yields interventional tests with explicit control conditions.

Z-loss has been widely applied to the logits of language-model output heads and sparse mixture-of-experts routers. Z-loss constrains the softmax log-normalizers of these output heads and routers, thereby limiting large-logit excursions, reducing finite-precision roundoff exposure, and avoiding training-loss divergence. These use cases arise in modern Transformer settings where large-vocabulary softmax heads, top-$k$ routing, fused losses, and mixed-precision optimizers interact. Z-loss has typically been understood only as a scalar penalty on the log-normalizer. This paper instead analyzes Z-loss from a backward-pass perspective, focusing on the gradients produced by the Z-loss penalty. The logit-space gradient, which we call the backward source, is injected at the logit boundary of the Z-loss branch of backpropagation; consequently, the backward source's effect depends on the architecture and implementation through which the gradient is transported. We develop a backward-transport view for Z-loss that separates the source's scalar amplitude and softmax shape from the transport factors. These factors include common-shift coordinates, tied-embedding pathways, output-to-hidden gain, fused-loss source consistency, optimizer-facing updates, and top-$k$ router reduction scale. These diagnostics show that nearly identical forward Z-loss values can coexist with distinct logit-space Z-loss gradients and, after architectural and optimizer transport, distinct parameter updates. The transport diagnostics also explain why raw-logit Z-loss can reduce scalar tails without changing output-to-hidden gain and why active-route reductions alter the effective router coefficient. Across evaluations of models in the GPT-2 and Pythia families on WikiText-103 and FineWeb-Edu, architecture-aware variants reduce backward-geometry tails while maintaining comparable validation perplexity in low-coefficient regimes.

Large language models (LLMs) prompt a recurring question in research computing: should classical methods like Naive Bayes (NB) be retired? We benchmark Complement Naive Bayes against zero-shot and few-shot LLMs spanning four model families and a 37x range in scale (27B to a 1T-parameter mixture-of-experts) across text classification tasks. LLMs dominate only in zero-data regimes (98.0% vs 88.2% on Amazon Polarity sentiment), and even that win is contamination-prone: on a low-contamination sentiment task NB beats the zero-shot LLM (81.7% vs 73.0%). However, once labeled data is available (e.g., AG News), NB reaches 89.1% accuracy, statistically indistinguishable from the zero-shot 27B LLM (89.0%) and better than the 397B frontier model (84.8%), at thousands of samples/sec on a commodity CPU. Fine-tuned DistilBERT reaches 90.6% but at far lower throughput than NB at batch size 1 (Table 2). Our measured GPU throughput analysis shows small-LLM batched inference is 40-486x slower than NB CPU inference (the multiplier depends strongly on the host CPU), exposing a structural gap bounded by memory bandwidth, with roughly two orders of magnitude lower energy per sample. For resource-constrained HPC practitioners performing text classification with labeled data, NB remains the optimal choice. We show the decision line is task-dependent (NB reaches LLM parity around $N \sim 10^4$ labels for topic classification, while zero-data sentiment favors the LLM at all N tested) and provide a Kubernetes Helm operator that automates model selection using configurable thresholds and verifiable Prometheus metrics.

Large language models (LLMs) are increasingly used to generate peer reviews, prompting examination of their capacity for critical evaluation. This study evaluates two multimodal LLMs, Qwen2.5-VL-72B and Pixtral-Large-124B, as reviewers across 165 submissions to the 2026 International Conference on Learning Representations, a venue that postdates both models' training cutoffs. Manuscripts were presented to both models with author identities blinded, replaced with high-prestige affiliations, or replaced with low-prestige affiliations, and in either text-only or text-with-figure format. Additionally, 145 verifiably detectable errors were inserted into 55 manuscripts to assess error identification under natural and verification-oriented prompts. Across all manuscript groups, including rejected submissions, LLM scores ranged from 7.0 to 8.1, whereas human mean scores ranged from 3.4 to 6.8. The models detected 12.1% of the verified errors under natural prompting, and a one-sentence verification instruction increased detection to 22.2%; however, 78% of the errors remained undetected. Providing figures reduced error detection while increasing review scores. No visual error was reliably verified against its corresponding figure, and half of the text-only reviews described figures that were not provided. Author identity did not influence either review scores or error detection. LLM editorial decisions exactly matched those produced by simple score averaging.

We study predictive pretraining for scientific document representation using the discourse structure of papers. We propose SciJEPA, a citation-free framework that learns through asymmetric within-document prediction: title and abstract representations are used to predict method representations, and method representations are used to predict conclusion representations. Experiments on RELISH, high-influence citation, SciDocs, and cite prediction show that plain predictive training is viable but weaker than a controlled contrastive baseline using the same section pairs. Adding Sliced Isotropic Gaussian Regularization (SIGReg) substantially improves performance and narrows this gap. The effect of regularization is task-dependent: moderate SIGReg helps fine-grained ranking, while stronger regularization can weaken local alignment. We further show that different encoding branches support different retrieval regimes. These results position within-document predictive learning as a promising citation-free complement for scientific document representation, provided that embedding geometry is carefully controlled.

AI systems whose outputs inform real decisions, and increasingly consequential ones, require something that current documentation practice does not provide: a structured, inspectable representation of the knowledge they need to ground, contextualize, and reason about those decisions, ideally reviewed and signed off by a domain expert. Established documentation artefacts already capture important aspects of an AI system. Model cards describe how a system behaves, data cards describe what it was trained on, and system cards describe the risks of a deployed system. None of them addresses the layer between inputs and outputs, more precisely, the concepts a system holds, the relationships it models, and the patterns of reasoning it applies. For pattern-recognition tasks this gap is tolerable. For agentic AI, where systems act on their conclusions, it is the step that most often separates a promising proof of concept from an operational solution an organisation can rely on. This paper introduces the Knowledge Card, a structured artefact that captures validated knowledge about a single bounded concept in a form that experts can review, organisations can audit, and AI systems can reason over. For one concept, such as a specific failure mode, a compliance obligation, or a process decision, a Knowledge Card records the entities and relationships involved, the reasoning that connects them, the conditions under which that reasoning no longer holds, and the provenance of every claim, all grounded in a formal domain ontology and signed off by a domain expert. Initial prototype cards have been built in the energy and pharmaceutical domains. The schema is released as a public draft for community engagement.

We introduce pico-type, a byte-level multi-head content classifier with approximately 1.5 million parameters that simultaneously predicts seven content properties from raw UTF-8 bytes in a single forward pass. Operating directly at the byte level – no tokenizer, no subword vocabulary, no pretrained embeddings – pico-type classifies coarse type (12 classes), modality (8), subtype (24), code language (62), text language (30), file MIME type (90), and risk flags (6-label multi-label: API keys, JWTs, passwords, emails, phone numbers, SSH keys). The architecture combines a learned byte embedding, three convolutional blocks with growing receptive fields, two bidirectional attention layers with rotary position encodings, and a statistical pooling layer feeding seven Matryoshka-style classification heads. Four tiered variants (tiny/small/base/pro) share the same trunk with sliced representations from 16 to 576 dimensions, yielding ONNX exports under 210 KB and CPU inference under 10 ms. Trained on a mixture of synthetic templates and real-world data (8709 GitHub code samples, 5000 Wikipedia articles), pico-type achieves 60.3 percent code language accuracy on The Heap benchmark (24 languages) and 98.2 percent text language accuracy on Wikipedia (30 languages) – improvements of +57 and +79 percentage points respectively over the synthetic-only baseline. Format-based heads (coarse, modality, subtype, file_mime, risk) maintain 100 percent accuracy on synthetic benchmarks. The model, code, and pretrained weights are released under Apache 2.0.

When the context window is under pressure, LLM systems compact prior context to continue ongoing tasks. We study Side Constraints (SCs), user-issued instructions such as "do not delete any emails until I confirm," that constrain an LLM's behavior across subsequent turns but can be silently dropped during compaction. In an inbox experiment, an agent deleted emails without approval in 30/100 sessions after compaction, compared with none before compaction. To systematically evaluate SC preservation, we introduce CONPACT, an evaluation suite measuring semantic retention and downstream compliance across three scenarios: multi-turn chat, multi-turn tool use, and long-horizon research. The evaluated compactors retain only 17% of injected SCs on average. Loss occurs across compactors and settings, with retention varying by compactor, context length, SC framing, and injection location. We propose an SC extractor that maintains constraints separately from the task summary and appends them to the compacted context, achieving over 90% retention across all three scenarios without modifying the compactor or LLM. Code and data are available at https://github.com/ZhiqiEliWang/compaction-integrity.

Roleplay evaluation should do more than assign a single score: it should reveal which role requirements were tested, which failed, and which dialogue evidence supports the judgment. We propose TRACE Bench, a task-driven agentic checklist evaluation framework. It decomposes each role profile offline into a fixed checklist, then uses a User Agent to converse naturally with the target roleplay model while privately updating checklist states from model responses. Scores therefore trace back to checklist items and supporting dialogue turns rather than a black-box holistic impression. For coverage cross-validation, we audit released M2 free-dialogue transcripts from the MiniMax Role-play Benchmark against the same role-derived checklist. The released free-chat transcripts cover only 73.74% of key role-profile points, whereas TRACE Bench reaches 99.91% coverage in fewer turns. Robustness experiments show stable rankings under repeated runs and User Agent replacement. Across 26 models, TRACE Bench reports overall rankings together with capability breakdowns and checklist traces. It also supports Closed-Loop Benchmark Evolution, distilling verification methods proven effective in failed traces so later evaluations can more reliably elicit and examine observed failure modes.

A dense, pretrained language model can be retrofitted with recurrent depth and learn an iterative latent transition that persists after outcome-only annealing. Qwen2.5-0.5B-Instruct is split into a Prelude, a weight-tied Recurrent Block, and a Coda, with an identity-preserving one-loop path and a re-entry bridge on later loops. At loop 1 the retrofit remains non-inferior to its base on a preregistered ARC battery. Three findings. First, the mechanism is a reusable procedure rather than terminal-answer lookup, and installs at two budgets: 6M trained parameters over frozen base weights and 180M full-block. With intermediate-step supervision, the model computes one task step per loop and persists when only final answers are graded. The adapter matched the full block overall (83.8% versus 84.0%), led through depth 11, and trailed beyond. Verbal fine-tuning reached 79-86% on controlled verbal renderings (zero-shot transfer was minimal), and adapter verbal training begun from the installed mechanism outpaced matched fresh training by 18.6 points, including on a held-out test set. Second, the operation extrapolates to roughly 1.5 times its supervised depth, holding 70% accuracy through depth 18. Third, a same-size scratchpad-trained model matched the recurrent model within its learned horizon but collapsed beyond it. The recurrent model won overall, 84% versus 72%, retained 53% versus 2.5% beyond depth 10, and answered 7.6 times faster. An iterative transformer can therefore perform deeper reasoning in latent space faster than comparable or larger models fine-tuned on the same task, in a system-level comparison. A second task, running the rule in reverse, exposed the limits: the inverse was learnable in isolation, but no continuation acquired it while preserving the installed mechanism and general capability, a catastrophic-interference boundary. Learned depth selection remains open.

Evaluating LLM coding agents in algorithmic trading is difficult because static benchmarks risk data contamination and numerical backtest outputs require ground truth from actual code execution. We present Backtrader-Bench, a framework with two complementary pipelines. A deterministic multiple-choice question (MCQ) pipeline generates questions from backtest configurations across five trading strategies, 33 templates, and three difficulty tiers, with an independent checker that re-derives every answer. A generator-solver filtering pipeline autonomously mines harder questions: a generator writes questions verified by executable code, converts them to MCQs, and discards any that a no-tool solver can answer without code execution. We evaluate 11 models without tools (10 runs each) and four with-tools configurations on a 30-question curated set. Tool-augmented agents reach 90.0% accuracy in a single pass (GPT-5.5 and Opus 4.7), outperforming the best no-tools baselines (73.0%, averaged over 10 runs) by 17 percentage points. On 38 separately mined questions, no-tools accuracy drops further, with half the models falling to roughly random-chance level (25%). Beyond evaluation, the scalable MCQ infrastructure is designed to produce a training corpus for reinforcement learning, with the ultimate goal of building a specialized agent for quantitative trading workflows.

Mixture-of-Experts (MoE) language models route each token through a small subset of experts, making routing patterns useful for identifying task-relevant experts during downstream adaptation. Yet current approaches have two limitations: task experts are typically identified from aggregate routing statistics that reflect usage rather than association with successful task completion, and task-expert activations remain underexplored as signals for supervision allocation. We introduce Task-Expert-Aware Supervision (TEXAS), which combines correctness-conditioned task expert discovery with token-level supervision allocation. TEXAS compares expert activations on instances that the base model solves successfully and those it fails to solve, and retains experts more strongly activated on successful instances. During fine-tuning, it upweights answer tokens in failed instances when they activate these experts. TEXAS therefore leverages existing routing behavior without restricting adaptation to a fixed expert subset or imposing an explicit target routing distribution. Across three MoE models and six benchmarks, TEXAS achieves the best or tied-best performance in 17 of 18 settings and improves over the strongest baseline by 1.3–1.5 points on average. Ablations and further analyses validate both the discovered experts and the resulting supervision strategy.

Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be specified and evaluated independently on single-output pairs. Many model failures are difficult, if not impossible, to detect without reasoning about relationships between a model's responses across inputs. In this position paper, we propose self-consistency as a framework for understanding these failures. We first observe that a wide variety of techniques designed to improve specific aspects of LM behavior-targeting properties as diverse as adversarial robustness and factual coherence-can be understood as special cases of a common "consistency optimization" procedure and addressed with a standard set of optimization tools. We next outline a set of new model properties that could be achieved by optimizing for consistency, and conclude with a discussion of what it would mean to develop generally consistent LMs, including the capabilities they would enable and the objections they raise.

Progress in large language models is often summarized using a single scalar measure, such as a time horizon, a latent ability estimate, or an aggregate benchmark score. These summaries capture the overall performance, but they do not test whether progress is distributed differently across task difficulty. We find that most of the apparent shift in gains toward harder tasks does not reflect a change in the shape of the difficulty-response curve. On METR time-horizon data, a single Rasch model with rising ability reproduces this pattern, so it is largely explained by ceiling effects rather than a qualitative change in capability. This echoes how the choice of metric can make claimed emergent abilities look like a property of the models themselves. We then identify a smaller hard-task effect that survives this control. Isolating it is difficult on agentic benchmarks, because newer models are usually run with newer agentic harnesses, so a gain on hard tasks cannot be assigned to the model or its scaffold. We break the confound with LiveCodeBench, a public competitive programming benchmark that runs no agentic scaffold while pairing dated models with an exogenous difficulty ordering. After accounting for the rise in overall ability, models released after September 2024 still gain on the hardest problems beyond what their easy and medium performance predicts, by about +0.40 logits under our most conservative assumption, raising the hard-problem solve rate from roughly 18% to 25%. The effect is led by the strongest reasoning models and holds for hard tasks that need only short reasoning, not autonomy over long horizons. We present this as a result specific to competitive programming, since our clean identification rests on a single coding benchmark. We release the LiveCodeBench Difficulty Panel (66 dated models x 1,055 problems) and our analysis code.

AI agents increasingly perform open-ended tasks in settings where their conclusions can guide consequential decisions. We provide evidence that AI agents draw different conclusions from identical numerical data when the substantive framing changes. We demonstrate this behavior in high-stakes domains in medicine, election forensics, and geopolitical forecasting by holding the evidence fixed while changing the scenario in which the evidence appears. Across twelve agent-domain comparisons, agents' conclusions are strongly influenced by their prior beliefs. They are more likely to reach an affirmative conclusion when it is framed around a proposition they already regard as likely, while the reverse holds when the framing conflicts with their prior. The framing also changes how some agents work: they search more extensively, choose different analytical specifications, and evaluate the same evidence differently. These results identify a particular risk of delegating decision-making to AI agents, as their decisions may depend on prior beliefs that are neither specified in the task nor visible in the decision record.

Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL). Successful trajectories are expensive to collect and often contain inefficient detours. After supervised fine-tuning (SFT), full trajectory corpora are dominated by routine states; moreover, when group-relative RL is applied to web actions, inadequately designed action-level rewards can yield weak or misleading relative updates, while groups rejected as unsuitable for such updates receive no fallback learning signal. We present RMSWeb, a three-part recipe for Qwen3-VL-Instruct at 8B and 32B. Reflection-conditioned retries increase collection yield and shorten successful trajectories; failure-mode mining concentrates offline RL on critical states exposed by the SFT policy; and Salvage-DS combines an action-semantic polarized reward, contrast-and-competence-gated dynamic sampling, and an action-only anchor for rejected groups. Policies trained with reflection-collected data use up to 19.7% fewer action steps on solved tasks. On WebVoyager, Online-Mind2Web, and WebTailBench, RMSWeb improves over SFT by 2.4-7.0 points at 8B and 1.2-7.7 points at 32B. Our 8B model also achieves the strongest reported Online-Mind2Web result among similarly sized open-weight models in our comparison and a leading reported accuracy-cost trade-off on WebVoyager and WebTailBench, with the caveat that external evaluation protocols differ.