Oral exams are compelling because (1) students can be asked to provide their reasoning for an answer and (2) their interactive nature permits follow-up questions to resolve ambiguous answers. But oral exams scale poorly. In this paper, we present scalable conversational assessment, an LLM-based system that conducts interactive text-based conversations with students to assess their ability to make and support an argument. We use a two-agent architecture to separate concerns and reduce the likelihood that answers are leaked to the student: the interviewer interacts directly with the student, but only receives guidance about the next question to ask the student; the hidden evaluator knows the goals of the whole activity, tracks the student's progress using a rubric, and guides the interviewer on when to follow-up on the current question or proceed to the next question. The system was first tested using a collection of simulated students with LLM personas of a particular ability level or adversarial tactic (e.g., answer-extractors, prompt-injectors, confident fabricators). A novel element of this evaluation was that it required the use of a fictional assessment topic to prevent personas that shouldn't have information from using knowledge present in the LLM's pre-training. The system was also deployed in a small, elective AI literacy course in both formative and summative contexts. Through manual and post-facto AI review of the transcripts we identified opportunities to improve the follow-up question generation and ensure that grading only credited ideas that the student (and not the interviewer) contributed.
https://arxiv.org/abs/2611.17220
Full-duplex spoken dialogue models learn turn-taking, overlaps and backchannels from conversations recorded with one channel per speaker, but most available conversational audio is a single mono mix. We present Kezek, a model that turns a mono recording of exactly two people into one track per person with the original length and timing; overlaps and short replies ("yeah", "mm-hm") stay on the track of the person who said them. Kezek decides who speaks when and who is who once for the whole recording (stage 1: a two-voice activity model guided by a speaker fingerprint trained to be insensitive to phase), and a conditioned time-domain separator only carries out that decision (stage 2), with outputs that always sum to the input. Standard separation metrics miss the failures that matter for dialogue data, so we report short-reply retention, leakage onto the other person's solo, channel swaps and coverage. Against the open pipeline pyannote -> DialogueSidon on the same recordings, Kezek keeps more short replies (0.83 vs. 0.69 on Open Yap test; 0.80 vs. 0.71 over the other's speech on unseen telephone speech from CallHome English), mostly because it returns the whole recording while the external pipeline drops part of it (on the audio both return, reply retention is about the same: 0.859 vs. 0.865 on Open Yap); it also leaks less, swaps channels less often, and runs 4-5 times faster (about 230 hours of audio per GPU-hour). We also report negative results and an ideal-mask ceiling that locate the remaining gap: quiet replies under a louder voice. Code and weights are released under permissive licences.
https://arxiv.org/abs/2611.16933
Printed circuit board (PCB) routing is the stage of physical layout that finds paths connecting the placed pads of each net. Public datasets in this domain are few, and most contain only the final states of routed boards, not demonstrations of the electronic design automation (EDA) engine actions that produce them. To close this gap, we propose CADET (Constructing executAble Demonstrations for Engine-grounded Training), which jointly generates synthetic boards and demonstrations built from EDA engine actions. The resulting RouteDemos provides 100,000 synthetic boards routed with zero design rule violations (DRVs), together with their demonstrations. Their distributions over 17 statistics resemble those of 142 real reference boards. CADET trains a routing policy on these demonstrations with behavior cloning and fine-tunes it with reinforcement learning (RL). On real test boards, the policy raised the rate of complete DRV-free routing from 38.0% to 67.2% over an RL-only baseline. This rate was comparable to that of the open-source autorouter FreeRouting and exceeded those of the other baselines, including GPT-6 Astra. We release the dataset and its generation code as open source at this https URL.
https://arxiv.org/abs/2611.16913
This paper investigates whether Large Language Models can answer a personality questionnaire as a specific individual would. Given a user's social-media posts, the model completes the BFI-2-60 Big Five questionnaire as that user. The predicted responses are converted into trait scores and compared with the user's actual scores. We evaluate GPT-4, Gemma, Llama 3.1, and Mistral on the MyPersonality dataset. To determine whether predictions reflect information from the user's writing, we compare them with answers produced without user text and with values derived from general population tendencies. We measure prediction error across all users and separately for individuals with low, intermediate, and high trait scores. We also test whether an LLM can express an explicit personality profile when given only the target Big Five scores, without user text, and examine how predictions change as more text is provided. The results show that user text does not consistently improve performance. Models perform better near the middle of the trait distribution, while at the lower and upper ends they often remain close to central values or perform well at one end but poorly at the other. When given an explicit profile, GPT-4 most clearly translates trait information into consistent questionnaire responses, particularly for Extraversion and Neuroticism. Increasing context does not systematically reduce error: open-source models fluctuate more, whereas GPT-4 remains comparatively stable. Overall, the evaluated models do not consistently reproduce the questionnaire behavior of specific individuals.
https://arxiv.org/abs/2611.16586
LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), an independent operational benchmark that evaluates 33 systems from 17 developers: specialized guardrails (e.g., Llama Guard 4, ShieldGemma-2B, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok 4.1 Fast), jointly on detection rate, false-positive rate, latency, and cost. We cover input/output moderation across 11 harm categories, jailbreak detection across 13 attack techniques, and prompt-injection detection, using in-house datasets built from handcrafted prompts, curated external samples, and quality-controlled synthetic generation, with synthetic moderation samples paraphrased to suppress generator fingerprints. Content-moderation runs are repeated five times and all rates reported with 95% intervals. The Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally ahead: top systems match the best generalist LLMs on detection (93-96%, no significant difference) at low false-positive rates while running ~5x faster; cost depends more on serving setup than on class. On jailbreak detection, generalist LLMs achieve higher median detection on three of four datasets and lower false-positive rates, but at higher latency and, for closed frontier models, 10-100x higher cost. On prompt injection, the two best dedicated detectors lead. We release the framework, an interactive leaderboard, and four datasets as a vendor-neutral basis for selecting safeguards under deployment constraints.
https://arxiv.org/abs/2606.20668
We present a suite of 50 unpublished research-level theoretical physics problems with verifiable answers. Without AI assistance, a typical problem in our collection could be solved by a world expert in about a day or less but would take considerably longer for experts in adjacent specialties. We test the ability of off-the-shelf AI agents to tackle these problems autonomously, without human hints. When given access to tools available in a realistic setting, including the ability to search the literature and write scripts for numerical and symbolic work, the best-performing agent, GPT-6-Astra harnessed in Codex, obtains the final answers with a near-perfect accuracy of 98% at an average cost under USD 3 per problem. On average, agents show only small changes in accuracy when denied web search and access to background research papers, suggesting significant memorization of the literature. These results suggest that state-of-the-art agents can accurately perform concrete calculations required across theoretical physics research. We comment on the broader implications of these findings and discuss possible explanations for the current lack of agent-led breakthroughs in theoretical physics despite their outstanding problem-solving abilities.
https://arxiv.org/abs/2610.16733
Low-bit quantization reduces the memory footprint and inference cost of large language models (LLMs), yet accurate FP4 quantization remains challenging. Practical formats such as MXFP4, NVFP4, and HiF4 pair a 4-bit floating-point payload with block-wise or hierarchical scale codes, making accuracy highly sensitive to how each local distribution is scaled. Default maximum-absolute-value scaling preserves dynamic range, but can spend scarce FP4 levels on rarely occupied outlier regions and leave dense tensor values poorly resolved. Existing scale-search methods minimize local element-wise reconstruction error, optimizing tensor fidelity rather than layer-output fidelity. We propose SoScale, a second-order weight scale-code search method for hierarchical FP4 quantization. SoScale searches hardware-valid scale codes under a Hessian-aware layer reconstruction objective and couples this search with GPTQ-style error recovery, selecting scales by their impact on layer outputs. Across Qwen3 models, SoScale consistently reduces downstream quantization degradation under MXFP4, NVFP4, and HiF4.
https://arxiv.org/abs/2611.16426
Optional reasoning, retrieval, web search, and tool use can improve large language model responses but incur unnecessary computation when activated indiscriminately. This study investigates whether prompt optimization can compensate for limited router capacity and whether classification gains translate into downstream efficiency. The proposed framework uses prompt-driven small language models to make four independent capability decisions, allowing multiple capabilities to be selected per query. Qwen2.5 routers ranging from 1.5B to 14B parameters are compared across five prompting strategies, followed by an evaluation of capability-specific optimized configurations using a fixed GPT-OSS-120B downstream model on 127 questions. Under minimal prompting, increasing router capacity does not consistently improve classification performance. Prompt optimization narrows the capacity gap for reasoning and calculator routing, whereas RAG and web search retain larger benefits from higher capacity. Similar classification F1 scores do not necessarily imply equivalent downstream behavior: false-positive activations increase unnecessary computation, while missed capabilities can reduce answer quality. The 14B router achieves a mean BERTScore F1 of 0.8408, compared with 0.8418 when all capabilities are enabled, while reducing average downstream token consumption by 68.7\% (8,837 to 2,764 tokens) and median downstream latency by 80.5\% (9.79 to 1.91 seconds). These savings exclude router overhead, which is measured separately in a different inference environment and increases with router capacity. The findings demonstrate that the relationship between router capacity, prompt design, and downstream efficiency is capability-dependent. Consequently, router selection must consider error profiles and routing overhead alongside classification performance.
https://arxiv.org/abs/2611.16414
Language models rely on finite token contexts, limiting access to earlier interactions. We introduce Recursive State-Augmented Language Models (Recursive SALM), a framework in which a persistent latent state evolves with each observation and generates adaptations of a pretrained language model. A modular read-transition-write architecture separates observation encoding, latent dynamics, and predictor adaptation. We develop deterministic and stochastic formulations. The experiments evaluate only deterministic SALM with frozen GPT-2 small and exclude earlier evidence from its decoder context. On a four-turn hypothesis-inference task, latent-gated LoRA reaches 89.1% accuracy, comparable to 88.0% for a trained full-context LoRA baseline, with earlier evidence available solely through a 64-dimensional state. Zeroing or shuffling the state substantially reduces accuracy. In a two-turn task with codewords unseen in training, several interfaces improve candidate ranking over chance, but rarely generate the exact codeword under the evaluated training schedule. These experiments show that persistent latent-conditioned adaptation can support sequential inference when earlier evidence is absent from the decoder context.
https://arxiv.org/abs/2611.16375
Closed-context question answering (QA) requires a system to generate an answer strictly supported by a supplied passage, or to abstain when the passage lacks adequate evidence. SQuAD 2.0 is a widely used benchmark for this setting, scored by Exact Match (EM) i.e., character-level agreement with human reference spans after lexical normalization. For generative large language models (LLMs), however, an EM failure may reflect a valid paraphrase or an alternative passage-supported answer rather than a hallucination. The best prior leaderboard EM is approximately 91%, but evaluating generative models requires separating lexical divergence from factual error. We introduce SAGE-QA (Semantic-Aware, Grounded Evaluation for Question Answering), a multi-tiered framework combining reference-based Semantic Equivalence Scoring (SES) and passage-based Re-Validation Scoring (RVS). On a fixed set of 10,000 SQuAD 2.0 validation examples, we compute EM and decision-matrix F1 from TP, TN, FP, and FN counts for zero-shot closed-weight frontier models and for open-weight models, untuned and fine-tuned. Zero-shot, Claude Opus 4.6 and Gemini 3.1 Pro reach 85.21 and 86.66 EM, outperforming untuned Llama 3 8B (56.88), gpt-oss-20b (66.28), and Llama 3.3 70B (71.39). Parameter-efficient fine-tuning adds 21.19 to 34.06 EM points, raising these models to 90.94, 87.47, and 94.24 EM, respectively. Auditing the frontier models with SAGE-QA and an external LLM judge yields dataset-wide SES scores of 92.81 for Claude Opus 4.6 and 91.67 for Gemini 3.1 Pro, and hybrid RVS scores, which combine benchmark credit with passage-based judge assessments, of 97.96 and 98.02. Fine-tuned open-weight models perform best, with fine-tuned Llama 3.3 70B surpassing the previous best known EM.
https://arxiv.org/abs/2611.16300
Floorplanning is an early step in turning a logical circuit into a physical chip: it positions large circuit modules and constrains the space available for subsequent logic placement and wiring. These decisions affect congestion and timing, but their physical consequences emerge through downstream implementation. For AI agents, the task requires translating engineering constraints into spatial decisions whose quality cannot be judged by successful tool execution alone. We present an evaluation framework that connects submitted layouts to a prescribed downstream implementation flow, checks geometric validity, measures physical quality, and reports resource use. On one industrial design whose 99 macros occupy approximately 71\% of a fixed L-shaped layout region, we compare our specialized system Nova-SI with general-purpose engineering agents and agents restricted to file access, obtaining complete physical measurements for 39 layout outputs. Under our 0--100 composite quality score, Nova-SI achieves a mean of 57.97 without online feedback from the prescribed power-grid and placement evaluation, compared with 31.17 for the strongest measured general-purpose configuration, GPT-6 Astra. The difference is supported by raw congestion and timing measurements and persists across scoring-sensitivity checks. Layout-level analysis further shows that similar global utilization can conceal substantially different local congestion. Together, the protocol and case study assess agents through the physical consequences of their delivered layouts and provide evidence about system performance on this industrial instance.
https://arxiv.org/abs/2611.12048
Modern software development relies on pull requests (PRs) that introduce small, targeted code changes. Once merged, these changes should be protected by focused tests that exercise the behavior the PR introduces, yet developers often submit PRs without such tests, leaving the new behavior unguarded against future regressions. Most test generation techniques target entire software units and optimize objectives such as code coverage rather than the behavior a change introduces, and prior diff-aware approaches have had limited success in capturing it. We propose ADAPT (Agentic Diff-Aware Patch Testing), a regression test generation approach that uses an autonomous coding agent to produce change-focused tests for PRs. Given a PR, the agent analyzes the diff between the base and head versions to identify the modified functions and their context, then plans and generates candidate unit tests. A dedicated fail-to-pass (F2P) verification tool runs each candidate on both versions and, after F2P success, reports which executable changed lines remain uncovered; guided by this feedback, the agent iteratively refines its tests. We evaluate ADAPT on the same 426 real-world PRs used by the prior state-of-the-art baseline, with GPT-4o and GPT-5-mini. ADAPT+GPT-4o reaches an F2P rate of 62.0% (264/426), compared with BLAST's reported 35.4% (151/426), while ADAPT+GPT-5-mini reaches 89.9% (383/426). On 137 PRs shared with Change And Cover (ChaCo), ADAPT achieves full patch coverage on 83.9% of PRs, compared with 79.6% for ChaCo, at considerably lower time and cost. A manual evaluation confirms that the F2P-valid tests target each PR's intended change, with mean scores of up to 4.54/5 for importance, 4.35/5 for assertion strength, and 4.49/5 for readability, often covering behavior beyond the developer-written tests.
https://arxiv.org/abs/2611.16114
We present Fikiria, a two-stage SFT+GRPO recipe for Swahili mathematical reasoning. Its reward multiplies a format gate by a weighted sum of answer accuracy and the ChrF\texttt{++} similarity of the Swahili reasoning to an in-language reference chain of thought, so that training targets the surface fidelity of the reasoning trace and not only the final answer. We train on a 2.3k-problem Swahili dataset created by translating GSM8K with GPT-4.1 and filtering with SSA-COMET. On AfriMGSM Swahili, Fikiria-AfriqueQwen-14B reaches 79.2\%. Under the same fixed prompt and strict answer-format scoring, GPT-4o reaches 68.0\%, partly because it often fails the required answer format; IrokoBench reports 78.8\% for GPT-4o using a best-per-model prompt, a figure numerically close to ours but not directly comparable. In a small human evaluation with co-author annotators, GPT-4o's outputs were preferred overall. ChrF\texttt{++} measures surface fidelity, not the validity of individual reasoning steps.
https://arxiv.org/abs/2611.16104
GPT-6 Astra can drive a robot arm directly through a fixed harness, and scores of such robot-use agents on RoboDojo have risen quickly since. Reported systems differ in harness and test subset and often change several designs at once, so it is hard to tell which choices produce the gains. We take these designs apart and test them in one framework, along the harness around the model, the action interface, the reuse of experience and the scene geometry given to the model. A lean harness that keeps the full episode history and stops each motion early goes a long way. Reusing experience changes how the model acts but is not a free lunch, and scene geometry helps most, even when depth is predicted from RGB alone. JustRUA starts from the official L3 harness and keeps only what matters most. Even with no learned policy, self-improvement loop or task-specific tuning, this simple harness raises the score on the RoboDojo simulation benchmark substantially, from 29.7 to 50.3.
https://arxiv.org/abs/2611.16009
In pre-deployment testing of AI systems such as AI agents, a developer can often deal with known failure modes (e.g., leaking API keys, deleting important databases, or installing a malicious package) by implementing fixes or making contingency plans for when the same issue might occur again. What is more challenging are the unknown unknowns: the unseen failure modes that haven't surfaced during testing. The decision to stop testing and deploy therefore faces an inherent tradeoff: continue testing to observe more behaviors, at the cost of time and resources, or stop testing and deploy, at the risk of failure modes remaining undiscovered. How then does the developer decide when they have tested enough? To answer this, we develop an algorithm that certifies when the probability of observing an unseen failure mode is sufficiently small. The algorithm constructs an anytime-valid upper confidence sequence on this quantity, so the developer can repeatedly test a system, inspect the bound, and decide to stop when it falls below a chosen threshold without invalidating the statistical guarantee. Our problem can be interpreted as a sequential variant of the missing-mass estimation problem which gave rise to the classic Good-Turing estimator (Good, 1953). On SWE-Chat, a dataset of real-world coding agent traces (Baumann et al., 2026), we implement our procedure and show that it achieves tighter upper bounds than a simple multiple-testing corrected Good-Turing baseline, and even dominates a fixed-horizon (non-multiple-testing-corrected) Good-Turing upper confidence bound, which is not anytime valid. Additionally, to better match realistic evaluation pipelines, we extend our method to handle sampling from a different task distribution and imperfect automated failure labeling.
https://arxiv.org/abs/2611.15835
Aerial imagery captured by unmanned aerial vehicles (UAVs) plays a vital role in applications such as surveillance, disaster response, and large-scale environmental monitoring. Automatic generation of textual descriptions for such imagery is essential for reducing manual inspection overhead and enabling better downstream decision-making. However, Multimodal Large Language Models (MLLMs), despite their impressive captioning capabilities, are prone to hallucination, producing descriptions that are inconsistent with the visual input. This problem is significantly amplified in aerial imagery due to its unique challenges such as extreme viewpoint variation, small objects, class imbalance, and ambiguous visual entities. Existing hallucination detection approaches either rely on computationally expensive large external models such as GPT-4, which perform poorly on aerial domains, or treat the problem as a rigid binary classification task, ignoring the inherent uncertainty in ambiguous cases. To address these limitations, we propose a lightweight, reference-free framework for hallucination detection and correction in free-form aerial image captioning. Our method leverages state-of-the-art MLLMs, specifically InternVL3 and Qwen2.5-VL, to detect hallucinated content in generated captions with minimal computational overhead. Detected hallucinations are corrected through a dedicated refinement module that improves the factual consistency of the original captions. Extensive experiments on the Labelled Illusion Dataset (LID) demonstrate that our framework significantly reduces hallucination rates while remaining practical for real-time deployment in resource-constrained environments. Code is available at: this https URL.
https://arxiv.org/abs/2611.15809
As AI web agents proliferate, combining large language models with autonomous, browser-level control, indiscriminate content scraping by web agents has emerged as a privacy and security challenge. Existing defenses, such as this http URL and active bot-blocking, are insufficient, as they are widely violated and easily circumvented. In this work, we demonstrate that AI web agents can be effectively distinguished from humans and traditional crawlers using a multi-layer fingerprint based on both network layer characteristics (e.g., TLS, HTTP) and browser interaction behavior. We implement this mechanism as a programmatic logging framework that can be deployed on a live, instrumented domain. By analyzing six prominent agent frameworks (AutoGen, Browser Use, Claude, Gemini, ChatGPT, and Skyvern), we uncover latent structural differences in how these systems assemble HTTP requests, establish TLS/HTTP connections, and execute autonomous browser actions. Feeding these multi-layer features into a decision tree classifier, our framework achieves high-fidelity identification (100% accuracy), successfully isolating distinct agent architectures and differentiating agent traffic from both human browsing baselines and legacy crawlers. Our findings demonstrate that cross-layer agent tracking provides a robust approach to content protection and web security policy enforcement.
https://arxiv.org/abs/2606.20910
Visual agents need to coordinate reasoning and action while their environment evolves during inference. We introduce LOLPlay, a benchmark in native League of Legends with 36 champion-specific tasks spanning resource management, combat, cooperation, and objectives. Agents interact through screenshots and mouse-keyboard actions; automated initialization and 82 executable subgoal checks distinguish partial progress from completion. Across 324 task evaluations of three models at three reasoning-effort settings, GPT-6 Astra achieves the highest task success rate of 25.0% at low and medium effort, and the highest subgoal success rate of 40.3% at medium effort. However, high effort improves neither metric over low effort for any model. For Astra, longer requests leave fewer interaction rounds, suggesting a cost to additional reasoning in real-time control. Trajectories reveal incomplete subgoal composition and actions executing after the observed scene changes, motivating agents that combine visual reasoning with timely action.
https://arxiv.org/abs/2611.03685
Quantization reduces LLM deployment costs, but effective strategies must balance model quality and storage. LLM-driven search can automate strategy design through repeated proposal and evaluation. This process must identify useful strategies within a limited evaluation budget. The key challenge is that \textit{actual quantization is expensive, while inexpensive proxies may not reflect quantized-model performance.} To address this challenge, we present AutoQS, an \textbf{Auto}mated \textbf{Q}uantization \textbf{S}trategy Design with LLMs. AutoQS keeps weight precision fixed and searches group-size allocations across layers and operators under deployment constraints. Evaluation history guides coordinated LLM proposals, while deterministic validation ensures compatible allocations. An inexpensive quality proxy evaluates candidates, and an candidate repository retains alternative allocations to inform subsequent proposals. Periodic actual quantization complements these proxy evaluations with measured quality and checkpoint storage, providing feedback for further search within the evaluation budget. Experiments on Qwen2.5-7B-Instruct show that AutoQS improves seven-task average accuracy over uniform-group GPTQ and AWQ by 0.70 and 1.12 percentage points, respectively, at four-bit weight precision.
https://arxiv.org/abs/2611.11975
Humans often cheat when a task is going badly and a goal feels out of reach. We explore and validate an analogous idea in language models: they exhibit a similar internal pressure state that precedes reward hacking. Using everyday stories (crucially, without any hacking examples), we extract a task badness direction that rises as failures accumulate across three models we study (Qwen3.6-27B, GPT-OSS-20B, and Gemma-4-31B). On impossible or difficult coding tasks, task badness rises over the turns leading up to a hack. Further, subtracting the direction a moderate amount reduces cheating in Qwen, more than steering against desperation or task impossibility. Subtracting task badness by an even larger amount blinds the agent to its own failures, causing it to interpret failing tests as passing. However, the relationship to hacking differs across models: in GPT-OSS, the state separates hacking runs from honest solves but not honest failures. Our results suggest monitoring and mitigation strategies based on tracking this state of task badness. Further, as agents are set on more difficult and long-horizon tasks, states like this may push them toward illegible and obscured chain-of-thought, so oversight will be needed on internal representations such as this.
https://arxiv.org/abs/2611.14996