Papers

Real-world time series are frequently driven by exogenous events and structural shifts, rendering conventional forecasting based solely on historical numerical observations insufficient. While language models can retrieve external news, standard retrieval-augmented approaches struggle with high noise, missing signals, and an inability to reason causally about event impacts. We propose SEER (Self-Evolving Event Reasoning and Retrieval), a closed-loop framework that dynamically optimizes event conditioning for time series forecasting. SEER translates prediction errors into two decoupled feedback mechanisms: (i) a reflective retrieval memory that refines subsequent search queries and filters spurious noise, and (ii) a persistent causal knowledge base that distills transferable domain dynamics. SEER enforces strict chronological boundaries across both event retrieval and reflection, preventing look-ahead bias and data leakage. Across six volatile time-series benchmarks, SEER consistently outperforms state-of-the-art time series foundation models and language model baselines.

Language models exhibit strong reasoning capabilities, yet adapting them to structured domains remains challenging and can yield inconsistent outcomes. We identify representation compatibility, the extent to which a model effectively processes a representation for a structured task, as a key factor in adaptation. We study this in chess, which provides a controlled testbed with precise semantics, computable optimal actions, and multiple state representations, including a symbolic encoding (FEN) and a spatial format (ASCII). We find that models often process semantically equivalent inputs substantially differently, affecting both learning and generalization. Building on this observation, we propose representation-aligned auxiliary supervision, which uses environment-derived tasks expressed in compatible representations to improve adaptation to structured domains. Across models and representations, auxiliary supervision consistently improves optimal-move prediction relative to target-only training under identical target data. Tasks that expose environment dynamics provide larger and most consistent gains than surface-level or static supervision, while remaining competitive with substantially increasing the amount of target-task data. Moreover, ASCII-trained models transfer more effectively to FEN than FEN-trained models do to ASCII, even surpassing the FEN target-only baseline on FEN evaluation. The gains also extend beyond optimal-move prediction to open-ended, factually grounded commentary generation. Overall, our results show that auxiliary supervision in model-compatible representations can enable effective adaptation in structured domains.

Despite rapid progress in automating scientific research, generating promising and well grounded research solutions remains a central challenge. We isolate research ideation as a standalone task and build our solution on the intuition that a challenge in one field can often be addressed by a mechanism that solved an analogous challenge in another. Accordingly, we introduce IdeaScientist, which decomposes ideation into gap finding, innovation, and report writing, and trains each role with reinforcement learning. These roles identify limitations in related work, draw solution intuitions from analogous problem settings, and develop those intuitions into complete research proposals. To facilitate discovery of insights across domains, we construct the Svalbard Idea Vault, a corpus of 2.77M decomposed research ideas for retrieval, training, and temporally controlled evaluation. Our evaluation restricts access to literature available before a cutoff date and assesses how closely proposed directions align with those later explored in 15K papers authored by human researchers. On Qwen3.6-27B, IdeaScientist outperforms the strongest open-source autoresearch baseline by 14.0%, driven mainly by gains in novelty. On this 27B open backbone, IdeaScientist even outperforms Claude Code SDK with Claude-4.8-Opus and Codex SDK with GPT-5.4, by up to 5.9%.

A language model reads long text in one quadratic forward pass, stops at the context window, and loses accuracy with length before reaching it. We ask whether the read can be factorized when deciding over a finite set: which document is relevant, which option is supported, which passage is the evidence. Periscope, a training-free inference method, arranges the $N$ chunks of a text on a $K{\times}K$ grid with $K{=}\lceil\sqrt{N}\rceil$ and asks a frozen model the same question about $K$ local spans of consecutive chunks and $K$ strided spans that sample the whole text, reading the log-odds of every answer at one token. Each answer takes its best local and strided score, and scoring every chunk by its two spans gives an evidence map at no further cost, whose peak is the chunk behind the answer. Every probe is about $\sqrt{sc}$ tokens for a text of $s$ tokens and chunk size $c$, so a window of $W$ tokens reaches $W^{2}/c$ tokens at $s^{1.5}$ cost. The map replaces the long read. On LongBench v2, reading only the $K$ chunks the map ranks highest, 9k tokens, matches the same model's best window read across windows from 32k to 1M tokens, and on InfiniteBench, where the median context is 150k tokens, it leads the best window read by 5 points. The same map ranks BRIGHT's long-document corpora with the best NDCG@10 of six methods. Each call caches only one probe, so a 27B model reads 4.5M-token contexts on one 80GB GPU, where a single pass would need 296GB of cache. A long read then needs a GPU that holds the model, not one that holds the text.

Localizing behavior to individual components of a language model is a central goal of mechanistic interpretability. However, scoring components one at a time misses context-dependent effects: a primary component can inhibit the activation of a backup, leading to issues with ranking components. Actual causality studies the structure of such interactions via witnesses: variables that provide contextual information to resolve interaction terms. However, estimation with witnesses typically requires combinatorial enumeration and is infeasible in practice. We introduce the witness-integrated set effect (WISE), a family of causal estimands that build on the witness mechanism while taking expectations over sets of causes and witnesses to remain computationally feasible. Building on this approach, we introduce JuntaLearner, a gradient-based circuit discovery method that learns to rank components by their causal impact across varying-sized sets of components and witnesses. Alongside faithfulness metrics, we introduce measures of necessity and task specificity, and the circuit recognition score (CRS) to summarize each metric across circuit sizes while emphasizing effects achieved by small circuits. Across tasks and models of increasing size, JuntaLearner achieves higher mean CRS compared to attribution baselines on all metrics. Since its cost does not grow with the number of candidate components, JuntaLearner scales to large models while accounting for set-level interactions and avoiding first-order approximations.

A trainable input embedding table assigns each vocabulary item an independently adjustable vector. We investigate whether this token-specific parameterization is required for substantial language-modeling capability, or whether a shared Transformer can learn from fixed token identities. We compare three decoder-only language models trained from scratch with the same tokenizer, contextual backbone, untied output-head architecture, and training recipe, with a target budget of 100 billion prediction tokens per model. Their input interfaces are a learned table, canonical 16-bit token-ID codes, and one fixed invertible recoding over GF(2). The fixed codes are repeated to model width without an additional trainable input projection. Both fixed-code models acquire substantial capabilities: canonical codes achieve 52.40% HellaSwag normalized accuracy, 70.51% PIQA accuracy, and 42.75% LAMBADA accuracy. The learned-input control performs better on several evaluations, including HellaSwag and LAMBADA, so these results establish viability rather than performance parity. The fixed interfaces remove 100.7 million trainable parameters, yielding 1.711B-parameter models, but parameter reduction is not the central result. These single-run experiments distinguish architectural necessity from empirical utility: independently trainable token-specific input vectors are not required for the observed capabilities. A fixed identity interface also provides a controlled setting for studying representation learning downstream of an immutable input, without establishing where particular capabilities are localized.

Large language models (LLMs) are increasingly used as synthetic personas representing survey respondents. Their validity as substitutes for particular respondents depends on whether they reproduce individuals' decisions. We examine what information helps synthetic respondents predict each individual's later choices, using five conditions that add progressively richer information: no personal information, demographics, personality traits, cognitive scores, and finally the respondent's earlier survey choices as behavioral history. We use a two-wave panel of 845 US adults who completed measures of 14 behavioral biases (spanning risk, time preferences, overconfidence, and reasoning), so each respondent's earlier answers provide a human test-retest benchmark; in the behavioral-history condition, all items that score the target bias are withheld. At the population level, the average number of biases per respondent in every condition is close to the human average (7.1-8.1 biases, against 7.1 for humans). This aggregate similarity masks differences in variance: persona descriptions recover only 53-67% of human between-person variation, whereas adding behavioral history restores it to approximately the human level. At the individual level, description-based personas achieve only 7-12% of the informedness observed in human test-retest responses, while adding behavioral history raises this to 28%. The condition including behavioral history has the highest estimated informedness in all 17 demographic groups, whereas description-based conditions provide little or no information for some groups. Synthetic responses also exhibit stronger education- and income-related differences than human responses. For LLM synthetic personas, a respondent's past answers add more to individual-level prediction than a description of who they are.

The BAIBAICHUCHU team participated in the Social Media Subtask of NTCIR-19 FinArg-3, ranking Chinese investor posts by Maximum Possible Profit (MPP). A three-track ensemble of lexical features, a FinArg-2-pre-finetuned MacBERT ranker, and an LLM judge reaches 0.734 in post-grouped development evaluation, but our best official run scores 0.517. All twelve submitted runs lie between 0.4598 and 0.5402, and our 26 unanimous three-track pairs score 0.500. A post-hoc audit finds that the submitted judge applied a long-only rule to bearish posts although MPP is stance-aware. Correcting it changes 28 of 87 official predictions without changing accuracy, yet lowers development accuracy from 0.680 to 0.622: a regime-specific semantic shortcut improved validation fit. We decompose ranking into directional text, volatility and horizon, and pairwise margin. A running-extremum model predicts $\sigma\sqrt{T}$ scaling, observed ex post in 502 price-aligned posts from a separate July 2026 collection. Pre-posting volatility scaled by the post-specific observation horizon, $\sigma_{\mathrm{pre},20}\sqrt{N_i}$, is associated with a later truncated-horizon MPP proxy (Spearman $\rho=0.320$; ticker-cluster 95% CI [0.133,0.466]) and correctly orders 61.4% of unequal-outcome pairs. Transferred text is nearly uncorrelated with the July outcome ($\rho=0.055$) and adds little conditional on this score and stance. Development reliability rises from about 0.60 to 0.93 as the labeled MPP gap widens. An earlier ERAI result of 0.6207 rules out universal unpredictability as a simple explanation. Evaluation should jointly consider text, market regime, historical volatility, horizon, and pair composition.

During fine-tuning, a language model can assign less probability to previously learned answers even when the current gradient acts to preserve that probability. With momentum, each update also carries gradients computed at earlier model states, and these stored contributions can push the model in the opposite direction. We investigate how this optimiser memory contributes to forgetting by separating old-task loss into confusion among its answers and leakage of probability outside the answer set. Across three language-model families, answer mass consistently declines while discrimination among old answers usually improves: the model becomes less likely to produce answers that it can still distinguish correctly. Decomposing Adam updates reveals opposing contributions to this loss of answer mass. Over training, accumulated history favours leakage, while the current gradient opposes it. Resolving history by age shows that the harmful contributions come mainly from older gradients of the new task, whereas recent gradients tend to protect the old answers. Changes in history's effect are dominated by its orientation relative to the old-task gradient. Interventions that reset momentum while matching the initial update norm establish that stored history affects retention, with state-dependent immediate effects and lower final old-task loss over longer Adam continuations, mainly through recovered answer mass. Finally, integration along finite updates shows that most sampled large loss increases are captured by local projections, while curvature along history amplifies some events. Together, these findings reveal how an optimiser's memory can erode learned behaviour even as its current gradient acts to preserve it.

General decision models, such as Jev, have recently emerged as efficient alternatives to LLMs for structured judgment and selection. But what kinds of decisions can these models reliably make, and how does their behavior change when individual decisions are composed into larger systems? To study this, we introduce JEVal, a bilingual benchmark comprising 11,257 instances from 36 datasets across 10 application domains, and evaluate 25 model configurations spanning general decision models and generative LLMs. Our results show that (1) general decision models are most competitive when decisions can be resolved from available evidence, but weaken when they require specialist knowledge or faithful uncertainty estimation: they can often identify the most likely outcome while substantially overstating its probability. (2) In more dynamic and realistic systems involving long-horizon, multi-step interactions, the advantages of fast local decision making are offset by reliability failures at the system level. on $\tau$-bench, faster local decisions reduce median episode time but lower task success as decision errors accumulate over long trajectories. (3) In large-scale social simulation, decision models approach strong generative LLMs on individual response prediction at substantially lower inference cost, yet remain weaker in user profiling and exhibit larger aggregate estimation errors and systematic bias. Finally, we propose SII-Decision-4B and SII-Decision-27B, which internalize an open-weight LLM's own reasoning into a single-pass first-token decision through Reasoning-to-Readout Self-Distillation, with SII-Decision-27B performing on par with Jev on JEVal while answering a typical query in about 0.1 s.

When documents supporting an agent's derived facts are revoked or replaced, should it repair memory or re-read current evidence? We introduce an evidence-revision evaluation on medication- and problem-list tasks from public ICU records. Under revocation, replacement and control events, we compare full and source-filtered re-reading with caching, rebuilding and graph-local repair across two 7B models. Memory is supplied in full without retrieval, and costs include ingest, revision and every use. On short records, local repair uses 5-10$\times$ fewer revision tokens than rebuilding, yet every memory pipeline costs at least twice full re-reading in held-out conditions. In a small pre-specified development sweep, adding task-ineligible documents extended records to about 10,000 tokens; at that length, memory's mean cumulative cost fell below full re-reading's after 2-14 uses, partly through truncated extraction, while source-filtered re-reading remained cheapest. In the replacement study, none of the four primary confirmatory tests reached statistical significance. These results show why the cost of agent memory after evidence revision must be assessed against source-filtered re-reading over the full pipeline.

A deployed LLM agent emits tool calls, queries, and code that can be silently wrong – by the time the error surfaces, the action has run. Frontier chat APIs hide the model's token probabilities; the agent's stated confidence barely beats chance on the mistakes that matter; and resampling does not help, since frontier models are highly repetitive, reproducing the same call across samples. We recover the missing signal from a low-cost open-weight surrogate run in parallel. It reads the same context, schema, and proposed action as the agent, then scores the call from its own log-probabilities through a family of complementary readouts: teacher forcing and request-PMI weigh the likelihood of each argument value, a discriminative verdict judges the call as a whole, and tool-choice competition tests the function against its siblings. One principle says which to trust: a generative likelihood localizes wrong argument values, while the verdict catches holistically wrong calls. When the error type is unknown, an ensemble is the low-regret default. The readout is training-free, needs no access to the agent's internals, and costs one prefill pass alongside the tool call. On difficult coding tasks it reaches AUROC 0.825 where the actor's stated confidence is near chance (0.598), and the generative readouts beat it by +0.07 to +0.28 across three further actors. Against self-consistency it gains +0.14 to +0.19 on near-deterministic actors, at 1/K the cost. The signal drives two deployment modes: a real-time gate escalating the least-trustworthy calls for review (+0.05 to +0.30 accepted-action accuracy at 50% coverage), and confidence feedback, returning the tool result with the score so the agent adapts its next step – lifting task success on live-execution benchmarks (+0.119 and +0.137, p <= 1e-4) and beating a random-value control where step errors are silent (+0.078, p = 0.003).

AI agents are becoming increasingly capable of generating scientific code, but generating code is not the same as improving the algorithms behind it. For numerical solvers, execution feedback can expose poor performance, but rarely reveals its underlying cause and how to address it. We introduce Auto-Diagnosis and Skill Discovery (ADSD), a framework that links numerical diagnosis to reusable solver self-improvement. ADSD follows a diagnosis-first paradigm that first explains why a solver performs poorly, then uses this diagnosis to guide the discovery of appropriate numerical methods. The resulting knowledge is packaged into reusable solver skills, turning solver improvement from trial-and-error editing into a structured process of diagnosis, discovery, and implementation. Across four challenging numerical domains–power flow equation, AC optimal power flow control, stiff ordinary differential equations, and heterogeneous diffusion PDEs–ADSD consistently improves solver accuracy, robustness, and efficiency. On GOC-500 power flow, for example, ADSD reduces mean solver error by nearly $71\times$, with improvements further transferring to unseen grid topologies and operating regimes.

Long chain-of-thought (CoT) traces impose substantial output-token costs. Under constrained budgets, compression must preserve answer-critical information, making boundary placement central. Token-level and fixed-length boundaries can fragment coherent spans such as phrases, formulas, and local derivations, whereas step-level boundaries can bind content requiring different compression actions. We introduce SynLat, a text-latent CoT framework that aligns compression boundaries with syntactic structure through non-overlapping Syntax-Aligned Units (SAUs). An answer-conditioned Teacher constructs progressive KEEP/LATENT targets for a single compression-conditioned Student, which generates mixed reasoning from only the question and requested compression level at inference. Across two Qwen3 Student scales, Standard-CoT and Long-CoT groups, and three compression levels, SynLat matches or exceeds the strongest evaluated baseline in all 12 task-group aggregates and strictly leads in 11 under the reported achieved-CR selection protocol. Overall gains reach 3.6/2.6 points at MEDIUM and 7.0/5.5 points at HIGH for Qwen3-8B/14B, with larger advantages under stronger compression, particularly on Long-CoT groups.

Real-world outpatient oncology notes contain specialised terminology, tumour staging expressions, treatment names, toxicity descriptions, and institution-specific de-identification markers that may not be represented efficiently by general biomedical or adjacent clinical language models. We developed and evaluated oncology-specific BERT-style encoders using a governed UK outpatient oncology corpus comprising 290,026 notes from 21,564 patients treated for lung and head-and-neck cancer. We compared RadBERT and PathologyBERT with two local strategies: OncoNote-RadBERT, produced by continued masked language model pretraining, and OncoNoteBERT, trained from scratch with an oncology WordPiece tokenizer. Models were evaluated using masked language modelling loss and perplexity on the validation set, tokenizer fragmentation metrics, clinical term tokenisation, masked-token probes, and exploratory representation analysis. Both external encoders fit the oncology corpus poorly in zero-shot evaluation (perplexity 113.04 for RadBERT; 2035.03 for PathologyBERT), while continued pretraining produced the strongest fit (2.10 for OncoNote-RadBERT). OncoNoteBERT achieved perplexity 2.83 but produced the most efficient tokenisation, with lower subword fertility and shorter normalised sequence length. It also returned a clinically acceptable prediction for 12 of 13 masked-token probes, compared with 7 of 13 for OncoNote-RadBERT. This divergence between corpus-level fit and masked-token performance was partly attributable to tokenizer fragmentation rather than learned semantics alone. Both locally developed models represented the institutional placeholder as a single learnable token. These findings show that continued adaptation and bespoke tokenisation provide complementary benefits, and that representation-layer design matters before adjacent-domain encoders are applied to oncology NLP.

Similarity scores are often offered as evidence that a model shares structure with the brain or across languages. We ask what such a score reads when that structure is removed, or when the instrument measuring it does not work, in two settings. Across seventeen languages, a grammaticality probe transfers worse between more distant languages (r = -0.66). But the probe's own accuracy falls along the same axis and is at chance in four languages, four of the five most distant. Dropping them halves the explained variance, and because it also narrows the range of distances, the design cannot say how much of the gradient is the instrument. Counting the 272 language pairs as independent gives p = 0.0006 for a steering effect that is null when the seventeen languages are the unit (p = 0.155). In brain alignment, a training objective raises a language model's similarity to fMRI responses from 0.10 to 0.34 (reliability ceiling 0.54), yet targets with the brain correspondence destroyed still reach 0.31. A shuffled target scores about k/n against the real one, for rank k and n sentences, a baseline that can be computed before any model is trained. Both interventions work, yet we detect no distance-graded steering effect and no syntactic benefit from alignment. Before reading a correspondence score, measure how much of it survives without the correspondence.

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations. We introduce Queen, a 4B-parameter chess-language model that can explain its moves and plans while playing at the level of a typical Grandmaster. Our novel framework enables domain-specific reasoning through complementary components: an encoder-decoder architecture and an iterative distillation algorithm. This architecture integrates a silent expert chess encoder with an instruction-tuned LM through cross-attention, which we train via a question-answering curriculum to extract chess concepts from the encoder's representations. Building on this domain-adapted model, we iteratively improve its explanations with a natural-language analog of the Bellman update: the model analyzes the positions after its top candidate moves and consolidates them into an explanation of the current position, which is then distilled back into the model. Over seven iterations, our model gains over 900 Elo points (1782 to 2697), substantially surpassing all frontier models on both playing strength and puzzle accuracy, despite containing three orders of magnitude fewer parameters. Furthermore, LM-based evaluations show that our explanations are fluent and approach GPT-5.6-Sol (high) in coherence. The generality of our architecture and training procedure suggests a recipe for applying language models to domains where silent expert encoders are available, like games, robotics, and computer use.

LLM-guided evolutionary methods, such as AlphaEvolve, have emerged as powerful approaches for challenging computational optimization problems, such as circle packing. However, prior work typically optimizes performance gain over a fixed number of iterations. We argue that practical optimization should maximize gain per unit cost. To this end, we propose FrugalEvo, a cost-aware evolutionary framework where a stronger, higher-cost LLM explores solution strategies, and a cheaper LLM implements them and iteratively refines the resulting code. We also design a cache-efficient evolution process, where our harness and prompts maximize the sharing of prefixes across different evolution steps, to improve cache reuse. To measure solution quality throughout a fixed cost budget, we introduce Budget-Aware Area Under the Curve (BA-AUC), defined as the area under the best-so-far evaluation score curve over cumulative LLM cost, up to the budget. Across 10 mathematical and systems optimization tasks, FrugalEvo matches or surpasses state-of-the-art baselines, including OpenEvolve, ShinkaEvolve, AdaEvolve, and EvoX, in final solution quality and achieves higher BA-AUC on 9 tasks. It also achieves higher average performance than these baselines on 10 algorithmic optimization tasks from ALE-Bench-Lite. Notably, on circle packing, FrugalEvo achieves new state-of-the-art performance with GPT-5.6 Terra and Luna for only 1.68 USD and with GLM-5.3 and its Flash variant for only 0.55 USD, matching or surpassing all baselines, including multi-agent methods such as CORAL and SwarmResearch, which cost approximately 50 USD on average.

Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and shape much of the response. Most post-training recipes for dLMs do not use this signal to decide which tokens to train on: they typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that shape the response. We introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments (pivots). Pivot-SD selects pivots using an information-gain metric measuring uncertainty reduction over the remaining masked positions. Pivots from successful trajectories are trained with cross-entropy, and pivots from failed trajectories with targeted unlikelihood, leaving the rest of the failed trajectory untouched. Using only 200 questions and four rollouts each, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks.

Physical fidelity has received increasing attention in world models and video generation, yet how video representations encode physical information remains less understood. We introduce the World Embedding Benchmark, comprising 8,000 controlled simulation cases from 80 families spanning fluid mechanics, solid mechanics, dynamics, and optics & electromagnetism. Each case pairs a rendered video with simulation-derived physical annotations, supporting three complementary tasks: text-video retrieval, physical-property regression, and multiple-choice video-description pair classification. We use these tasks to distinguish cross-modal physical alignment from the recoverability of quantitative physical information. Evaluated pre-trained omnimodal embedding models show weak retrieval and near-chance within-family pair classification, while lightweight probes recover useful physical information from frozen video embeddings. Continual contrastive training with physics-specific video-text pairs improves retrieval and pair classification but degrades physical-property regression, revealing a trade-off between alignment and quantitative information recoverability. Finally, we use the embeddings to retrieve reference videos for retrieval-augmented generation with MiniMax-H3. Retrieved references improve the physical fidelity of generated videos, with stronger retrieval models yielding larger gains in our experiments. Together, these findings highlight the need to evaluate physical alignment and property recoverability jointly, and demonstrate the utility of physical representations for improving video generation.

Relational databases are among the most widely deployed forms of structured knowledge, and natural language access to them requires grounding language onto schema entities and relations while handling the ambiguity inherent in how people phrase requests. Existing synthetic NL-to-SQL data generation methods largely ignore this ambiguity and produce oversimplified queries that fail to prepare models for the complexity of real-world structured knowledge access. We present FALCON, a framework that generates realistic, ambiguity-aware NL-to-SQL data matching the complexity of challenging real-world benchmarks, at low cost using compact open models. Our approach combines reserved-word SQL seeding and persona-based prompting to generate structurally complex queries, while alignment-based filtering preserves difficulty by distinguishing genuinely incorrect examples from complex but valid queries. Human evaluation confirms consistent high quality across model sizes, and our generated data exceeds existing benchmarks in both SQL complexity and natural language richness. Difficulty-stratified analysis shows models trained on FALCON data increasingly outperform baseline-trained models as query complexity increases, validating our pipeline's success in generating challenging training data. When combined with a small proportion of existing benchmark data, mixed training recovers performance on simpler queries while preserving these advantages on complex ones. The model- and database-agnostic design enables organizations to generate high-complexity NL-to-SQL training data locally without external APIs.

We describe the Writerslogic team's participation in the CLEF 2026 SimpleText shared task, addressing Task 1 (text simplification) and Task 2 (complexity spotting). For Task 1, we develop a multi-candidate generation pipeline using GPT-4o-mini that produces five simplification candidates per sentence at varying temperatures, then selects the best candidate using a reference-free scoring heuristic that rewards compression, source word retention, Cochrane Plain Language Summary vocabulary usage, and lexical simplicity. On Task 1.1 (sentence-level simplification), our Claude Sonnet 4 submission achieves SARI 47.43 and BLEU 14.21, the top-ranked sentence-level system (3rd on the combined Task 1 leaderboard, behind two document-level submissions). For Task 2, we fine-tune a DeBERTa-v3-large NLI model on 350K labeled (source, sentence) pairs, framing hallucination detection as natural language inference. The model reads the most relevant source sentence as premise and the candidate as hypothesis, directly learning to distinguish grounded from hallucinated content. On Task 2.1 (binary overgeneration identification), our fine-tuned DeBERTa system achieves 0.8081 document-level macro F1 (0.8085 in our best ensemble), the top-ranked entry within the identification track and 2nd among teams overall, behind AIIR Lab (0.8197). On Task 2.2 (multi-class error classification), our best submission reaches 0.804 multiclass accuracy, ranking 2nd among unique teams behind AIIR Lab (0.827). We evaluate both tasks on English and multilingual biomedical text from Cochrane systematic reviews.

We describe the Writerslogic systems for three PAN at CLEF 2026 shared tasks (Reasoning Trajectory Detection, Voight-Kampff Generative AI Detection, and Multi-Author Writing Style Analysis), unified by a shared analytical framework: feature robustness under distribution shift is governed by support overlap between training and test distributions, not by training-set effect size. This yields a taxonomy (domain-anchored, domain-portable, domain-invariant) that explains why generator-specific features die under domain shift while vocabulary fingerprints (hapax ratio, Yule's K, Heaps' exponent), compression measures, and character n-grams survive. On Reasoning Trajectory Detection, where training was entirely mathematics and 84 percent of test was unseen domains, the framework guided system design to 1st place in source detection (0.85 macro F1 via Opus-Sonnet agreement) and 3rd place in safety classification (0.66 macro F1 via query-refusal decomposition). For Voight-Kampff, we built a calibrated ensemble of DeBERTa-v2 (ONNX), multi-seed LightGBM with 44 domain-portable stylometric features, and SVM on n-gram TF-IDF, combined via learned stacking with isotonic calibration; the best configuration achieved 0.891 on the PAN 2026 test set with balanced sub-metrics (0.853 to 0.902 across all evaluation dimensions). For Multi-Author Writing Style Analysis, we describe a system fusing spectral clustering over character n-gram similarity graphs, normalized compression distance for local boundary detection, and SmolLM-135M perplexity for neural change-point detection; a platform mix-up meant our run never reached the official evaluation, so we report the design and its a priori predictions. Across all three tasks, features measuring generation process properties are designed to outperform features measuring generated content properties under domain shift.

Authorship Attribution (AA) requires capturing fine-grained stylistic characteristics, making it particularly challenging in zero-shot (ZS) settings where no task-specific supervision is available. In this work, we investigate the effect of author representations on ZS AA by evaluating a label-only prompting baseline together with three author representation strategies: representative writing samples, LLM-generated descriptions, and style embeddings (LISA). The first three approaches perform attribution using LLM prompting, while the embedding-based approach uses style embeddings with cosine similarity. We investigate the influence of prompt design and propose a two-stage embedding-based attribution framework that combines candidate space reduction with embedding-dimension selection. The results show that label-only ZS AA is ineffective, while incorporating author-specific representations consistently improves attribution performance. Among the evaluated approaches, the proposed two-stage LISA framework achieves the strongest overall performance, whereas LLM-generated style descriptions provide a substantially more compact representation of author style at the cost of some attribution performance. These findings demonstrate the importance of author representation in ZS AA, while indicating that current open-source LLMs remain insufficient for robust attribution without more effective representation learning.

Distillation has become a core primitive of large language model training, but its properties are not yet well understood. We take an entropic perspective, studying how the entropy of the student depends on the data and the divergence that define the distillation objective. We prove that forward KL inflates the entropy of the student above that of the teacher. Since cross-entropy training is a special case, this yields an identity that we verify quantitatively in pretraining and supervised finetuning. Other divergences come with no such guarantee: reverse KL deflates entropy until the gap between student and teacher gets too large, and interpolating between the two changes entropy smoothly early in training but abruptly at convergence. The lower entropy of on-policy distillation comes from token-level reverse KL, not from on-policy sampling. The divergence therefore acts as an implicit entropy regularizer, whose role is clearest in self-distillation: as conditioning on privileged information deflates entropy, the divergence hyperparameters that work best are those that compensate for it.

Table-to-report generation refers to the task of automatically generating article-level analyt- ical reports from relational tables and is an essential capability for automated data science and decision support. Its central challenge lies in systematically discovering verifiable com- posite insights across tables, attributes, and analytical perspectives, and organizing them into coherent, complete, and traceable evidence chains. Existing methods primarily rely on sequential, reactive data agents or direct Large Language Model(LLM) generation. They suffer from exploration bias: early local observations constrain subsequent actions, causing models to focus prematurely on local analyzes and miss cross-table or cross-dimensional evidence. We propose ComInsight, which reformulates insight discovery as the composition of atomic evidences. We first define an atomic insight as the smallest executable analytical unit conforming to a predefined analysis pattern and enumerate all valid atomic insights from database schema and content. These atoms are then organized into a multi-relational insight graph, where nodes represent verified data facts and edges encode logical, temporal, or hierarchical relations. Finally, a set of composition operators systematically fuses atomic nodes into higher-order composite conclusions. Every composite output is accompanied by executable SQL and fine-grained provenance, ensuring full verifiability. Across three benchmarks InsightBench, DDR-Bench, and T2R-Bench, ComInsight consistently outperforms strong baselines in factual correctness, novelty, and structural completeness. We believe ComInsight offers a reliable, efficient, and explainable path toward table-to-report generation.

On-policy self-distillation (OPSD) uses reference solutions as privileged hindsight to supervise student-generated reasoning trajectories. However, reference-based guidance may explain a correct solution without addressing why the student's own reasoning fails. This reasoning mismatch between the guidance provided and the correction needed can encourage the student to borrow correct conclusions while leaving its reasoning errors unresolved. Moreover, applying the same hindsight throughout the trajectory risks a distillation trap, where unnecessary constraints on valid reasoning compete with correction of substantive errors. To address these issues, we propose Root-Cause-Guided On-Policy Distillation (RC-OPD), which uses repairs of the student's own reasoning to provide guidance that addresses its specific errors while building on valid progress. For each failed attempt, RC-OPD locates the earliest substantive error, develops a local correction, and uses the corrected intermediate result as an anchor for the valid prefix. An iterative diagnosis–repair–continuation process tests the repairs through student continuation, identifying further errors within a fixed repair budget. For repair chains that reach a correct answer, root–cause–guided distillation uses failure diagnoses and corrective goals to supervise the erroneous segments, while anchor-guided distillation supports the corresponding valid prefixes with reasoning chains leading to the repaired intermediate results. We evaluate RC-OPD across multiple datasets and model scales. Extensive experiments and analyses show that it mitigates reasoning mismatch and the distillation trap, yielding substantial performance gains.

Hallucination detection is particularly important for medical language models, but repeated-sampling approaches are computationally expensive. A faster alternative is a single-pass uncertainty head that predicts hallucination risk from a frozen generator's internal signals. Existing uncertainty heads consume backbone-specific features and tokenization, motivating adaptation when the backbone or language changes. We study two Persian medical models, Gaokerena-V and Gaokerena-R, on a 168-question Iranian medical entrance examination. Across five generations per question, Gaokerena-V produces the same option on only 14 questions and Gaokerena-R on 37, compared with 168 for Med-Gemma, indicating substantial response variability in the Gaokerena models. We therefore adapt the LLM Uncertainty Head (LUH) framework to these models and construct two paired claim-level hallucination datasets directly in Persian, with 1,600 responses per backbone. Lightweight heads are trained on frozen-backbone attention maps and token probabilities. On held-out test splits, the heads achieve PR-AUCs of 0.4820 and 0.4652, corresponding to 2.30 and 2.66 times their respective random baselines, and ROC-AUCs of 0.7852 and 0.7810. The resulting detectors require neither retrieval nor repeated sampling at inference.

Post-training with verifiable rewards can induce reward hacking, motivating the use of monitors within the training objective rather than solely for offline auditing. We show that a low monitor readout does not identify whether such an intervention controls behavior. In a code-generation environment whose dominant exploit is available at the start of the reasoning trace, we train policies against three monitors that pass the same offline gate: an in-domain activation probe and two penalties conditioned on how early the policy commits to its own final answer. The probe score is at its numerical floor from the first recorded training step, and the trained-score median is zero for every prefix-trained run at the endpoint. These readouts estimate different quantities, and we do not compare their scales; within each monitor family, however, low values do not establish behavioral control. Within one fixed configuration, prefix-trained runs with the same zero-median trained score range, by seed alone, from a mixed regime with a low hacking share to near-pure reward hacking. All probe runs reach the hacking regime, but their floor-level readout reflects a mismatch between the position where the probe was validated and the position where it was read during training, not a second instance of this ambiguity. Text-level analysis identifies a prefix failure mode: generic planning and filler shells postpone the exploit past the cut without eliminating it from the final output. Low measured commitment therefore does not distinguish a low hacking share from delayed commitment to the exploit. Offline discrimination and low monitor-aligned readouts are insufficient evidence of behavioral control; an out-of-band behavioral check is required. We characterize the endpoint readout, not its evolution. Code is available at https://github.com/zhezhou1106/spoof-cost.