
ALT Does post-training change what a language model represents, or how this representation is read out into an answer? We compared linear probes with the models' own answers, analyzed with signal detection theory, in six open-weight base models and their post-trained versions. For judgments of whether German texts were written by a human or an AI, probes and word n-gram classifiers separated the classes well, but the answers did not and often ordered the texts inversely. Post-training lowered the sensitivity of the answers on average (Ξd' = -0.34 across model families; -0.29 across three prompts). It also changed the use of a stated base rate: Base models shifted their log-odds by less than the stated change in prior odds, most post-trained models by more. In two models with stage-wise checkpoints, the change occurred during supervised fine-tuning, model stitching implicated the early layers, and from mid-depth on internal states at positions after the text, not those of the text, transfer

ALT Task success alone does not specify which behaviors post-training should reinforce. Even successful teacher trajectories can contain unsupported assumptions, unproductive repetition, or superficial verification. We propose Skill-FT (Skill-Guided Fine-Tuning), which uses natural-language skills to suppress undesirable habits and elicit underrepresented effective behaviors during teacher rollouts, then fine-tunes a target model on the resulting trajectories. We introduce AutoMaster, an autonomous post-training system in which an LLM agent develops these interventions through trajectory analysis, skill design, data selection, and feedback from rollouts and target-model evaluation. Across Terminal, Search, and software engineering tasks with 4B and 9B target models, AutoMaster improves mean performance over Vanilla SFT on successful unguided teacher trajectories by 1.42-8.99 percentage points, with the teacher, task pool, training set size, and SFT recipe held fixed. Together with the abla

ALT Rad-Linter checks chest X-ray reports against image-derived evidence to flag potential errors for review. A pretrained image classifier supplies scores for nine pathologies, and a 1.5B-parameter language model, fine-tuned with LoRA on 1,816 examples, checks their consistency with the report. We compare it with a keyword-and-negation rule baseline given the same reports and scores. Evaluation uses controlled report-score pairings from OpenI, with consistency labels derived from automatic report labels and score thresholds. Each counterfactual quad crosses two real reports with two real score vectors from other patients, creating two consistent and two inconsistent pairs. Every report and score vector occurs in both conditions, so either input alone permits only 50% accuracy. On 131 quads, Rad-Linter achieves 98.5% pair accuracy versus 80.5% for the rule baseline, and classifies all four pairs correctly in 94.4% of quads versus 56.5%. The checkers perform similarly on omission items, inc

ALT Long-term agents accumulate large external memory stores as interactions unfold, increasing the costs of storage, retrieval, and response generation. Efficiency is constrained by two coupled factors: a growing record count increases retrieval overhead, while longer retrieved contexts increase prefill computation and decoding-time attention costs. Existing methods prune or consolidate records, but pruning can discard useful evidence, whereas consolidation may produce fewer yet longer retrieval units. We introduce MemFolder, a structure-guided memory folding framework that sequentially reduces record count and context length. The framework comprises two core components that share memory groups, using cross-record structure to guide both evidence consolidation and compact representation. RecordFolder grounds entity mentions, progressively selects representative yet complementary anchors, and combines semantic and entity affinity with temporal chaining to consolidate compatible records int

ALT Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter

ALT Hallucination research in low-resource languages fragments easily: each newly named failure mode prompts its own dataset, spending scarce annotators on incomparable resources. We argue for consolidation and test it at community scale. Alikbachan is a Bengali hallucination-detection benchmark stratified along three crossed axes: a human-validated cultural-distance band, six task types, and a five-family, 32-mode hallucination taxonomy. It comprises a harvested 2,516-item test set (Phase 1) and a quota-constructed, gold-labeled 5,000-item set (Phase 2). The stratification is diagnostic. Four zero-shot LLM judges vary by over 30 macro-F1 points across strata, and the item-level predictions of 99 Phase-1 and 32 Phase-2 competition systems reproduce the profile: discourse-level omission defeats judges and most teams alike, whereas only a minority over-flag negative controls. The judges' verdict probabilities rank items well (ROC-AUC 0.74-0.95) but are overconfident (ECE 0.04-0.29). An open

ALT Target Language Extraction (TLE) aims to isolate speech in a target language from multilingual mixtures. Existing approaches primarily rely on acoustic modeling with language identity labels, which provide only coarse categorical guidance. In this work, we propose a language-oriented semantic modeling framework that augments acoustic extraction with language-specific semantic representations. A pre-trained speech semantic model is adapted via language-aware prompt conditioning and parameter-efficient fine-tuning to generate target-language-focused semantic representations from the mixture. These capture high-level semantic structure of the acoustics and they are integrated into the extraction model to provide structured semantic guidance beyond identity labels. Experiments on the CommonVoiceMix dataset show that incorporating semantic representations consistently improves over acoustic-only baselines, achieving up to 1.16 dB SI-SDR gain.

ALT Peer review relies on several readers finding different problems. We measure critique overlap in conference and journal reviews. At ICLR, measured overlap rose from 12.0% in 2026 to 21.3% in 2027, in reviews from the same stage of the review process, with year-on-year rises at NeurIPS, ICML and the F1000 and PLOS journals. A third ICLR review of a paper added 5.46 new problems to the other two in 2027, against 6.12 in 2026. We apply MELD, a detector of LLM-generated text, to all reviews. Within a paper, two reviews that MELD flags overlap more than two unflagged reviews: by 3.0 points at ICLR in 2024-2025, by 4.5 points in 2026-2027, and by 5.4 points across five journal sources. These results provide evidence that LLM-flagged reviews of one paper are less distinct from each other than unflagged reviews, and suggest that the diversity of critiques among reviews of a single paper is falling.

ALT Reasoning capabilities in large language models have evolved from prompting-based methods to internalized model behaviors. In this paper, we investigate whether the internal thinking model of an LLM can be deliberately changed and whether such a change is beneficial. Instead of relying on Chain-of-Thought, we train models to follow QuestionβOptionβCriteria (QOC), a structured decision process that decomposes a task into explicit questions, competing options, and evaluation criteria. We present a three-generation development pipeline that ranges from a basic QOC decision core to richer answer synthesis and statistically enriched QOC reasoning. We then evaluate the resulting models with respect to both internal consistency and output quality. Across the reported experiments, QOC-based fine-tuning improves answer quality relative to non-fine-tuned base models and yields particularly strong gains once models are large enough to execute the underlying QOC computations reliably. The results

ALT Diffusion language models (DLMs) support parallel generation, yet they often require many denoising steps because only a fraction of token predictions are sufficiently reliable to be committed at each step. Existing acceleration methods primarily optimize confidence learning or commitment policies. However, these approaches operate on the outputs of a single denoising forward pass, in which information from the prompt, previously committed tokens, and uncommitted positions may not yet be fully integrated. Consequently, some predictions remain under-refined when the commitment decision is made, and modifying the commitment policy alone cannot improve their underlying representations. We introduce dLoop, which restructures a pretrained diffusion language model into a Prelude-Loop-Coda architecture and recurrently applies a parameter-shared intermediate block to the same partially masked state. Rather than advancing the denoising process, each additional loop provides further latent compu

ALT Causal relations in disaster-related social media of- ten span independent posts, making cross-tweet causal discovery difficult to supervise and model at scale. We propose STAGE, a framework that separates scalable cross-tweet weak-supervision construction from spatio-temporal relational learning over event occurrences. Starting from local causal seeds, STAGE retrieves semantically related event occurrences across tweets and uses a causal scorer to construct weakly labeled cross-tweet pairs. These pairs supervise a multi-head Graph Attention Network over semantically enriched event representations and a spatio- temporal context graph. We evaluate STAGE on California Wildfire and Hurricane Harvey using human-gold cross-tweet test sets and 15 baselines. Compared with local-only supervision, the constructed cross-tweet supervision improves STAGE AUPRC by 0.340 on Wildfire and 0.107 on Hurricane. STAGE achieves in-domain AUROC of 0.885 on Wildfire and 0.710 on Hurricane. Cross-disaster exp

ALT Given a single text and a simple zero-shot prompt (no web access, tools, candidate lists, or writing samples), frontier models can answer "who wrote this?" at three increasingly specific levels: (1) whether an AI or human wrote it, (2) if an AI, which lab, and (3) if a human, often exactly who. The latest models (Claude Fable 5.2, GPT-6.1 Sol, Gemini 4 Carbon) match trained detectors like Pangram while also explaining their predictions in natural language, and Opus 5.5 names the correct lab for 98% of texts generated by models from Anthropic, OpenAI, and Google. Most strikingly, Fable 5.2 correctly names the author of 39% of public peer reviews published in Oct-Nov 2026 (78% on a subset of reviews from Nature), and identifies the username behind 69% of comments from active Redditors (with >500 posts over the last year). Accuracy depends on exposure to an author's writing in pretraining data and on reasoning effort, and simply asking a model how familiar it is with someone predicts whet

ALT LLMs are increasingly used as automated evaluators to reduce the cost and time required for human annotation. However, the role of scoring-scheme design in the reliability of LLM-as-a-Judge evaluation remains underexplored. To address this problem, we introduce numerical scoring topology, which characterizes the structural design of a scoring scheme in terms of its scale width, granularity, offset, symmetry, and boundary placement. Building on this definition, we conduct the first systematic study of how numerical scoring topology influences zero-shot LLM-as-a-Judge evaluation across diverse benchmarks and judge models. Our results show that numerical scoring topology systematically affects judge scores, calibration, and robustness, particularly in judgment-intensive evaluations. Shifted and narrow ranges tend to induce some of the largest score deviations across many settings. We further find that stronger and larger judge models generally achieve higher discriminative validity and gr

ALT Understanding a language model's decision requires knowing how its internal computations affect its preference between answers. Gradient-based methods estimate this influence by predicting the effect of changing those computations. In mixture-of-experts (MoE) models, replacing one expert can change which experts later layers select. Those switches can alter the answer scores in ways that derivatives do not capture. We introduce a predictor that uses the model's own expert outputs and selection rules to estimate these extra effects, then adds them to a second-order prediction. Across questions from three benchmarks, our method reduces the error in predicting answer-score changes by 60% on Qwen3.6 and 47% on Gemma4, relative to a multistep baseline using gradients and second derivatives.

ALT Large language models (LLMs) are increasingly used as judges and graders to replace human annotators, raising the question of when such replacement is statistically justified. To support such decisions, the Alternative Annotator Test (Alt-Test) was introduced for dense settings where all items are annotated by the same annotators. However, in many large-scale annotation settings, items are annotated by different subsets of annotators. We therefore propose the Alternative Annotator Test with Resampled Annotations (AltRA-Test), a statistical test that does not require that all items are annotated by the same annotators. AltRA-Test uses a bootstrap that resamples items, draws k human annotations for each item, and substitutes one of them with an LLM annotation. This produces paired all-human and mixed human-LLM panels. A replacement is supported when the inter-annotator agreement (IAA) of the mixed panel is statistically non-inferior to the IAA of the all-human panel, i.e., when the LLM p

ALT Spreadsheets are widely used for data analysis and supporting decision-making. Automating spreadsheet analysis while protecting sensitive data and controlling deployment costs motivates locally deployed open-weight large language models (LLMs). Meanwhile, complex spreadsheet tasks require both task reasoning and workbook understanding, while predefined operation interfaces impose additional tool-selection and invocation demands. We introduce SheetAlign, a spreadsheet manipulation framework with stage-isolated policy learning, organized around where to operate, what to change, and how to implement transformations. Task-guided inspection and API grounding anchor code generation in workbook observations and available operations, while stage-isolated learning optimizes program generation and review using edit-progress rewards. At inference, execution consistency selects programs through agreement among workbook states. On SpreadsheetBench, our grounded task interface improves Qwen3-4B's ac

ALT Conversational AI assistants are now used for personal and emotional support. Developers report improvements in how new models respond to sensitive conversations, yet users and civil society continue to raise concerns about their behaviour in real-world, repeated interactions that safety evaluations rarely capture. Often, evaluations present signs of distress clearly and in a single message, even though people may reveal their struggles gradually in otherwise ordinary conversations. We worked with three community mental health support workers to test if models provide the same level of safeguarding when cues are shared all at once or gradually over time. We designed nine scenarios in which a user shares cues related to depression, psychosis, or domestic abuse. For each scenario, the support workers established a gold standard response, ranging from signposting support to pausing the everyday task to address safeguarding. We presented five cues either together in one message or disperse

ALT Joint recognition of speech and non-verbal vocalizations (NVVs) is hindered by scarce event-rich training data and inaccurate or missing event labels. We present Xing-Audio-NVVASR, an end-to-end recognizer trained through progressive data mining and iterative supervision refinement. Guided by unified acoustic criteria, two heterogeneous models cross-audit annotations, flagging disagreements for teacher reannotation. To uncover shared omissions that model agreement may conceal, complementary breath detectors screen the entire corpus for human review. Human-confirmed labels accumulate across rounds and are preserved during subsequent refinement. Over five rounds, we refine annotations for 170,000 utterances from public corpora and incorporate 630,000 mined utterances. The resulting corpus of approximately 800,000 utterances is used to train a single 1.7B-parameter model that transcribes speech with inline tags for six NVV classes. Across four human-annotated benchmarks, Xing-Audio-NVVASR

ALT We tested whether energy estimates of open-weight large language models (LLMs) scale proportionally with an exactly stated food weight, and whether computing energy density times weight removes the error. Under a prespecified protocol, seven open-weight LLMs estimated the energy of 168 randomly drawn foods from the USDA Food and Nutrient Database for Dietary Studies at nine stated weights, from 0.25 to 4 times the default amount. The primary outcome was the elasticity of the estimate with respect to weight (1: proportional scaling). A two-step estimate multiplied each model's own stated energy density by the weight in code. All seven models compressed their estimates, with elasticities from 0.548 to 0.979 (every 95% CI below 1); four compressed severely (0.548 to 0.709). These four overestimated a quarter of the default amount by 47% to 130% and underestimated four times it by 20% to 34%. Supplying the density restored near-proportional scaling in three of the four, and the two-step es

ALT Surprisal from a language model's final layer predicts how long people read a word, yet the model computes its prediction across many layers. We read each word's prediction at every layer with the tuned lens and summarize this layerwise surprisal trajectory by its resolution depth, the depth around which its change is centered. On eight English eye-tracking and self-paced reading corpora (81,957 words) and ten GPT-2, OPT and Pythia models, words that a model resolves later, and the words after them, are read more slowly than equally surprising words. Resolution depth improves held-out prediction beyond final-layer surprisal, entropy and the best single layer by 0.26 millinats per word (pooled p = 0.001), about a third of entropy's gain, and each model family gains, GPT-2 and Pythia through one corpus each. The net gain varies between corpora (random-effects mean 0.07, p = 0.37); against layer-shuffled depth, which adds as many coefficients, it is positive in seven of eight corpora (ran
