We conducted an interview study with twelve students on their use of generative AI in academic communication. Students delegated professional messages to AI most where the pressure to sound professional is highest: email to instructors and administrators. AI involvement ranged from correcting the writer's own text to working out and writing the message outright, and students checked AI-written text against two criteria: whether it looks like AI and whether it sounds like them. Building on these findings, we model the AI-mediated process of writing a student--instructor email at the highest level of involvement we observed, and compare it with an unaided model of writing the same messages, built from participants' accounts and a classic model of the writing process. Three differences emerge: the learning loop that builds writing skill is removed, the message is no longer written for its specific recipient, and the confidence a successful exchange returns goes to using the system rather than to the writer's own ability. From these differences we derive two risks, that individual capacities never form and that authenticity and trust in communication become work. Design can respond to both but is unlikely to be enough, so the risks also need research and policy attention.
Rubrics aim to make language-model evaluation transparent by decomposing response quality into interpretable criteria. However, natural-language rubrics are often ambiguous, require LLM judges, and typically assume criteria aggregated through linear weighted sums, limiting their ability to capture dependencies, alternatives, penalties, and override conditions. We propose ExecRubrics, a framework for representing rubrics as compact executable programs. ExecRubrics encodes evaluation logic as verifiable Python scoring functions, giving natural-language rubric intent an operational semantics: a fixed decision procedure that can be inspected, executed, and edited. On three long-form response benchmarks -- HealthBench, HelpSteer, and ArgQuality -- we show that ExecRubrics can recover substantial preference signal without an LLM judge at evaluation time. On ArgQuality and HelpSteer, the strongest executable variants are within 1.1 and 4 percentage points, respectively, of the direct GPT-5.5 agentic baseline. Executable rubrics are also considerably faster, achieving a 192x average speedup. We show that incorporating external logic and resources from text processing libraries such as NLTK and spaCy can further improve preference accuracy. Our results suggest a novel way of approaching automated evaluation, by offering a faster, more explainable, and less ambiguous alternative to black-box rubric evals, particularly in high-stakes domains such as healthcare and banking where precision and auditability are critical.
Kaustubh D. Dhole, Charles L. A. Clarke, Eugene Y. Agichtein· 0 citations
Human-like cognition does not select past experience by topical similarity alone: affective significance and unresolved conflict also shape what becomes accessible. We present PsychoAgent, a cognitive architecture for LLM agents that separates factual and affective memory and integrates both through a conflict-aware executive controller. Affective memories are first filtered by semantic relevance and then re-ranked by salience, preserving topical fit while allowing emotionally important traces to enter the prompt. Across three controlled conflict scenarios, the full architecture retrieved more conflict-critical memories than semantic-affective and single-memory RAG baselines (0.933 vs. 0.500 and 0.667), with a small semantic-similarity cost. Five blinded raters evaluated 27 outputs. After within-rater standardization, the full architecture had the highest overall mean (+0.22 SD), but corrected pairwise differences were not significant. A three-day illustrative trace further shows persistent affect, offline memory recombination, and selective memory reweighting. The findings support affect-sensitive retrieval as an inspectable mechanism for modeling human-like conflict effects in LLM agents.
Mohammad Amanlou, Parham Abed Azad, Farbod Davoodi et al.· 0 citations
A benchmark score is a measurement instrument, yet most benchmarks read each item at a single canonical surface form. We ask whether that reading is faithful: when an item's intent is held fixed and only its meaning-preserving surface form varies, does the canonical-form score estimate model behavior well, and how much of any variation is decoding/judge noise rather than signal? We instantiate this in safety, a high-stakes setting with no gold label to average toward. To avoid prior confounds, we pre-author the reformulations (refusal-free, mostly non-LLM: machine back-translation and a Matrix-Language-Frame code-switch generator) so an identical surface form reaches every model, score all responses with one human-anchored, vendor-neutral judge (Claude, kappa = 0.86 vs. human on unsafe compliance, stable across languages, cross-checked by GPT-4o), and verify intent preservation. On 370 seeds x 5 surface forms x 5 models, no single transformation is uniformly most dangerous (6 of 20 per-transformation McNemar tests survive correction, most protective). Yet evaluating only the canonical prompt underestimates unsafe compliance: the union of unsafe outcomes across forms exceeds even the worst single form by 3.3-12.9 pp, with bootstrap 95% CIs excluding zero for all five models, and 5-13% of seeds safe on canonical are unsafe under some reformulation -- above a zero stochasticity floor (canonical resampled five times at temperature 0 gives 0/370 new exposures). The size of this gap is model-dependent (largest on Gemini 2.5 Pro). One form recovers only ~53% of a model's observed unsafe surface and about three reach 85% -- a redundancy characterization of this form set, not of a defined population. A benign control (XSTest) suggests the instability is bidirectional, though the benign and harmful pools are not item-matched. We release the dataset, code, and per-response labels.
Yongxi Zhou, Junwei Yao, Yuanzhe Liu et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Large Language Models (LLMs) are increasingly used for emotional support tasks, such as negative thought reframing. This task relies on modifying cognitive appraisals, the subjective interpretation of events that elicit negative emotions, which is typically conceptualized along multiple discrete dimensions. Current LLM-based frameworks model cognitive appraisal by exhaustively evaluating all possible dimensions, but they fail to account for the varying saliency of these dimensions across different contexts. In this work, we investigate a vital yet overlooked question: "Can LLMs infer the salient appraisal dimensions from emotional support conversations?" To address this question, we introduce the AppraiSal benchmark, containing 996 emotional support conversations with human-annotated mental states, including salient cognitive appraisal dimensions. Furthermore, we propose PRISM, a multi-agent probabilistic framework grounded in Bayesian Inverse Planning, designed to improve LLMs' ability to identify context-specific appraisal dimensions. Experimental results show that PRISM brings improvements to LLMs across various sizes, particularly in identifying the most salient appraisal dimensions.
Hainiu Xu, Zhaoyue Sun, Hanqi Yan et al.· 0 citations
Large Reasoning Models (LRMs) produce long, explicit chains of intermediate steps before generating a final answer at inference time. These intermediate traces dominate latency, memory usage, and serving cost, even though the final answer correctness is not causally related to the trace correctness and the trace length is not a reliable indicator of the problem complexity. This raises a natural question: can the computation expressed in these intermediate tokens be internalized into the parameters of a language model, enabling it to produce answers directly (or with much shorter intermediate traces)? We introduce \textit{masked distillation}, a knowledge-distillation framework in which a student LLM is trained to predict only the solution tokens conditioned on the question, while a reasoning teacher provides feedback on the student's responses after conditioning on the question and its own CoT trace. We instantiate this framework in two settings: (i) a \textit{self-distillation} setting, in which the same model serves as the teacher in thinking mode and as the student in non-thinking mode, and (ii) a \textit{dual-model} setting, in which a larger reasoning teacher supervises a separate smaller non-thinking student over the solution tokens. By treating intermediate tokens as a scaffold which reasoning models use to fit over the solution tokens, We additionally vary the length of intermediate-token scaffolding the student is supervised on, interpolating between full internalization (the student emits only the solution) and no internalization (the student emits the full trace before the answer). We evaluate the framework through controlled experiments on two reasoning domains: GSM8K (grade-school arithmetic) and Countdown (a number-puzzle search task).
Enhancing the reasoning capabilities of large language model (LLM) agents requires effective orchestration of diverse expert models and tools. However, existing frameworks typically call APIs, based on coarse-grained matching between tasks and the functions of expert models or tools, while overlooking critical factors such as performance variability and cost efficiency among functionally similar alternatives. To address this, we propose Agora, a framework that uses a confidence-calibrated auction to dynamically allocate tasks to expert models and tools. By treating reasoning steps as tradeable items, Agora bases allocation on calibrated competence rather than raw confidence. Across five main benchmarks, Agora improves or remains competitive with single-model, routing, and cascade baselines under matched candidate pools.
Most agent benchmarks put facts, tools and permissions behind one interface. Real organizations spread them across people. Incognita asks what happens when the task and success criterion stay fixed but access does not. We transform eighteen customer-service tasks into three settings: direct access, one known intermediary, and six role-isolated participants whose capabilities must be discovered. Across 864 trials with four models, social access reduced success for every model; the pre-specified intervals excluded zero for two. The latest tested model, gpt-5.6-sol, achieved the highest social-access success at 0.65, a 0.11 decrease from centralized indirect access with an interval that included zero. Exploratory comparisons separated five of six model pairs under social access, while neither centralized setting separated any at this sample size. In post-hoc task-blocked tests, three pairwise interaction $p$-values remained significant after multiplicity adjustment. A reference-relative reader associates the wider gaps with failures to obtain needed information. Because the data cannot distinguish ineffective requests by the evaluated agent from inaccurate replies by simulated participants, the reader's labels describe where trajectories stopped, not why models differed.
Agent benchmarks for measuring memory largely study textual cases, in which information is deliberately extracted from the environment, written down, and then later retrieved. In other words, they assess what agents elected to record, not what they happened to see. We introduce DMV-Bench (code: https://github.com/yyyujintang/DMV-Bench), the first interactive benchmark for visual memory in multimodal agents, to study this often-neglected property. DMV-Bench is built on (1) a controlled home-furnishing e-commerce environment, supported by a catalog of 1,000 product variants, and (2) a text-leakage contract which ensures that the primary discriminative signal of each task is solely in the pixels. In DMV-Bench, agents undergo chains of autonomous shopping sessions in which every visited product image carries a unique, pre-rendered incidental cue that the agent is later asked to recall. We show that conventional solutions struggle with this task. Inspired by dual-coding theory, we propose a memory architecture that uses parallel visual and verbal codes, which we call DualMem. On DMV-Bench, DualMem outperforms a caption-only baseline and three recent multimodal agent-memory systems across multi-session chain lengths on multiple models. These gains persist even adjusting for memory-bank size and encoding-position bias. Further experiments also reveal an asymmetric division of labor between the two codes; a weighted coding scheme is often strongest. We view this as a step towards memory systems that preserve a richer record of agents' observations.
Yujin Tang, Chenming Shang, Ruize Xu et al.· 0 citations
Hybrid reasoning models can answer directly or spend extra tokens on extended thinking. A practical router should choose between these modes for each query, so easy problems avoid unnecessary reasoning and hard problems receive enough budget to finish the answer. Existing routers move in this direction, but they typically require labeled training data or fix thinking budgets up front, ignoring answer-level evidence from the model itself. We introduce DART, a training-free routing framework that samples two cheap no-think drafts, accepts direct answering when the drafts agree, and predicts a thinking budget from draft entropy when they disagree. Across the main comparisons, DART preserves or improves always-thinking accuracy in most settings while reducing thinking-token use. Accuracy improves by up to +9.0 points on Olympiad-level math and by up to +22.5 points on code under execution-based equivalence, while thinking-token use drops by 32-73%. The Stage~1 signal extends across model scales (0.6B--32B), model families, and API-only hosted settings, with no labeled data and no gradient updates required. Our code is available at https://github.com/js-lee-AI/DART.
Jungseob Lee, Seongtae Hong, Seungjun Lee et al.· 0 citations
Large language models are increasingly proposed as educational tutors, yet stronger task-solving ability does not necessarily imply stronger learning support. Motivated by recent calls to measure the social impact of NLP systems in practice, we study whether public LLM tutoring benchmarks distinguish learning-supportive behavior from mere answer production. We propose a lightweight diagnostic based on the gap between solving-oriented and pedagogy-oriented benchmark performance. Using public MathTutorBench leaderboard results, we show that these dimensions are only partially aligned: across eight publicly reported models, the correlation between solving and pedagogy composites is 0.421, and several models shift meaningfully in rank when evaluation moves from solving to pedagogy. We then analyze the public TutorBench sample and show that agency-relevant behaviors are explicitly encoded in benchmark rubrics, especially in active-learning settings that reward guiding questions, calibrated hints, and non-disclosive scaffolding. Together, these findings suggest that educational-impact evaluation should not treat task success as a sufficient proxy for learning support. We argue that public tutoring benchmarks can better support positive-impact evaluation by reporting solving-oriented and pedagogy-oriented scores separately and by making disclosure-sensitive, student-agency-preserving criteria more explicit.
Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind), cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs often become systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter: male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.
Arnau Marin-Llobet, Simon Henniger, Mahzarin R. Banaji· 0 citations
