Embodied instruction following in interactive environments requires agents to interpret natural language commands and complete long-horizon tasks. Existing planners either rely on large labeled trajectories, or use few-shot and zero-shot LLM prompting that produces a whole plan in a single serial reasoning process. These approaches depend on linguistic common sense, commit to a plan from the initial instruction, and only update after failures, which weakens robustness and state awareness. We present AtomTree, a hierarchical framework for state-aware task planning with LLMs. AtomTree incrementally builds a tree of subgoals and actions, where the LLM decides at each step, from the current symbolic state, whether to further decompose or to output a grounded low-level action. A lightweight memory module summarizes past outcomes for reuse in later decisions. When a low-level action stalls, AtomTree triggers hierarchical backtracking replanning to return to the parent node and regenerate the subtree plan from the latest state, preventing continued drift along an incorrect trajectory. In zero-shot evaluation on the ALFRED benchmark, AtomTree outperforms prior LLM-based planners. Project page: https://zhaihaotian.github.io/atomtree-website/.
Mainak Ghosh, Michael E. Rose, Sebastian Erhardt, Erik Buunk, Dietmar Harhoff
PaECTER is an open-source document-level encoder specific for patents. We fine-tune BERT for Patents with examiner-added citation information to generate numerical representations for patent documents. PaECTER performs better in similarity tasks than current state-of-the-art models used in the patent domain. More specifically, our model outperforms the patent specific pre-trained language model (BERT for Patents) and general-purpose text embedding models (e.g., E5, GTE, and BGE) on our patent citation prediction test dataset on different rank evaluation metrics. PaECTER predicts at least one most similar patent at a rank of 1.32 on average when compared against 25 irrelevant patents. Numerical representations generated by PaECTER from patent text can be used for downstream tasks such as classification, tracing knowledge flows, or semantic similarity search. Semantic similarity search is especially relevant in the context of prior art search for both inventors and patent examiners.
Frontier language models express coherent preferences, but behavioral evidence alone cannot say whether these belong to the model or to a character it portrays. Following Hofstadter's view of the self as a self-description, we ask whether a model's description of itself works as a portable specification of its preference profile, a quine test: can the text a model writes about itself regenerate its behavior elsewhere? Ten models from nine labs write self-descriptions; we place each one in the prompt of other models (receivers) and of fresh instances of the author, and measure how closely their choices on a 110-item preference battery match the author's. First, the tested native self-descriptions (native: the profile measured under a generic assistant prompt) show no average reconstruction gain across variations in length, framing, wording and prompt placement: on a 0 to 1 fidelity scale where a model's fidelity to a fresh instance of itself is 0.979 and to unrelated models about 0.89, the one-sided 95% upper bound on the pooled gain is below +0.01 in every tested system-prompt condition, and descriptions based on the author's observed choices do no better. Second, the channel itself is strong: explicitly listing the author's answers lets receivers reproduce its profile at fidelity 0.955 to 0.960 across three model grids, within 0.03 of same-model repeatability. Third, self-descriptions written under an explicit preference policy transfer the assigned preferences almost as well as the policy text itself. Fourth, reading its own self-description shifts even the author away from its baseline, modestly more than unrelated prose of the same length. For the preferences measured here, native self-descriptions influence behavior without reliably specifying how to reproduce it in another model. We release the pipeline, battery, raw call logs, and 198 model-written descriptions.
Chia-Hsuan Chang, Brian Ondov, Jiang Lv, Lan Jiang, Yunmai Wang, Haerin Song, Xueqing Peng, Qiaozhu Mei 等 10 位作者
AI co-scientists have expanded the use of large language models for generating scientific hypotheses and guiding experimental investigation. Here we investigate whether temporal patterns in the semantic representations of complete biomedical conclusions can guide hypothesis generation. We introduce ForeSCITE, a framework that identifies thematic clusters of published conclusions, estimates a direction separating earlier from more recent findings within each cluster, and projects existing conclusion embeddings along these directions. An embedding language model trained to reconstruct conclusions decodes the projected coordinates into natural-language hypotheses. Across 407 clusters derived from approximately three million pre-cutoff conclusions, we identify temporal structure and illustrate its retention after decoding. In retrospective, time-sliced evaluations, ForeSCITE increases semantic coverage beyond the historical corpus, biomedical entity-pair novelty and post-cutoff pair re-observation relative to five prompt-based and retrieval-augmented baselines. Expert review of prioritized CAR-T hypotheses identified 37 candidates as both novel and biologically plausible, and expert justifications for five selected hypotheses illustrate their biological rationale and subsequent literature support where available. These findings support temporal guidance in semantic space as a scalable mechanism for proposing biomedical hypotheses for subsequent scientific evaluation.
Hankun Wang, Bohan Li, Shi Lian, Xiaoyu Gu, Jing Peng, Da Zheng, Yiwei Guo, Colin Zhang 等 10 位作者
Speech editing for content creation requires precise control over both what an edit should do and where it should apply. Free-form natural language provides a flexible interface for expressing edit requests, but its ambiguity may leave the intended operation, parameters, or target region underspecified. We study a semi-formal transcript-grounded edit instruction that specifies operation categories and parameters and binds them to transcript spans or boundaries. This semantic timeline avoids explicit timestamp alignment and provides an externally inspectable contract for compositional edits. We instantiate the interface in dots$.$tts$.$edit, a speech editing model built on a continuous autoregressive speech generation architecture. Four representative speech-creation controls cover lexical content, affective expression, pitch and speaking-rate delivery, and temporal phrasing through text, emotion, prosody, and pause editing. Task-specific data pipelines construct operation- and scope-controlled pairs while retaining source-derived context outside each target region. We further introduce doteBench, a bilingual benchmark that measures precise instruction following, local preservation, and audio quality across the four controls and their composition. Experiments show leading overall instruction following and local preservation across its five editing categories, while audio quality remains comparable to existing open-source systems. Across three Seed-TTS-Eval shards, the model shows negligible differences from the base model in zero-shot TTS recognition error rate and speaker similarity.
Verifying numerical claims in research written by AI agents is an important but challenging problem. Recent benchmarks show that many agent-written papers fail integrity checks, but such verdicts, often produced by a single LLM judge, identify failure without explaining its nature, and existing checks either test a paper for internal consistency or test whether a reported value appears somewhere in an agent's output. We re-analyze all 540 released judgments from a recent public benchmark of AI research agents (nine agents, fifteen open research problems) and manually label 256 numerical claims from 18 submissions. Failures concentrate on alignment between the paper and the underlying work rather than overt cheating, and about four in five unsupported claims correspond to values in the agent's own files that come from the wrong experiment, metric, or run. We introduce a deterministic claim-to-evidence checker that scores whether the surrounding context identifies the claimed result and calibrates each claim against numerical decoys. With its rules fixed before labeling, it certifies claims with over 90% precision against blinded annotators (78 of 85; Cohen's kappa = 0.89), while verifying only a subset of claims; most claims it cannot verify are nevertheless traceable, so failure to verify is not evidence of falsehood. An LLM verifier given the same evidence recovers more traceable claims but accepts about three times as many misattributed ones, so the two are best used together. Applied to 99 submissions, the checker reveals large differences in traceability across agents, and a second judge from a different vendor preserves the official agent ranking while passing twice as many papers. These findings motivate provenance-first auditing of AI-produced research.
Ranking aggregation from multiple retrievers is a powerful yet under-explored technique in modern retrieval and RAG systems. We introduce LLM-IRA, the first extension of the quantum-inspired interactive ranking aggregation framework to document reranking. By leveraging large language models (LLMs) solely for pairwise judgments on high-disagreement document pairs, our method learns fusion weights in a principled, interactive way. The final ranking is a simple, ultra-fast pointwise linear fusion of scores, no full pairwise matrix or expensive listwise prompting is required. Experiments on the BEIR benchmark and TREC Deep Learning tracks with an open-weight instruction-tuned judge (Llama-3.1-8B-Instruct) show that LLM-IRA matches or numerically exceeds strong static fusion baselines on 5 of 6 evaluated benchmarks. We report two weight-update variants: momentum blending and a full accumulated-history refit, which together reach statistical significance (Wilcoxon signed-rank, p<0.05) over the best unsupervised aggregator on 4 of 6 benchmarks (DL19 under momentum; DL20, DBpedia, and NFCorpus under accumulated refit, which is the stronger of the two configurations overall). On the benchmarks we profiled (DL19, DL20), it already clears the best unsupervised baseline within as few as 5 LLM calls per query, well before exhausting its typical ~25--30 call budget. LLM-IRA also achieves cost-effectiveness comparable to state-of-the-art zero-shot LLM rerankers at an order of magnitude fewer LLM calls, and requires no training data or model fine-tuning.
Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable. However, common approaches rely on privileged information to construct a stronger self-teacher, limiting their use when only unlabeled prompts are available. In this paper, we introduce Weakening-Induced On-Policy Self-Distillation (WISD), which instead constructs a weaker counterpart of the original model and uses the contrast between the two as a training signal without requiring privileged information. We explore two ways to construct the weaker model: low-bit quantization and magnitude pruning. We evaluate WISD across diverse architectures and scales on mathematical and coding reasoning tasks. Across these benchmarks, quantization and pruning improve average accuracy over the original models by 4.0 and 5.5 percentage points, respectively, achieving state-of-the-art performance even without privileged information. Notably, pruning raises Qwen3.5-4B's accuracy from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6. Quantization raises Qwen3.6-35B-A3B's accuracy from 68.5% to 74.6% and from 61.8% to 69.5% on the same benchmarks, respectively. These results establish WISD as a novel and effective self-distillation framework for large reasoning models.
Lu Wang, Zhuoran Jin, Longxiang Wang, Junhong Chen, Yupu Hao, Yao Xu, Daojian Zeng, Yubo Chen 等 11 位作者
Recent advances in multimodal large language models have increased interest in interaction paradigms for video understanding. The evolution from offline to streaming and proactive interaction has gradually shifted models from passive responders toward more proactive agents that can decide when to answer. However, existing proactive paradigms typically assume that the information required to answer a query is complete, leaving models to only wait or answer. In the real world, however, information is often incomplete due to uncertainty from either the video or the user. In proactive paradigms, such uncertainty can cause the model to wait forever or generate highly hallucinatory responses. We therefore argue that proactive interaction should be redefined with a third action: ask, allowing models to actively ask for missing information. To study when and what models should ask before answering, we introduce AskStream, the first multi-turn benchmark spanning video and user uncertainty for streaming video understanding, with four task categories and eight subtasks. It contains 438 videos with 2,714 manually annotated frame-level time windows across three types: wait, ask, and answer. Across both multiple-choice and open-ended settings, even the strongest model attains only 52.18% on our proposed Ask-F1 metric, highlighting substantial limitations in proactive information seeking under uncertainty.
Financial assistants answer questions that sound factual but depend on what the user leaves unsaid: what a data figure measures, which product variant is meant, and the investor's situation and the rules that apply. An assistant that ignores these can still answer with a real number from its data, one that answers a different question. We call this frame collapse. We study it in Indian mutual funds, where product variants and tax rules are formally defined, so the correct answer can be computed for every reading. One study tests each challenge. Given an ambiguously labelled figure, models used it, often under the user's label. Given every alternative labelled, they listed them all, but the layout decided which came first. Missing a fact about the investor, weaker agents assumed a default or an outdated tax rule; requiring the fact did not guarantee its use. Because such answers match their data, checks against the data pass them. An AI judge caught mislabelled figures only when told what each field means, and still passed answers that silently chose another reading. We show how these failures arise, not how often, and propose controls for financial assistants.
Jev enables efficient decision-making with large language models (LLMs) through typed querying primitives, allowing the same semantic state to be expressed in different formats, such as binary and multiple-choice queries. However, the consistency of decisions across primitives has not been systematically evaluated. This work introduces PrimaCons, a four-axis benchmark for assessing cross-primitive consistency while preserving the semantics of the query. Evaluations across three Jev sizes reveal consistency violations across all four axes. Our analysis shows that these violations follow systematic patterns and that cross-primitive discrepancies predict decision failure with a ROC-AUC of 0.91. To mitigate these inconsistencies, we propose the Cross-Primitive Residual Calibrator (CPRC), which attaches a lightweight residual network to a frozen backbone and jointly optimizes predictive accuracy and cross-primitive consistency. Under a leave-one-domain-out protocol, CPRC reduces cross-primitive inconsistency by 58.57% and improves accuracy by +2.97%., demonstrating its effectiveness in both in-domain and out-of-domain settings. Our findings establish cross-primitive consistency as a key dimension for evaluating Jev and provide a practical approach to improving the reliability of its decisions. Code and benchmark is available at https://github.com/67L1/PrimaCons-CRPR.
When an LLM agent fails, repair methods choose which action to replace and what to put in its place. They often score candidate repairs with greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. We ask whether these evaluators pick repairs that help the agent that will actually execute them. Holding failed trajectories and candidate single-step repairs fixed, we label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, pre-specified tests show that greedy replay loses 0.08-0.13 in expected net success gain in ALFWorld and 0.06-0.13 in net reward in WebShop, relative to selection with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B with Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own, and fresh rollouts confirm that the two models prefer opposite repairs on selected disagreement cases. Between similar-sized models from different families, the losses appear in both directions in both environments (0.07-0.17). The scoring executor also changes repair-location labels and method rankings. Repairs should be verified with the executor that will run them, and repair labels should name their executor.
A binary success indicator is a natural representation for discriminative tasks such as image classification under adversarial perturbations, where each input is mapped to a single prediction. However, generative tasks are fundamentally different: an LLM jailbreak prompt induces a distribution over open-ended trajectories, while a sampled response reveals only one path. Yet LLM jailbreak evaluation typically retains the binary response indicator underlying attack success rate, obscuring the surrounding structure of possible generations. We introduce a continuous, context-conditioned risk landscape over first-token branches, estimated by fixing candidate openings and repeatedly sampling their continuations. This continuous risk-landscape perspective reveals two complementary properties: current decoder exposure and safe rerouting capacity, which have been obscured by binary-indicator-based ASR. The resulting representation also provides a common basis for comparing clean prompts, attacked variants, and attack strength against a known adaptive defense. Across multiple target models and jailbreak attacks, we uncover risky branches beneath apparently safe outputs, identify qualitatively different ways attacks reshape the model's opening distribution, and show that attacks strongest before a defense do not necessarily remain the strongest after the adaptive defense reacts. These findings recast jailbreaks as deformations of the surrounding continuation-risk landscape, establishing a more informative foundation for understanding jailbreak behavior and developing safer language models.
This paper reports a methodological pivot from the Open Khipu Repository to the Khipu Field Guide (KFG) database, prompted by stalled progress on the OKR, and a full re-analysis of the resulting 650-khipu sample. Three findings recur across every analysis attempted. First, khipu size confounds nearly every derived structural count, explaining 42-70% of its variance; all subsequent models control for it. Second, the two-letter name prefix (AS, UR, HP, QU, JC, MM, KH) identifies the original cataloger, not a site or period, which turns documentation completeness into a second confound that must be modeled explicitly. Controlling for both, khipus cataloged by Hugo Pereyra's Proyecto Quipu (HP) show significantly fewer Ascher sum-relation types than Gary Urton's (UR) khipus (p = 0.044), while the Ascher-cataloged (AS) khipus are too collinear with their own near-total lack of side/direction documentation to assess (VIF = 5.0). Third, generic structural clustering (UMAP+HDBSCAN) does not separate the pooled 650-khipu sample from two independent null models (permutation and Gaussian), corroborating an independent conclusion already reached by the KFG itself; the same test restricted to a single cataloger (UR) proved not reproducible across runtime environments and is reported here as a methodological caveat rather than evidence of structure. A targeted replication of the KFG's own S/Z-knot spatial-block heuristic for calendar/astronomical khipus recovers 8 of 10 published candidates and proposes 34 additional ones under the same criterion.
Large audio language models (LALMs) can process speech directly, but it remains unclear how well they use the non-lexical acoustic information that speech carries. We study this question through pain, using TAME-Pain, which contains pain-neutral Harvard sentences and spoken pain ratings recorded at different self-reported pain levels. We introduce HEAR, a four-stage diagnostic framework that traces pain-related information from the speech signal to the model's prediction, and apply it to seven closed-source and open-weight LALMs. On the Harvard sentences, where pain can only be inferred from the voice, all models perform at chance level, and same-speaker references or few-shot examples improve balanced accuracy by no more than six points. On the spoken ratings, the same models reach 87.5-99.7% balanced accuracy. Layer-wise probing of open-weight LALMs shows that pain can be decoded from early audio representations (74.1% balanced accuracy for binary prediction and 45.0% for three-class severity), but that this information, along with most pain-associated acoustic cues, decreases in deeper layers. In an acoustic perception test, weaker models cannot distinguish even large changes in these cues, whereas stronger models can but still fail to infer pain. Finally, we recover pain-related acoustic cues and directly reinject them as structured text into the prompt, which improves pairwise pain prediction by up to 32 points. These results indicate that current LALMs encode paralinguistic information in early layers that is largely unavailable when the model produces its answer.
Retrieval-augmented generation (RAG) systems ground large language models (LLMs) by retrieving external evidence and conditioning generation on the retrieved context. Although dense retrieval is effective for local semantic matching, independently ranked retrieval can return redundant passages, miss cross-chunk dependencies, and provide incomplete evidence for multi-hop reasoning. We propose SpectRAG, a spectral graph-augmented retrieval framework that combines hybrid candidate seeding, weighted chunk graph construction, graph-based expansion, spectral relevance propagation, and token-budgeted context selection. Document chunks are represented by dense semantic embeddings and graph-spectral coordinates derived from a weighted chunk graph whose edges encode semantic similarity, document adjacency, section hierarchy, shared entities, citation links, and similarity between LLM-generated utility questions. At query time, SpectRAG builds a high-recall seed set using exact inner-product search and hybrid relevance signals, expands it over the chunk graph, re-scores candidates using graph proximity and spectral information, and assembles the final context under a fixed token budget. The goal is not merely to retrieve individually similar chunks, but to construct a context that is relevant, diverse, structurally connected, and sufficient for grounded reasoning. Experiments on MultiHopRAG, UltraDomain, and long-document question-answering settings show that SpectRAG improves broad evidence recovery and achieves competitive generation quality against dense, hybrid, HyDE, LightRAG, and NaiveRAG baselines. The results further indicate that graph-spectral structure is particularly useful for inference-oriented and cross-document retrieval, while highlighting trade-offs in early precision, computational cost, and domain-dependent performance.
Personalizing AI assistants requires adapting their behavior to users' preferences and task contexts. Existing work elicits assistant personas through persona prompting or finetuning on persona-conditioned generations. However, persona descriptions alone do not identify what to change or preserve in a particular response, or which behaviors suit the current request. We introduce $\textit{Persona Learning from Feedback (PLF)}$, a framework that learns reusable, preference-grounded assistant personas as independent LoRA adapters on a shared frozen backbone. During training, a frozen self-teacher conditions on persona specifications and response-specific feedback to provide token-level supervision on each adapter's own generated prefixes. A feedback gate focuses corrective updates on responses judged to need improvement. At deployment, a supervised router uses the request or dialogue to select a single adapter or a fixed composition, without user-specific finetuning. We evaluate PLF on single-turn personalization with model-generated critiques and multi-turn tutoring with simulated students' reactions. Across both settings, PLF improves model-judged preference fit over fixed-persona policies and strong baselines including persona prompting and persona distillation, demonstrating how feedback-trained behaviors can be reused for adaptive assistant personalization without degrading general capabilities.
In high-risk text interactions, a system may need to estimate whether the current speaker is likely to be a minor before applying age-conditioned safeguards. Many text-based dialogue systems lack strong identity signals such as account information, verified identity, facial data, or device data, and must rely on weak cues that emerge across multiple turns. We formulate implicit minor detection under weak identity information as a high-risk user-attribute inference task and propose a constrained framework comprising a large language model, a Harness, and a Maker-Checker protocol. The framework controls when inference is invoked, requires structured dialogue-linked evidence, and subjects validation-selected Harness revisions to version-level review. On a controlled Chinese benchmark comprising 8,328 multi-turn dialogues, our method improves average precision and minor-class F1 by 7.60 and 6.37 percentage points, respectively, over direct inference using the same LLM family.
Giuliano Martinelli, Abelardo Carlos Martinez Lorenzo, Satvik Garg, Linxi Wang, Riccardo Orlando, Piriyakorn Piriyatamwong, Samuel Paul Fraiberger
AI systems that ground their answers from retrieved documents usually need only a few relevant passages, but applications such as systematic reviews and legal discovery need nearly all of them, because their conclusions rest on the full body of evidence. These applications have traditionally searched with curated taxonomies, which can only find what their vocabulary describes. We compare taxonomy-based retrieval with an alternative neural retrieval based on sentence embeddings and followed by language model screening in ImpactAI, an evidence synthesis application built on 7,354 randomized trials. Measured against the studies included in 29 systematic reviews, the share of studies found rises from 45% to 85%, and the screening model agrees with human annotators as well as annotators agree with each other. This gain comes at a cost in speed: at the median, retrieving and screening evidence for a query takes 77 seconds, against 6 seconds for taxonomy retrieval. The results suggest finding evidence with embeddings, organizing it with taxonomy labels, and setting how many candidates to screen according to the time available for each query.
Zhuoyuan Hao, Zhuo Li, Wu Li, Fangming Liu, Min Zhang, Jing Li
Recent work improves reasoning in large reasoning models (LRMs) by test-time scaling through self-consistency and parallel thinking, adding generic "thinking tokens", or presenting the question twice, as in re-reading (RE2). However, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain, and often ignore, the spontaneous repetition that many LRMs exhibit at the head of their long chain-of-thought (CoT). In this work, we analyze and harness the model's tendency to restate the question, dubbed Echo of Prompt (EOP), as a front-loaded, compute-shaping mechanism, which can be viewed as self-generated prompt repetition inside the reasoning trace rather than in the input. Specifically, we formalize its probabilistic cost by casting echo removal as rejection-based conditioning, and define the Echo Likelihood Gap $Δ\mathcal{L}$ as a computable proxy linking early repetition to likelihood gains and downstream accuracy. We find that (1) in a causal intervention, reinserting an echo into failed GSM8K traces raises exact match by 7.9-10.4 points for two reasoning models but not for a non-reasoning base model; and (2) in length- and suffix-controlled likelihood analyses and layer-wise attention studies conducted to show that EOP's benefits go beyond verbosity, EOP increases answer to answer-prefix attention in middle layers, consistent with an attention refocusing mechanism. Building on these findings, we develop Echo-Distilled SFT (ED-SFT) to instill an "echo-then-reason" pattern through supervised finetuning, and training-free Echoic Prompting (EP) to re-ground the model mid-trace. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and find consistent gains over baselines, including thinking-token insertion. Our code is publicly released at https://github.com/hhh2210/echoes-as-anchors.
LLMs increasingly fact-check online content with the poster's identity in view. However, it remains unclear which post-training changes how much a model's truth judgement depends on the poster's social class, and how such a change happens inside the model. This paper investigates job-based doubt, i.e., judging a claim false more often when the poster has a lower-status job, using fact-checked claims with randomised poster profiles in base and post-trained checkpoints of six model families. Our results reveal that: 1) post-training creates such doubt only with the open Tülu-3 recipe (OLMo-2 32B, Tülu-3 70B), not with Meta's, Alibaba's or Mistral's; the effect is under one percentage point but statistically reliable, exceeding training drift and replicating on independent claim sets; 2) in OLMo-2 32B, the model does not start noticing the poster's job more, but starts interpreting it differently: a lower-status job pushes the base model toward "true" and the post-trained model toward "false"; and 3) this flip is carried by a single residual-stream direction and happens at supervised fine-tuning. The mechanism was confirmed on 400 new claims with pre-specified predictions. None of seven deployed commercial models doubts lower-status posters more. Our work underscores the need to audit speaker bias per training recipe.
Hazel H. Kim, Jakub Binkowski, Ekaterina Fadeeva, Juan David Guerra, Yijie Hao, Zheng Yi Ho, Baixiang Huang, Buyun Liang 等 25 位作者
Hallucination remains one of the most critical challenges in deploying large language models (LLMs) reliably and safely. Yet the term is used inconsistently across the research community. We argue for a strict definition of hallucinations in LLMs. First, hallucinations should be confident incorrectness, kept separate from expressed uncertainty (UNK)1. UNK reveals that the model may be wrong, whereas a hallucination does not reveal it and is therefore more dangerous in practice. Conflating the two obscures their distinct behaviors and makes them hard to study. Second, hallucinations should be split by a simple re-elicitation test: does any semantically equivalent prompt recover the correct response? This test probes whether the relevant knowledge2 is accessible. If none does, the knowledge is likely absent or inaccessible, and the error type is a confabulation (CF) that likely requires external knowledge or retraining; if some does, the knowledge is accessible from the model but applied inconsistently, and the error type is a semantic illusion (SI) that simple prompt or decoding changes can mitigate3. Re-evaluating 14 recent hallucination methods under this taxonomy yields three findings. (1) Detection methods flag UNK more often than hallucinated responses (i.e., SI and CF). (2) Mitigation methods are far more effective for SI than for CF. (3) Latent signals are more informative than output-level signals for CF. An audit of recent hallucination papers shows that most conflate these categories. We recommend reporting CF, SI, and UNK separately, and introduce a practical labeling protocol that needs only a modest number of semantically equivalent prompts.
Jiyoun Kim, Muhan Yeo, Eunhye Jang, Jeewon Yang, Hangyul Yoon, Su Ji Lee, Hee Jo Han, Hee-Jae Jung 等 17 位作者
Discharge summaries are crucial clinical documents containing the context of a patient's overall hospital stay, and are routinely reviewed by medical experts for patient readmission, ongoing care, and diagnostic decision-making. When reviewing them, medical experts often must iteratively synthesize information across multiple discharge summaries while verifying the evidence supporting each answer. Although large language models (LLMs) are increasingly explored for clinical question answering, existing benchmarks do not sufficiently reflect this setting: they often evaluate exam-style medical knowledge or focus on single-turn question answering with limited evidence-grounding evaluation. We introduce EHRNote-ChatQA, the first benchmark for evidence-grounded multi-turn clinical question answering over patients' multiple discharge summaries. Built from de-identified MIMIC-IV discharge summaries, EHRNote-ChatQA contains 967 patient-level multi-turn samples spanning one to five notes and 16,072 medical-expert-verified QA pairs (8,036 content questions, each paired with an evidence-grounding question) across eight clinical categories. The benchmark is constructed through an expert-informed pipeline combining discharge-summary structuring schema, expert-curated multi-turn QA templates, and LLM-based generation, followed by review and revision of every single QA sample by 12 medical experts (11 physicians and 1 nurse). Benchmarking 24 proprietary and open-weight LLMs reveals several challenges, including that LLMs struggle more with evidence grounding than content answering, multi-turn errors compound across turns, and single-turn clinical QA performance does not reliably transfer to this setting. These findings establish EHRNote-ChatQA as a rigorous and practical benchmark for evaluating clinical QA systems.
Parametric knowledge in large language models (LLMs) is a cornerstone of their success, yet remains poorly understood. Existing knowledge benchmarks typically rely on predefined questions (e.g., "What is the birth date of M.L. King?"), evaluating only knowledge that benchmark designers explicitly choose to query, a problematic availability bias. In this paper, we introduce open knowledge evaluation, a new paradigm for benchmarking the stated knowledge of LLMs. Instead of asking narrow questions, it evaluates models on the knowledge they choose to surface in response to open-ended elicitation prompts (e.g., "Tell me everything you know about M.L. King"). This shifts the focus from predefined answer retrieval toward characterizing the knowledge models naturally express. We instantiate this paradigm with BeQu (Beyond Questions), a benchmark of 10,000 entities paired with reference corpora for statement verification. Using BeQu, we evaluate a broad range of language models and analyze the effects of reasoning effort, model scale, prompt format, and knowledge domain. Data and leaderboard are available on this work's GitHub repository and at bequ.scads.ai, the benchmark website.
Video captioning models are increasingly evaluated with metrics designed for single-sentence image captions. We test whether those metrics recover human judgement on dense, multi-window video description, and find that they do not. Six frontier models captioned 196 short clips. Trained annotators post-edited every caption against a failure taxonomy fixed before annotation, once as a whole-clip overview and once window by window, and a quality-control reviewer checked each annotation, yielding 1,144 overview verdicts and 4,672 window verdicts. Reference-overlap metrics do not separate the captions reviewers accepted from those they rejected: BLEU-4 reaches an AUC of 0.510 and a CIDEr-style tf-idf cosine 0.528, and restricting references to reviewer corrections leaves BLEU-4 at 0.479 (95% CI 0.431-0.529), indistinguishable from chance. Review granularity changes what is detected. The same annotators record camera and shot errors on 2.7% of captions when reading the overview and on 12.8% when reading window by window, a 4.7x gap that holds across every robustness check we ran, on the mode most likely to need a near-total rewrite. Correct-overview rates span 54.1-72.7%, but only 4 of 15 pairwise gaps survive Bonferroni correction (5 after adjusting for reviewer and clip effects), and accuracy and derived severity order the models differently. Models with statistically indistinguishable accuracy have sharply different failure profiles, which no single scalar shows. We argue that caption evaluation should record what was wrong rather than only how often, and we report the taxonomy, the per-model profiles, the corpus properties that predict difficulty, and the code that regenerates every number in the paper.
Why do multilingual language models generalize across languages even without shared vocabulary? The Platonic Representation Hypothesis posits that models trained on different data, objectives, and modalities converge toward a shared statistical model of reality. Extending this intuition to representations inside a single multilingual model, we propose the Relational Identifiability Hypothesis: relations in a shared world can promote entity alignment by ruling out alternative cross-lingual correspondences, making corresponding entities more likely to become directly readable without an additional transformation. We test this hypothesis in a controlled bilingual Transformer whose two token-disjoint languages describe the same modular-arithmetic world. We find that the languages can develop a shared answer subspace without aligning corresponding numbers. Cross-lingual alignment emerges selectively when the available relations constrain number correspondences: subtraction makes zero task-level identifiable, inducing selective alignment, while an order-sensitive operation produces broad alignment across numbers. These patterns predict success in zero-shot cross-lingual transfer. Together, our results show that relational structure in a shared world can promote entity alignment and enable transfer.
Large Language Models (LLMs) often struggle with temporally grounded reasoning because their static parametric memory is misaligned with the dynamic evolution of real-world knowledge. Although RAG and Parameter-Efficient Fine-Tuning (PEFT) provide practical ways to enhance LLMs, they remain insufficient for temporal reasoning. Existing retrieval methods typically prioritize semantic relevance while overlooking temporal validity, event order, and evolutionary dependencies, resulting in evidence that is relevant in isolation but temporally fragmented. Meanwhile, conventional PEFT methods usually employ fixed and input-agnostic low-rank updates, making it difficult to capture time-dependent entity semantics and relation drift. To address these limitations, we propose Temporal Evidence-Parameter Co-Alignment (TEPCA), a framework for LLM reasoning over TKGs. TEPCA aligns LLM reasoning with temporal knowledge at both the evidence and parameter levels. At the evidence level, it introduces time-constrained Dynamic Relevance-Guided Path (DRGP) pruning to preserve globally coherent temporal paths, together with Mutual-Information (MI)-based pruning to retain query-informative local facts under temporal constraints. At the parameter level, it integrates Dynamic-Gated LoRA (DG-LoRA), a timestamp-conditioned low-rank adaptation mechanism that dynamically modulates the effective adaptation subspace according to temporal context. This design enables LLMs to better capture fine-grained entity evolution and time-dependent relation semantics while preserving parameter efficiency. Experimental results on two TKG benchmarks demonstrate that TEPCA achieves state-of-the-art performance. Additional evaluations of DG-LoRA on eight commonsense QA datasets further verify its general input-adaptive capacity, scalability across LLM backbones, and interpretability.
Mid-training (MT) is a training phase in LLM training pipelines that continues pretraining (PT) on smaller datasets of higher value. Taking the PT and MT phases as a whole, training usually starts with only PT data and then transitions completely to MT data. In this work, we view this complete transition from PT to MT data as a special case of a broader class of data mixing schedules that specify how the proportions of PT and MT tokens in each batch change over training. Under fixed total training compute and MT token budgets, we ask: Is this complete transition to MT data optimal for minimizing final MT loss? We show that the answer depends on the learning rate (LR) schedule: one data schedule may outperform another under one LR schedule, but changing only the LR schedule can reverse the outcome. Specifically, when the LR decays throughout continued training, introducing MT data earlier and mixing it with PT data lowers the final MT loss compared with placing all MT data at the end of training, but this result is reversed if we use constant LR. We therefore study how to jointly choose when to introduce MT data and when to decay the LR through ablation studies, and find that it is better to decay the LR only after MT data enters training. Finally, we develop a theoretical model that explains why the optimal data schedule for a fixed MT token budget depends on the LR schedule.
Xiaoxi Kang, Lizhen Qu, Lay-Ki Soon, Adnan Trakic, Terry Yue Zhuo, Patrick Charles Emerton, Genevieve Grant
Large Language Models (LLMs), such as ChatGPT, have drawn a lot of attentions recently in the legal domain due to its emergent ability to tackle a variety of legal tasks. However, it is still unknown if LLMs are able to analyze a legal case and perform reasoning in the same manner as lawyers. Therefore, we constructed a novel corpus consisting of scenarios pertain to Contract Acts Malaysia and Australian Social Act for Dependent Child. ChatGPT is applied to perform analysis on the corpus using the IRAC method, which is a framework widely used by legal professionals for organizing legal analysis. Each scenario in the corpus is annotated with a complete IRAC analysis in a semi-structured format so that both machines and legal professionals are able to interpret and understand the annotations. In addition, we conducted the first empirical assessment of ChatGPT for IRAC analysis in order to understand how well it aligns with the analysis of legal professionals. Our experimental results shed lights on possible future research directions to improve alignments between LLMs and legal experts in terms of legal reasoning.
Chumeng Liang, Linxuan Wang, Xinyu Peng, Huabin Liu, XiaoXu Guo, Yuxin Chen, Ge Liu, Guang Lin 等 10 位作者
Speculative decoding accelerates language model inference by verifying multiple draft tokens in a single target-model pass. Recent parallel drafters have achieved breakthrough performance in frontier production models, but their effectiveness deteriorates as the entropy of target distributions increases due to insufficient draft diversity. To overcome this bottleneck without sacrificing parallelism, we introduce UBTree, a parallel drafter that couples a Unigram proposer with a Bigram selector to construct drafting Trees. The unigram proposer is trained with the standard cross-entropy objective to generate candidate tokens independently for each position, while a lightweight bigram selector predicts transition scores between adjacent candidate pairs. Unlike the proposer, the selector is trained with a renormalized KL objective on high-temperature data. This tree-native training broadens the supervision beyond the greedy path, encouraging plausible alternative branches that improve the chance of accepting additional tokens during tree verification. Across seven standardized benchmarks with Qwen3-4B and Qwen3-8B, UBTree achieves an average speedup of $5.84$--$6.94\times$ over autoregressive decoding and outperforms DARTree in all 28 comparisons. Production-scale evaluation further demonstrates UBTree's advantage over frontier baselines such as DSpark.