Papers

Enterprise conversation analytics asks many questions of millions of interactions. Each question can require reconstructing what people mean and identifying which information matters, repeating costly interpretive work across the same transcripts. We propose a simple principle: clarify the text, then focus the reader. Statement normalization transforms dialogue into short, speaker-attributed statements with source references and semantic tags. The statements make meaning more explicit; the tags support selecting evidence for a particular question. Downstream models can use the full representation or a relevant subset, depending on what helps them make the decision. In an offer-suppression task on customer-service calls, normalization improves a supervised classifier without selection, while weaker prompted readers benefit from both normalization and selection. A small model can learn the normalization contract, while lightweight encoders handle tagging and downstream decisions. Sharing this preparation across questions supports an inference pipeline built entirely from small models, making analytics over millions of conversations substantially less expensive.

Conversational task disambiguation over tabular data uses dialogue to resolve missing information about a user's intended task before producing a solution over tables or databases. Existing evaluation and training lack a leakage-aware foundation. Task success mixes the agent's disambiguation and solution-generation capabilities and can also reflect oracle leakage, that is, information that a user simulator reveals beyond what a real user would. Existing datasets also lack a shared representation of ambiguities and access boundaries. We introduce the notion of an ambiguous verifiable task, which formalizes ambiguities and resolutions, decomposing the agent into an asking policy and a solution policy, and the environment into an oracle and verifier. This framework provides baselines and metrics for evaluating task disambiguation separately from solution generation, formal definitions of oracle leakage, judge-free leakage diagnostics, and a training objective for the asking policy. We instantiate the framework in text-to-SQL with AmbiTab, a benchmark suite that unifies six ambiguous datasets under a common representation specifying what the agent, oracle, and verifier may access. We evaluate clarification strategies and oracle leakage, and train an asking policy with reinforcement learning. The trained asker improves our disambiguation metrics on all six datasets and task success on five, and our leakage diagnostics measure how training affects oracle leakage.

Our work explores learning a compressed latent representation of text, at the intersection of data compression and representation learning. We propose an autoencoder architecture that performs residual downscaling and upscaling of hidden representations along the time axis, with a residual low-dimension discrete bottleneck. We analyze our approach for different quantization methods, training objectives, and datasets. For different levels of compression, we evaluate the similarity between the original and reconstructed text both at the surface-level (BLEU) and at the semantic-level (LLM-based judge). Additionally, we evaluate our models on downstream question-answering and semantic text similarity benchmarks. Our approach results in compressed representations which are on par with lossless text compression algorithms at 2.24 bits per byte on web text data, while having good reconstruction and downstream task performance.

How does the human mind represent semantic categories? Why do natural languages favor certain meanings over others? Prior explanations have relied on logical definability and complexity, but these are highly sensitive to the choice of logical language, rendering some design choices unmotivated. In this article, we propose that machine learning provides a somewhat more agnostic approach to measuring semantic complexity. We review emerging evidence that logic and machine learning often yield converging results on relative complexity and its resulting effects in semantic typology. Where they diverge, learning appears to be a better explanation than logical complexity. We argue that treating machine learning models as “cognitive thermometers” enables a unified approach to complexity that bridges symbolic logic and connectionist AI.

Work zones are critical yet hazardous components of transportation infrastructure, requiring carefully designed Transportation Management Plans (TMPs) to ensure safety and mobility. However, TMP preparation remains labor-intensive and heavily dependent on practitioner expertise. This paper proposes a Large Language Model (LLM)-assisted framework to automate TMP content generation, leveraging the WisDOT WisTMP system as the application context. The framework fine-tunes multiple open-source LLMs across different model scales and deploys them locally to ensure data security. To support model training, we construct a domain-specific dataset from historical WisTMP documents by converting PDF files into structured question-answer pairs in JSON format. Experimental results show that fine-tuning significantly improves performance across standard text generation metrics. Further section-wise and strategy-level analyses reveal that, while LLMs achieve strong overall performance, they tend to over-generate strategies and struggle to produce project-specific justifications and accurate cost estimates. In addition, scaling from 7B/8B to 14B yields limited gains. These findings demonstrate the potential of LLMs to improve TMP preparation efficiency while highlighting remaining challenges in LLM-assisted TMP development. The source code and demo videos will be publicly available at https://zihaosheng.github.io/TMP-LLM/.

Looped Language Models (LoopLMs) offer a parameter efficient approach to scaling reasoning by reusing shared parameters across recurrent computation steps. Despite their promise, effective post-training of LoopLMs remains challenging. Existing approaches either provide reward based supervision that is sparse or costly to extend across loops, or rely on external teachers or privileged information, leading to limited teacher availability or teacher-student context mismatch. To address these limitations, we introduce LoopOPD, a cross-loop on-policy distillation framework that uses additional recurrent computation within a LoopLM as its own source of supervision. LoopOPD uses a frozen terminal loop policy as a compute privileged teacher for an intermediate loop student on student generated rollouts, providing dense supervision without an external teacher or privileged information. We further propose Dynamic LoopOPD (D-LoopOPD), which continually refreshes the terminal loop teacher as the shared model parameters are updated, enabling recurrent self-improvement. We characterize how distillation updates propagate across loop depths and derive sufficient conditions under which a single update yields simultaneous local improvement at both loop depths. Experiments on Ouro-Thinking models show that LoopOPD improves mathematical reasoning, while D-LoopOPD yields further gains through dynamic teacher updates. Despite being trained only on mathematical data, the resulting models also improve on general reasoning and code generation benchmarks, demonstrating that recurrent computation can serve as an effective source of supervision for LoopLMs. Our code and model checkpoints will be released upon acceptance.

Large language models can now write complete, interactive 3D worlds as code, but grading those worlds automatically is unreliable. Existing judges take one view of the output: a vision-language model scores a few rendered snapshots, or a language model reads the source. On worlds written by five frontier models we find that the two views disagree on 32% of required items, mostly code that no frame shows, and that fixed views miss small close-up contents. We present WorldBench, a benchmark and judge for open-ended, LLM-generated Three.js worlds. From one prompt describing a floating voxel island with ten biomes, physics, and day/night and seasonal cycles, the judge explores the running world, controlling its clock, orbiting it, and sending a navigator agent to frame each biome, and reads the code for what it sees. Neither channel is trusted on its own: a code quote counts only if it is text the source contains, and visual claims are checked against measured pixels where the property is measurable. A mutation test, in which we remove features by construction, shows that code-only judging gives full credit to four of five removed features, because their code remains in the file. Our judge cuts the points kept on removed features by a third (5.44 to 3.55 of 7.11), and what it still credits is mostly code that exists but never runs. We evaluate five frontier models: Claude Fable 5.1, GPT-6 Astra, Kimi K3, Grok 4.7 and Gemini 3.1 Pro. Code, prompt, tests and judge configuration are available at https://github.com/KrishBakshi/worldbench

Modern language models undergo reinforcement learning with verifiable rewards (RLVR) on top of already-trained checkpoints. A key promise of RLVR is the discovery of new reasoning strategies. In principle, a model can sample novel ideas absent from its prior training data. In practice, however, augmenting RLVR with strong novelty incentives has seen limited success and can degrade model quality. Because verifiable rewards supervise only a narrow slice of the model's knowledge and behavior, such degradations are difficult to recover from. Instead, we decouple exploration from optimization in a framework we call Exploration-Distillation (ExpDis). We train one or more explorer policies with a novelty bonus in the reward, filter their trajectories for correctness and quality, and distill them into a separate student policy. The student policy is then trained without a novelty bonus. We repeat the above procedure for several rounds, alternating between exploration and optimization. This decoupling allows us to aggressively scale exploration without degrading the student policy. Across seven mathematical reasoning benchmarks and two model families, ExpDis outperforms DAPO at the same wall-clock budget. Moreover, we observe improved pass@$k$ scaling, indicating that ExpDis produces models that generate more diverse correct solutions.

Conditional memory architectures such as DeepSeek Engram use input n-grams to look up learned embeddings, expanding the capacity of large language models (LLMs) with limited additional computation. Beyond model scaling, this architecture has demonstrated the potential to decouple factual knowledge storage from general-purpose computation, offering a promising route to updating factual knowledge while keeping the Transformer backbone fixed. Realizing this potential is challenging because different expressions of a fact may activate different n-gram embeddings, while updating shared embeddings can unintentionally change the model's predictions about other facts. We propose EngramEdit for decoupled knowledge updates through conditional memory. EngramEdit first computes target memory representations that make the model predict the updated fact across multiple expressions. It then jointly updates the shared n-gram embeddings to match these targets across expressions and edits, penalizing updates to frequently reused embeddings more strongly to preserve unrelated knowledge. Experiments show that EngramEdit enables independent factual knowledge updates through conditional memory, achieving near-perfect editing success. Revised knowledge is usable across unseen expressions and in multi-hop reasoning, with nearly three times the strongest baseline's accuracy under chain-of-thought (CoT) prompting. Unrelated knowledge and general capabilities are largely preserved even as factual updates accumulate. These findings show that EngramEdit turns conditional memory into an editable knowledge interface, extending its role beyond model scaling to support decoupled knowledge updates.

Vision-language-action models (VLAs) are strikingly sensitive to instruction phrasing and do not inherit the language robustness of the vision-language models they are built on. A one-word edit can move success by tens of points: $\pi_{0.5}$ turns on a LIBERO stove 100% of the time for "switch on the stove" and 2% for "switch on the hot plate", and a $\pi_0$ checkpoint finetuned with rephrase augmentation still shows swings of up to 61 points. We characterize this sensitivity with statistically tested single-edit swings and an oracle phrase search, which shows that phrasing alone nearly closes the 21-point gap between in-distribution and out-of-distribution tasks. We then reduce it without modifying the policy. Because the sensitivity is systematic, it can be expressed as explicit rules: we score many phrasings of a few training tasks, have a large language model distill the evidence into ten to twenty rephrasing rules, and at deployment rewrite each incoming instruction once under these rules. The rules improve the frozen $\pi_0$ by 16 to 27% relative on twelve held-out tasks across adversarial, VLM-generated, and human-generated phrasings, with gains concentrated on out-of-distribution tasks. The pipeline replicates on $\pi_{0.5}$ and LIBERO, lifting in-finetune success from 93.6% to 97.8%. The method requires no retraining and no per-step verification, and applies zero-shot to unseen tasks and instructions. Project website: https://sttawm.github.io/rephrase-before-you-act

Prompted embedding models have recently received increasing attention, particularly for retrieval, where detailed retrieval instructions are provided as part of the retrieval prompt. Several new datasets and studies have examined this setting, showing that the current embedding models often struggle to follow such instructions reliably. In this paper, we study the mechanism of how instructions actually affect the representations of retrieval queries in asymmetric retrieval tasks. We show that models can fail to follow even simple task instructions when query-side distractors are included in the evaluation. We hypothesize that this behavior is driven by the training setup of current embedding models and their evaluation, and show that fine-tuning with added query-side distractors leads to substantial improvements, with minimal effect on other tasks.

Many of the questions now put to large language models have no correct answer to score against: what a policy is worth, which option a user should choose, how to weigh competing values. Stated-preference economics has faced this problem for decades. It judges survey responses without knowing the true value, through a framework of validity and related concepts: content, construct, and criterion validity, reliability, incentive compatibility, and consequentiality. We argue that this framework is a general method for evaluating language models, and we set out what each concept means for LLM evaluation. We demonstrate the approach using a published water-quality stated preference economic valuation survey (Vossler et al. 2023) administered to six models. In this economic application, the validity tests take the form of predictions from economic theory: demand should slope down, and willingness to pay should respond to the scope of the good and to income. The tests separate the models sharply. Two older models fail the most basic test at a household income level of \$75,000, and the two newest pass every test of theoretical validity we can score, but diverge on convergent validity. Passing validity tests shows that a model's answers are coherent, not that they are correct.

Hallucinated information can propagate through multi-stage LLM systems and become part of the context for subsequent reasoning. Existing studies of post-hallucination reasoning (PHR) mainly characterize changes in final outcomes and aggregate reasoning dynamics, leaving how models resolve hallucinated premises at the response level insufficiently understood. In this work, we introduce PHRBench, a controlled benchmark for behaviorally structured PHR across four domains and 18 large language models. PHRBench characterizes each reasoning trajectory independently of final-answer correctness through Hallucination Compliance, Hallucination Avoidance, and Heuristic Correction, and defines an insightful trajectory as successful correction that ultimately reaches the correct answer. Across 4820 controlled instances, we find that successful recovery remains relatively rare and is associated with more frequent belief updates along the reasoning trajectory. We further find that properties of the hallucinated prompt contain substantial predictive signal for successful recovery, with a lightweight predictor achieving an AUROC of 0.847. These findings provide a behavioral view of post-hallucination reasoning, characterizing how LLMs resolve erroneous context and when successful recovery is likely to occur.

Much knowledge work produces new deliverables from files a workspace already holds, and LLM agents are beginning to take such work over. Through direct corpus interaction, an agent can search and read any of those files from a terminal with no indexing, and producing a deliverable from many of them in this way is what we call direct workspace interaction (DWI). Reaching the files, however, is only half the task: nothing keeps track of what the task asks for, what has been read, and what was listed but never opened, all of which slip through the context window without leaving a trace, so an agent may extract a figure and still deliver a report without it. To address this, we present RunningTab, a framework that equips direct workspace interaction with an environment-side tab: a per-task record of what the task still owes, kept by the environment alongside the agent. Specifically, the agent adds its requirements, while the environment records every file read as an excerpt with its provenance and every listed but unopened file as a candidate; the agent can then see each requirement beside its best-matching excerpts and top unopened candidates, resolve it against matching content or set it aside with a reason, and, should it try to finish with requirements still open, receive them in a finish check. We validate RunningTab on three benchmarks with three LLMs, where it consistently outperforms plain DWI and baselines that keep the record in the model, while its tab usually holds the values a deliverable needs once seen.

Terminal-agent capability depends jointly on model weights and the runtime harness that formats prompts, binds tools, and handles error recovery. Existing harness-model co-evolution approaches improve both components, yet often treat trajectories produced during harness search as an undifferentiated replay buffer. This practice overlooks that a trajectory's value for model training depends on the harness under which it was generated. To systematically analyze this interface, we establish an alternating co-evolution framework that decouples harness search and policy training through component-wise promotion decisions. Within this framework, we introduce CoTrace, a harness-aware data recipe that explicitly governs trajectory routing, provenance matching, and curriculum refresh. Under CoTrace, recurring execution failures guide harness synthesis, while policy training is strictly conditioned on verified rollouts matched to the adopted runtime for supervised fine-tuning (SFT) or fresh online interactions for reinforcement learning (RL). On the Tmax promotion split, CoTrace advances Qwen3.5-9B from 78 to 88 solved tasks under supervised fine-tuning while an online reinforcement variant reaches 90. Specifically, a compact harness-matched corpus produces steady model gains at substantially lower compute than much larger corpora pooled across sibling harnesses. Furthermore, evaluations on Terminal-Bench 2.1 and SWE-bench Lite show that out-of-distribution transfer depends fundamentally on harness compatibility, where maintaining consistency between training and evaluation runtimes prevents procedural execution breakdowns observed under foreign scaffolds.

When reinforcement learning teaches a language model a new behavior, can we find the training rollouts that taught it? And when an attribution method says it can, how do we know the answer is real? We study both questions on online RL fine-tuning with GRPO, using a planted behavior with a known cause. We release BehaviorTrace, an open evaluation harness that combines full-gradient sketching, the planted-behavior setup, and controls for gradient magnitude, fluency, headroom, and variation across seeds and generation draws. Across three seeds on Qwen2.5-1.5B, much of the apparent attribution signal comes from confounds. A control that ranks training steps by gradient size alone, with no behavior target, reaches 4.2 to 4.5 times chance and matches or beats the best targeted estimator on two of three seeds. At saturated checkpoints, model fluency predicts the behavior label at least as well as every gradient method we compared it with. Once fluency is controlled, the per-rollout results change from seed to seed and from one generation draw to the next, so a single run cannot settle the question. One signal does hold on all three seeds. The gradient of the trigger tokens aligns with a target built where the behavior actually occurs. We turn these findings into a checklist for evaluating attribution in RL. We test existing estimators, including GAS (renormalized TracInCP) and a TRAK-style estimator, and do not propose a new one.

Speculative decoding accelerates large language model inference by using a low-cost draft model to propose tokens that the full-size target model verifies in parallel. Parallel and semi-autoregressive (semi- AR) drafters improve drafting efficiency by proposing an entire block in a single forward pass, but training them raises a new difficulty: the draft distribution for a given position depends on where the decoding round starts, and where rounds start depends on how many tokens earlier rounds accepted. Existing training objectives typically rely on block-local surrogates that ignore this cross-round coupling, and therefore do not directly optimize the global decoding efficiency. In this work, we develop a theoretical framework for training and evaluating these drafters by representing speculative decoding as a Markov reward process. This formulation yields the Expected Decoding Rounds (EDR) objective, which weights local rejection costs by state occupancies and exactly equals the expected number of decoding rounds. Unlike prior surrogate objectives, EDR introduces no auxiliary hyperparameters. We then derive an exact temporal-difference gradient that supports unbiased stochastic optimization from target-model rollouts. The same framework also yields an exact offline evaluator for round counts, enabling paired drafter comparisons on shared target rollouts without running speculative decoding. Finetuning two state-of-the- art drafters, DSpark and DFly, with EDR consistently improves mean accepted length and outperforms existing training objectives across nine benchmarks spanning math reasoning, code generation, and chat.

Monitoring the chain-of-thought of reasoning artificial intelligence (AI) models remains a key approach to detecting deception and other forms of misbehavior in such models. However, semantic chain-of-thought monitoring depends on reasoning traces being legible and sufficiently faithful to the underlying computations that produced the model's behavior, not to mention accessible. Moreover, there is increasing evidence that chain-of-thought outputs may soon become illegible or unfaithful, if they even remain accessible. Based on cognitive load theory, we investigate a lower-bandwidth signal–the number of reasoning tokens generated–which does not require access to the content of the reasoning trace. Three reasoning-capable large language models answered 210 multiple-choice questions–across analytic, descriptive, and normative reasoning types as well as moral and non-moral domains–under system prompts instructing them to respond truthfully, falsely, or without regard for truth. Across all three models, truth-directed responses elicited fewer reasoning tokens than both lie-directed and truth-indifferent responses. In addition, in one model, untruthful instructions increased the likelihood of reasoning at all. These findings show that explicitly prompted untruthful response policies can produce robust group-level differences in test-time reasoning-token use. While not yet establishing reasoning-token count as a detector of spontaneous deception or general misalignment, our results are a proof of concept that it can serve as a simple, content-independent black-box candidate signal for differentiating untruthful from truthful model behavior when raw reasoning traces are unavailable or unreliable. Future work should test instance-level detection rates, out-of-distribution generalization, learned deceptive policies, hidden objectives, and robustness under adversarial pressure.

Document-level text simplification involves transformations that go beyond sentence-internal edits, addressing discourse coherence, anaphora resolution, and cross-paragraph consistency. Despite advances in sentence-level simplification for high-resource languages, document-level simplification in morphologically rich, low-resource languages such as Estonian remains largely unexplored. This study presents a comprehensive evaluation of five state-of-the-art multilingual large language models (LLMs) for document-level simplification in Estonian. Three prompting strategies are examined: single-pass generation, pipeline-based modular agents, and guideline-augmented pipelines. The evaluation framework integrates automatic metrics assessing readability, semantic preservation, and discourse coherence, alongside a structured manual annotation protocol. The findings indicate that Gemini-2.0 and LLaMA-3.3 produce outputs with near-native fluency and strong meaning preservation, whereas other models display notable grammatical and semantic limitations. This work contributes novel document-level coherence metrics, evidence-based prompting strategies, and publicly available resources for reproducibility.

Adaptive computation aims to improve language-model inference by tailoring execution to each input. For layer programs, oracle evaluations use known answers to estimate the potential gain from this flexibility, before a practical selector is available. However, a gain from selection does not by itself explain why the chosen programs help. This study examines this distinction using 32 layer-skipping and repetition programs on two models and 4,413 multiple-choice items. The analysis compares their gains over a fixed action selected without the evaluation prompt with those of input-blind perturbations at the same sites, re-evaluating selections on another prompt. With shared option order, the controls give 10.2-11.8 and 15.6-19.4 percentage points of headroom on Qwen3-4B-Base and Llama-3.1-8B, exceeding the real programs' 9.0 and 10.1 in all three random-direction draws per model. They match answer-change rate only, and the ordering depends on the menu: in post hoc comparisons, real programs lead on Llama's repeat-only menu in every draw. A smaller KL-calibrated comparison, including an input-dependent control, favours real programs in point estimate, with inconclusive corrected tests. Fixed letter offsets produce headroom of similar scale. Rotating options sharply reduces both families' headroom, while leaving positive real-minus-control differences of 1.4-2.3 and 3.7-4.5 points; their magnitudes and statistical support depend on further adjustments and the reference. A supplementary generated-answer test finds that search-selected programs keep a 26.0-point advantage over programs selected for other problems after rewording, without a placebo comparison. These results show that substantial headroom can persist across prompts with shared option order without establishing a benefit specific to the selected layer computation; neither ordering against these controls identifies that benefit.

Learning from large-model demonstrations offers a way to train small agents that can complete recurring tasks without calling a large model at every step. A central design choice is what to retain from teacher trajectories that contain reasoning, actions, and information about task progress. We introduce Task-Progress Distillation (TPD), an offline approach that pairs each demonstrated action with a short label describing the current task stage. The student learns these compact targets and selects actions by jointly scoring admissible stage–action pairs, which a deterministic harness executes in the environment. On ALFWorld, a 1.7B student trained with 404 demonstrations achieves 72.4% mean unseen task success with either TPD or action-only supervision, compared with 48.3% for a reasoning-trained student using constrained action selection. Explicit stages provide an additional benefit at 200 demonstrations, improving success from 48.0% to 67.7% over action-only supervision. With more demonstrations, the action-only student closes the gap, and both approaches reach 76.9% at 808 demonstrations. Shared-history analyses link part of TPD's local advantage to better decisions when moving between subgoals, particularly from object acquisition to processing. These results show that compact supervision can train effective small task agents, while explicit task progress provides additional guidance at an intermediate demonstration budget.

Social media is a rich source of real-time public sentiment, but widely used sentiment analysis tools are often applied without understanding their limitations. In this study, we evaluate the inter-rater reliability of three bespoke sentiment analysis tools (TextBlob, VADER, and Twitter-roBERTa-base) and three large language models (LLMs: Qwen3-32B, GPT-OSS-120B, Llama-4-Maverick-17B) against six human raters across 100 tweets. We measured agreement using two statistical measures: Cohen's kappa for pairwise comparisons and Fleiss' kappa for multiple raters. Even among the human raters, our results showed only fair agreement, highlighting the subjectivity of sentiment analysis. Higher agreement was observed under the binary sentiment classification (negative vs. non-negative and positive vs. non-positive) than under the three-class classification across both humans and automated tools. The Twitter-roBERTa-base model showed the strongest alignment with human ratings, outperforming both bespoke sentiment tools and LLMs, particularly in distinguishing negative versus non-negative sentiment. LLMs showed substantial agreement among themselves and moderate to substantial alignment with humans, performing better in positive vs. non-positive classifications. Our findings underscore that domain-specific fine-tuning remains crucial for reliable social media sentiment analysis, and human-centered evaluation remains essential for establishing gold-standard labels.

We study whether latent sequence compression of prompt prefixes preserves the capabilities that large language models rely on during inference. We introduce SemanticFold, a compression scheme that folds prefix hidden states at learned boundaries, and evaluate it across five model scales: Qwen3-1.7B, Qwen3-8B, SmolLM2-1.7B, Pythia-1.4B, and Pythia-6.9B. We use a fixed-target protocol: a frozen prefix is executed natively or compressed, and both arms teacher-force identical continuation tokens. This design rules out target-selection explanations for likelihood changes. We examine five endpoint families: fixed-target negative log-likelihood, finite-label reasoning accuracy, linear probe accessibility, open-ended generation, and systems-level memory and latency. We find that compression moves these endpoints non-monotonically and that they do not share a single compression threshold. On Qwen3-1.7B at compression ratio R=1.7, compressed-minus-native mean NLL decreases by 0.135 under paired bootstrap with 10000 draws. On SmolLM2 at R=1.2, the mean change is 0.013 higher than native. On both Pythia checkpoints, NLL is effectively unchanged. An NLL decomposition separating sequence shortening from the learned residual transform shows that the favorable Qwen likelihood is attributable primarily to residual adaptation rather than to shortening alone. MLP-only, which applies the transform without shortening, achieves 0.082 lower NLL than Full SemanticFold. Linear probe accuracy and macro AUC change by less than 0.03 in absolute value across conditions, with confidence intervals crossing zero. We conclude that preservation under latent compression has no single scalar certificate: language-model fit, decodability, and reasoning behavior answer different questions and can move in different directions under the same compression operation.

An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for jailbreak repairs, where the goal is to correct a specific unsafe behaviour without changing unrelated behaviours. A patch may block exact evaluation prompts yet fail on close harmful variants, or suppress harmful behaviour by over-refusing benign prompts that share its wording or structure. Existing protocols primarily test whether models can be broken, while aggregate metrics (attack success, refusal rates, global capability) cannot distinguish selective repairs from broader local suppression. To address this gap, we introduce PatchBench, a benchmark of empirically observed model-specific jailbreak failures inducing actionable harmful answers. Starting from 27,870 prompts from 37 public datasets, we curate 15,314 English prompts and query 8 open-source instruction-tuned models. Combining WildGuard filtering, pairwise Elo ranking, and manual verification, we retain a curated bank of 400 high-confidence jailbreak failures. We further introduce PatchBench-Local, an evaluation protocol testing whether a patch is behaviourally precise. For each harmful source prompt, PatchBench-Local generates three families of local neighbours: harmful variants preserving malicious intent, benign prompts with matched structure, and benign prompts reusing key harmful terms. It evaluates harmful-neighbour correction and benign-neighbour preservation, distinguishing selective repair from broader local suppression. Evaluating four activation steering methods with PatchBench-Local and MMLU shows that global capability can remain nearly unchanged while local benign regressions are severe, confirming aggregate metrics miss important collateral damage. PatchBench-Local provides a more precise basis for developing and comparing jailbreak repair methods.

Large language models (LLMs) have already been shown to match or exceed human experts in persuasion. While their persuasive capabilities hold promise for beneficial uses such as education and health communication, they can also be used to manipulate and misinform, making their evaluation a growing priority for developers and regulators. That evaluation, however, remains fragmented: studies differ in what they treat as persuasion, and broad claims often rest on narrow, situation-specific assessments. Automated methods, often modelled on human studies, offer a way to compare such assessments directly, as they can be run on the same models at scale and can include high-risk forms of persuasion that would be difficult or unethical to test on people. In this study, we adapt nine published automated methods to a shared setup, run them on the same fifteen LLMs, and ask whether their rankings agree and why. We find that the methods agree only weakly (mean Spearman $\rho = 0.25$). Our analyses point to two contributing factors. Models that refuse some tasks but not others, directly or indirectly, lower agreement by about a quarter, and these refusals fall mostly on manipulation tasks. General capability also plays a part: most rational persuasion (non-manipulative) methods track it, whereas most manipulation methods do not. Together, these findings suggest that agreement depends more on the task a method sets than on how it scores persuasion, although this pattern is only indicative given the eight methods available for analysis. More broadly, our results suggest that persuasion scores combine a model's ability to persuade with its willingness to do so. A single score is therefore informative about its own setting, but says little about a model's persuasiveness across tasks.

While Large Language Models (LLMs) possess rich world knowledge and impressive generalization capabilities, their direct application to tabular data classification is hindered by high inference costs and limited interpretability. In contrast, decision trees are fast and transparent but often underperform in low-data regimes. In this work, we propose a novel framework that bridges these paradigms by distilling LLM knowledge into interpretable decision trees under a few-shot learning setting. Instead of directly prompting the LLM to generate full trees, which is often unstable and inefficient, we develop a three-stage paradigm that prompts the LLM to generate rules and organize the rules into a tree. Experiments on multiple real-world tabular datasets demonstrate that our method achieves superior accuracy and interpretability with significantly lower prompting overhead compared to existing baselines.

Joint spatial-geometric and analytic function reasoning requires translating a perceived spatial configuration into a symbolic function whose executed curve satisfies geometric constraints. We present GAGR-Lab, a framework for measuring this capability through Cartesian game scenes, explicit function semantics, and authoritative Rust trajectory execution. It distinguishes spatial perception, metric grounding, geometric relations, function interpretation, function construction, and constrained synthesis. We specify four configurable scene-difficulty presets and a prospective 24-cell diagnostic design, while reporting only the subset actually evaluated. A bounded pilot of one hosted model (Llama 3.2 11B Vision Instruct) using two API credentials as execution replicas yields 72 balanced games with 432 attempts, 429 valid provider responses, and no target hits; exploratory ordinary-function prompt variants also fail to hit, while the structured localization interface yields no scoreable outputs. A privileged analytic search control independently succeeds on 600 directional cases from 300 generated scenes, with exact repeatability and 1,200 successful vertical-reflection or translation checks. The framework separates serving reliability, symbolic compliance, and geometric success, and preserves exact model-visible inputs and realized paths. A staged protocol outlines diagnostic calibration, held-out replication, multi-model comparison, and paired robustness tests. The contribution is an operational research framework with an executed pilot and a clearly identified prospective study plan; the full difficulty matrix and comparative model results remain untested.

Search agents enable Large Language Models (LLMs) to iteratively retrieve and use information for complex multi-hop questions. Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising approach for post-training such agents, but its reliance on sparse, outcome-based supervision can make credit assignment difficult and limit learning efficiency. In this paper, we systematically investigate how intermediate supervision can improve reinforcement learning for search agents. We study a range of reward-shaping and credit-assignment strategies that provide learning signals from intermediate retrieval steps. Building on these insights, we develop a training framework that combines intermediate signals with final outcome rewards to improve learning from multi-step search trajectories. Experiments across multiple benchmarks under matched training conditions demonstrate improvements in aggregate search-agent performance and show that both the choice of intermediate signal and where its credit is assigned affect training behaviour. These findings show that reward design and credit assignment are important design dimensions for training effective search agents.

We introduce Hyperbolic Symplectic Positional Encoding (HySPE), grounding positional attention in non-compact symplectic transformations. While canonical Rotary Position Embedding (RoPE) parameterizes the compact, elliptic branch of $\Sp(2,\R)$ via rotations, HySPE operationalizes its hyperbolic branch via a damped symmetric composition of dual shears, yielding a conformally symplectic contraction with two spectral decay rates per channel pair. To eliminate the exponential representation drift inherent to naive absolute factorizations, we diagonalize the operator in its invariant eigenbasis and introduce blockwise coordinate rebasing with adaptive centered execution. This guarantees length-independent numerical bounds while matching cached RoPE forward latency (7.21\,ms on an RTX 4090). On TinyShakespeare, HySPE-UltraLong maintains an invariant perplexity of 4.810 up to $16\times$ zero-shot extrapolation ($L=4096$), whereas RoPE degrades to 131.198. Scaled to a 51M-parameter subword Transformer on WikiText-103 ($L_{\text{train}}=512$), HySPE closely matches RoPE in-domain while robustly extrapolating to length 8192, reducing tail perplexity by 83.9% over RoPE. While these controlled experiments establish HySPE's extrapolation robustness and numerical stability, evaluating its scaling behavior on large-scale foundation models remains an important direction for future investigation.

We present InterView-C, a German multimodal corpus of 27 survey interviews conducted entirely in virtual reality, with both interlocutors represented by avatars. The corpus aligns spoken interaction with synchronized behavioral data, including gaze, head and body movement, facial behavior, hand and finger tracking. Its reference transcripts and linguistic annotations provide a reliable interface between this multimodal spoken interaction and predominantly text-based NLP methods. This interface is important because automatically transcribing speech can distort linguistically relevant information, while downstream models trained on existing resources may additionally face transfer challenges when applied to transcribed spoken data. InterView-C therefore provides word-timed and manually post-edited verbatim transcripts for all 54 recordings, interview-item timings, questionnaire responses and negation cue and scope annotations for 1,422 sentences, 1,398 of them doubly annotated ({\alpha}=0.87 for cues; {\alpha}=0.81 for scopes). We demonstrate both challenges empirically: nine open-weight ASR systems disproportionately misrecognize short closed answers and number words, while negation models trained on existing corpora show lower and highly variable performance on our transcribed interviews than a model trained on the InterView-C annotations. InterView-C thus enables linguistic analyses of spoken interaction while retaining their alignment with rich multimodal behavior.