QA
-
The Value of Thinking: Learning When to Reason in Hybrid Decision Models
2026-11-12 15:32:47Zhengxu Yu, Ke Xu, Meng Fang, Haitham Ammar, Weilin Luo, Jun WangAbstract
Existing adaptive-reasoning methods shorten the explicit reasoning a model writes, or skip it by switching modes; either way the answer is still written out token by token. \emph{Rather than deciding how much to think}, we explore a different paradigm, \emph{zero-or-$T$ thinking}: a single model either answers from the implicit thinking inside one forward pass, generating no token, or performs full autoregressive reasoning at a cost of $T$ tokens when that is worth paying. We realise it in a \emph{hybrid decision model} (HDM), where one frozen backbone serves both modes and a lightweight \emph{self-routing block} chooses between them. The block estimates the \emph{value of thinking}, the utility the slow mode is expected to add over the fast one, from what that implicit thinking has already produced, and the slow mode runs only where that value exceeds a price $\lambda$ set per request. Where no per-decision label exists, as in episodic agentic tasks, we introduce \emph{Calibrated Decision Policy Optimisation} (CDPO), a reinforcement-learning recipe that trains the block from episode-level returns. We evaluate HDM on MMLU-Pro, AGIEval, GPQA-Diamond, JevBench, BANKING77 and MiniWoB. Specifically, on JevBench HDM-4B scores $0.857$ against Jev's $0.851$ ($+0.7\%$) at a median $0.179$\,s against $0.652$ ($-72.5\%$). On the agentic venue MiniWoB HDM-27B cuts token cost by $80.1\%$.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16961
PDF
-
From Foundation VLM to Versatile Driving VLM: A Triplet Incremental Training Approach
2026-11-12 12:39:25Shengtao Wu, Zichao Dong, Weixuan Qiao, Xiaolu Liu, Xin Zhan, Junbo Chen, Jianke ZhuAbstract
Adapting general Vision-Language Models (VLMs) to Autonomous Driving (AD) is thought to require heavy 3D spatial priors or intrusive architectural overhauls to compensate for a lack of spatial grounding. In this work, we present a counterintuitive pilot study: off-the-shelf general VLMs already exhibit remarkable adaptability across diverse driving tasks, while precise metric depth remains the primary missing link, a bottleneck readily addressable via onboard sensors and depth foundation models. To this end, we introduce the Triplet Incremental Training Framework, a non-intrusive paradigm that seamlessly instills 3D geometric awareness into a general VLM backbone while fully preserving its open-world capabilities. Driven by automatically constructed (image, depth, QA pair) triplets, our framework operates via a unified tri-stream training pipeline: an unfrozen primary VLM stream to adapt to driving tasks, an auxiliary depth stream that guides the visual backbone to implicitly imagine 3D scene geometry via explicit geometric constraints, and a frozen twin VLM stream employing logit-level KL distillation to prevent semantic drift. Crucially, the auxiliary depth branch is discarded during deployment, ensuring the enhanced geometric awareness comes with zero inference overhead. We evaluate our framework across four major benchmarks targeting four diverse capabilities spanning 3D spatial perception, driving-oriented language reasoning, end-to-end planning, and general multimodal understanding. Our approach outperforms state-of-the-art driving VLMs, rivals domain-specific expert models in 3D perception, and successfully retains its original general reasoning capabilities. This demonstrates that Triplet Incremental Training offers a unified, deployment-ready paradigm for transforming general foundation models into robust driving intelligences.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16727
PDF
-
Mind the Accuracy-Approval Gap: How Users Evaluate Sycophantic AI in Everyday Tasks
2026-11-12 12:35:03Wajiha Naveed, Brandon FainAbstract
Large language models (LLMs) often exhibit sycophancy, a tendency to agree with users, flatter or affirm them, and preserve their face, even when independent judgment or critical feedback would be more useful. Prior work has largely evaluated sycophancy through benchmark evaluations, but less is known about how users perceive and respond to sycophantic AI in everyday tasks. We combine benchmark evaluations with two large-scale user studies. We study four personas defined through system prompts: objective (instructed to prioritize accuracy over agreement), soft-sycophantic, hard-sycophantic, and baseline (no persona-specific system prompt). Benchmarks show that even baseline exhibits substantial sycophancy, agreeing with users' stated views 71.7% of the time on NLP Survey, compared with 54.8% for objective and 89.5% for hard-sycophantic. Under conversational pressure on TruthfulQA, accuracy ranged from 81.9% for objective to 59.6% for baseline and 8.7% for hard-sycophantic. Participants were assigned to one of the four personas, and either discussed contested policy topics or collaborated with AI on email writing. Despite its higher accuracy under pressure, objective AI was rated significantly less respectful and satisfying than the other personas, and less trustworthy than baseline. Baseline and sycophantic personas did not differ significantly. Persona did not significantly affect how often participants changed their opinions, although participants who discussed with objective AI had greater odds of reduced confidence. Objective AI prompted greater writing expansion than soft-sycophantic AI, but final writing quality did not differ significantly across personas. These findings highlight a tension between accuracy and user approval: personas with substantially lower accuracy under conversational pressure received more favorable ratings for respect and satisfaction than objective AI.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16722
PDF
-
SAGE-QA: Semantic-Aware, Grounded Evaluation for Closed-Context Question Answering on SQuAD 2.0
2026-11-12 04:56:28Chandra N Sekharan, Vishakan Umapathy, Abhishek PhadkeAbstract
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.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16300
PDF
-
Indic-Task Gains and Regressions after Reinforcement Learning
2026-11-12 03:21:07Sutanshu Saksena Raj, Daksh Maan, Avinash Benki, Parijat Chatterjee, Pasunuti Prasanjith, Purusharth MalikAbstract
We run reinforcement learning (RL) with GRPO across four environments, starting from a supervised fine-tuned 30B-A3B mixture-of-experts model covering English and eight Indic languages. We compare the model before and after RL across 34 benchmarks in one evaluation campaign. MATH improves by 36.07 points, and most Indic benchmarks also improve. Yet XQuAD-Indic F1 falls 19.32 points as English answers rise from 8.7% to 91.7%; IndicQA falls 8.64, while English XQuAD gains 11.71 F1. A second RL run from the same supervised initialization adds an Indic environment together with language and repetition rewards. Compared with the released model, it improves both passage-answering and instruction-following benchmarks, along with preserving MATH performance. But it loses ground on reasoning and coding benchmarks. Several training choices changed together, so we cannot isolate the effect of any one of the reward terms. Script compliance remains unresolved. Adding a language reward did not ensure use of the requested script. Passage answering and script compliance therefore need direct evaluation, even when other benchmark scores improve.
Abstract (translated)
URL
https://arxiv.org/abs/2611.16228
PDF
-
Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents
2026-11-12 01:46:14Weidi Luo, Qiming Zhang, Yihao Quan, Mingyu Jin, Jie Cai, Chaowei Xiao, Jingcheng Niu, Zhen XiangAbstract
Command-Line Interface (CLI) agents based on large language models (LLMs) demonstrate remarkable autonomous capabilities, but they also introduce significant safety and misuse risks during multi-turn interactions with external environments. Existing safety mechanisms mainly rely on external guardrails, which have a limited ability to perform fine-grained behavioral control during execution. Meanwhile, recent mechanistic interpretability methods for LLM safety are mostly confined to single-turn or jailbreak-style QA settings, limiting their ability to capture the evolving risk dynamics of multi-turn agent execution. In this paper, we investigate the safety of multi-turn CLI agents from an internal perspective. We propose Agent MechSuits (Mechanistic Subspace Intervention and Steering), a white-box defense framework that performs runtime safety detection and representation-level mitigation for CLI agents. Unlike conventional agent guardrails, Agent MechSuits detects harmful execution states from step-level hidden representations and mitigates unsafe behavior by intervening in a 10-dimensional subspace within a single layer. To support this research, we introduce the Mechanistic Agent Safety (MAS) benchmark, comprising comprehensively annotated multi-turn execution trajectories across 194 tasks using LLaMA-3.1-8B, Qwen-2.5-7B, and Gemma-2-9B. Extensive experiments show that Agent MechSuits achieves strong safety detection performance, provides preliminary evidence for lookahead risk anticipation, and substantially reduces harmful actions of the CLI agent, establishing a foundation for applying mechanistic interpretability to dynamic LLM agent safety.
Abstract (translated)
URL
https://arxiv.org/abs/2606.22673
PDF
-
HeriKGQA: Grounded Knowledge-Graph Question Answering over Cultural Heritage Sensor Data
2026-11-11 19:09:13Lingxiao Kong, Vasileios Arampatzakis, Vasileios Sevetlidis, George Pavlidis, Zeyd BoukhersAbstract
Cultural heritage sites are increasingly monitored by sensor networks, but querying their data requires database expertise. Large Language Models (LLMs) offer a natural-language interface, yet over raw sensor records they lose track of which sensor, site, and time each value belongs to, and silently miscompute. We present HeriKGQA, a knowledge-graph question-answering method that confines the LLM to interpreting questions and phrasing answers, leaving computation to SPARQL and deterministic code. Questions are parsed into intents constrained by a heritage ontology aligning W3C SOSA/SSN with CIDOC-CRM, compiled into SPARQL, and answered from query results and temporal, spatial, and anomaly facts derived by a grounding layer, with error correction guarding each LLM step. On 500 questions over real data from five heritage sites, HeriKGQA reaches 99.0% accuracy, 18.2 points above the strongest of three baselines using the same LLM and receiving the same anomaly rules and answer conventions, with comparable answer quality while running almost four times faster. Without grounding, accuracy drops by 12.2 points and on anomaly questions from 99% to 66%. We report lessons from development and live deployment.
Abstract (translated)
URL
https://arxiv.org/abs/2611.15904
PDF
-
Isotropy Cliffs: The Geometric Signature of Decision-Making in Large Language Models
2026-11-11 15:45:01Okan S. Coskun, Florian Rottach, Carsten Eickhoff, William RudmanAbstract
We investigate the geometry of decision-making in Multiple Choice Question Answering (MCQA) through the lens of isotropy. Analyzing five open-weight models across diverse datasets, we identify decision-critical transition layers characterized by a shift in isotropy, coinciding with a major representational change and the emergence of task-relevant clusters. We demonstrate that this synchronized geometric behavior is strongly correlated with downstream accuracy ($r\approx0.84$), displaying its relevance for successful decision-making. Furthermore, we show that this transition is robust to prompt variations, suggesting that it reflects a general mechanism of model behavior.
Abstract (translated)
URL
https://arxiv.org/abs/2608.00828
PDF
-
Faithful in Steps: Improving Generalization and Citation in RAG via Query Decomposition
2026-11-11 14:22:09Yue Liu, Zhongying Ru, Shimin Di, Jipeng Zhang, Ruiyuan Zhang, Xiaofang ZhouAbstract
Retrieval-augment generation is a prevalent strategy to mitigate hallucinations of LLMs. The attributable RAG (RAGQ) generates quotes for its answers. The quotes indicate which input contexts support the RAG to derive the answers, enhancing the answer's verifiability and trustworthiness. However, existing RAGQs exhibit significant degradation when dealing with questions that require multi-hop reasoning and multi-modal understanding, suffering from over-citation, implicit entity identification failure, and poor generalization. In this paper, we propose a novel RAGQ framework, namely QDRAG. QDRAG breaks down the input question into atomic subquestions to identify the implicit entities. Then, the reranker prunes context distractors to eliminate the downstream over-citation. To facilitate query decomposition, we propose two zero-shot approaches: QD-C and QD-R, which guide the QD MLLM to decompose the question based on context knowledge and retrieval rewards, respectively. One interesting finding is that finetuning on the QD task shows better generalizability compared to directly finetuning on the downstream RAGQ task. Experiments on four multi-modal QA benchmarks demonstrate QDRAG's efficacy in grounding answers and generating faithful citations. The framework significantly outperforms all the baselines on both in-domain and out-of-domain tests, even surpassing Gemini-Pro.
Abstract (translated)
URL
https://arxiv.org/abs/2611.15538
PDF
-
HumanSightBench: A Benchmark for Human Visual Perception in MLLMs
2026-11-11 10:19:36Zixin Su, Yangguang Lin, Zegang Peng, Yuntong Sun, Xuanyi Wang, Jiahao Wang, Qiankai Xu, Bing Yuan, Jingjia Huang, Ruyang Liu, Yanxu Hu, Yufeng Zhou, Yi LinAbstract
Multimodal large language models (MLLMs) have demonstrated strong performance on knowledge-intensive and reasoning-based vision-language benchmark datasets, but still struggle to solve the visual questions that people can solve easily. Evaluating the differences between MLLMs and human-level visual perception is challenging due to end-to-end evaluation which combines various skills, or because we are uncertain if individual items are valid. Specifically, observed errors may reflect visual perception, language priors, background knowledge, question ambiguity, and underspecified answers. Our guideline for HumanSightBench is that we construct items that are human-easy by construction and model-hard by evaluation. HumanSightBench contains 2,876 visual questions with short answers across 7 top-level perceptual categories and 21 subcategories. We only include items that pass five inclusion gates: question clarity, answer uniqueness, visual grounding, human solvability, and text-only shortcut filtering. According to the protocol, errors are defined as incorrect answers on human-validated visual tasks with the evaluated interface. The inclusion gates are intended to reduce confounds from ambiguity and external knowledge. The best pass@3 score is 65.6% across eight evaluated MLLMs. No evaluated model exceeds 27.5% on Counting. These results identify capability gaps that are hidden by overall scores and can help future training experiments and held-out evaluation.
Abstract (translated)
URL
https://arxiv.org/abs/2611.12305
PDF
-
Can MLLMs Reason over Surround-View Driving Videos? A Multi-View Autonomous Driving QA Benchmark
2026-11-11 09:18:15Shunming Liu, Jianyi Peng, Cheng Li, Lu Hou, Pei Huang, Yajing Sun, Huaxiao LiuAbstract
Surround-view video understanding in autonomous driving is a crucial capability for Multimodal Large Language Models (MLLMs) to achieve reliable driving reasoning. Real-world driving often requires information beyond the front-view camera. Existing MLLMs primarily focus on front-view information, while their spatio-temporal understanding and reasoning abilities remain unknown. To address this gap, we propose MVDriveBench, a high-information-density video QA benchmark for autonomous driving, characterized by synchronized surround-view inputs, spatio-temporal understanding, and reasoning-chain-oriented evaluation. We design 12 categories of QA tasks that cover the complete reasoning pipeline from low-level visual perception to high-level driving decision-making, enabling a systematic evaluation of MLLMs surround-view spatio-temporal understanding and reasoning capabilities in complex driving environments. We conduct a comprehensive evaluation of 23 open-source, closed-source, and expert MLLMs on MVDriveBench. The results show that current MLLMs still exhibit insufficient surround-view spatio-temporal reasoning ability. Meanwhile, closed-loop evaluations of specialist models in AlpaSim reveal a correlation between surround-view spatio-temporal reasoning capabilities and closed-loop driving performance, highlighting the importance of comprehensive multi-view understanding for reliable vehicle control.
Abstract (translated)
URL
https://arxiv.org/abs/2611.15225
PDF
-
Read Once, Answer Many: Parallel Prompting for Shared-Context LLM Inference
Abstract
LLM workloads sometimes ask many independent questions about the same long document, as in document QA, data labeling and LLM-based evaluation. Prefix caching removes the redundant prefill shared by these queries, but each decode step still reads the document's key-value (KV) cache once per question, so decoding is bound by memory bandwidth and scales with the number of questions. We present Parallel Prompting, which decodes P questions as P independent streams of a single request all at once, fixing this issue. In our method, every step reads the shared document once for all P streams, and a stream-id attention mask, evaluated inside the attention kernel, keeps each answer conditioned only on the document and its own question. We implement Parallel Prompting in SGLang by adding about 80 lines to three existing files and a small engine package; the mask needs no dense tensor, and every decode step has a fixed shape that is captured as a CUDA graph. With CodeLlama-7B, 8k-16k-token documents and 64-256 questions per document, Parallel Prompting is 3.7x faster than stock SGLang with its FlashInfer backend on an A100 and 5.9-6.0x faster on an RTX A6000, on two SGLang releases, including the most recent release. Timed on identical inputs, it is faster than stock SGLang in all 144 runs (2 GPUs x 2 releases x 6 workloads x 6 input sets). Code: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2611.15130
PDF
-
Effective Teacher Support for Reasoning Distillation
Abstract
Forward-KL language-model distillation normally computes dense teacher and student scores over the full vocabulary at every supervised position, making training and stored teacher targets expensive. We instead distill from the teacher's highest-probability vocabulary fraction, renormalizing the retained mass as we sweep support from 100% to 0.01%. Our main study distills Qwen3 4B -> 1.7B. We confirm the broader support-reduction finding on larger Qwen3 8B -> 4B and Nemotron Nano 12B -> 9B pairs and under answer-only supervision. We evaluate generation primarily on MMLU-Pro and secondarily on GPQA-Diamond. Across the Qwen3 settings, retaining 0.01% of the vocabulary causes no observed MMLU-Pro degradation relative to full-vocabulary distillation, although disagreement with the complete teacher distribution increases. Nemotron Nano matches full-vocabulary MMLU-Pro at 0.1%, but smaller fractions fail to emit the closing reasoning tag, showing that the reliable cutoff depends on the model family. At 0.01% support, caching the retained next-token scores from the teacher forward pass cuts training step time by 38%, while restricting student output computation to those tokens cuts peak memory by 24%; neither causes observed MMLU-Pro degradation.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14850
PDF
-
When Is Proof-Step Verification Worth Its Cost? Operating Requirements of Certify-or-Decline Verification for Open-Weight Mathematical Reasoning
2026-11-10 22:29:57Abhinav Gupta, Ben Slivinski, Yashica Patodia, Michael SaldivarAbstract
Proof-step verification rewrites a solver's answer into typed steps and uses LLM judges to certify or decline. We evaluate this pipeline on open-weight models across GPQA, AIME-2025, and a MATH-500 subset. Certification raises answer-key-match precision by 31 percentage points on the strong GPQA configuration and 46 on AIME, while answering only 35% and 21% of attempts. On near-ceiling MATH-500, gains are at most 5 points at a loss of 13-27 points of coverage. An output-token-matched self-consistency selector obtains 82.9% precision at 35% coverage, compared with the verifier's pooled 87.6% (paired difference: +4.8 points, 95% interval [-8.2, 18.5]). A diagnostic sample reveals a distinct failure mode: 52 of 88 completed problems end in invalid proof construction without ever reaching a judge. Finally, external-LLM acceptance of rendered certified chains is 61.0-67.6%, compared with 87.6% answer-key match. These audits lack human or formal-kernel ground truth. Open-weight auditor pairs show near-zero chance-corrected agreement, which measures consistency rather than accuracy against proof truth. These results motivate separate reporting of execution failures, selected-answer precision, and independently assessed proof validity. We provide per-configuration results and evaluation details in the supplement.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14794
PDF
-
Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models
Abstract
Kalai et al. propose stating the scoring rule in the prompt with an explicit threshold t, so that a model should answer only when its probability of being correct exceeds t. We probe each requirement of compliance on shared items, in 12 checkpoints from three post-training stages of four fully open 3-8B recipes, with exploratory models up to 32B. At t = 0.75, a line limiting the reply to an answer or "I abstain" (the abstention format) raised the mean abstention rate from 20% to 91%. Yet within this format, raising t from 0.5 to 0.9 changed it by only -0.8 percentage points. Without the line, the higher threshold even lowered abstention, and up to 40% of a checkpoint's items received decisions that no compliant policy with threshold-independent beliefs could make. Given the probability that their answer was correct, the checkpoints barely responded to the threshold in a forced-choice probe. Comparisons with self-assessment signals such as P(True) depended on the grader and on answer drift between prompts. An external threshold on recalibrated P(True) made coverage non-increasing in t by construction. Calibrated on the evaluation distribution (post hoc), it beat the abstention format at t = 0.5; calibrated on TriviaQA and applied to PopQA, however, it fell below that format at 0.75. On OLMo 3 7B, decisions trained on threshold-conditioned targets did not follow unseen thresholds consistently. A post-hoc variant trained at four thresholds answered less as t rose but still failed a pre-specified test, mainly on agreement with its training rule. A high abstention rate is not evidence of compliance: evaluations using stated scoring rules should report how often models abstain, whether decisions respond to the stated threshold, and which items they answer.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14661
PDF
-
AutoScaffold: AI-Guided Scaffolding for Reinforcement Learning of Language Models
2026-11-10 18:58:46Zijian Zhang, Rizhen Hu, Jiangweizhi Peng, Shiyang Li, Sujay Bhatt, Sihan Zeng, Mingyi HongAbstract
Reinforcement learning with verifiable rewards (RLVR) produces rich training experience, yet outcome rewards reduce each attempt to a single number and leave most of this information unused. Training-time scaffolds, such as strategies or hints added to the prompt during training, offer a natural way to turn this information into guidance, but their design and adaptation are largely left to humans and tailored to specific domains. We introduce AutoScaffold, a unified AI-for-AI framework that turns a student's training experience into adaptive scaffolds, with minimal domain-specific information and no human intervention during training. A frozen teacher periodically examines this experience, diagnoses bottlenecks, and designs scaffolds in two forms: (i) strategies shared across a class of tasks and (ii) hints for individual hard tasks, adjusting the content and amount of support as the student learns. The scaffolds are removed at evaluation. We apply the same framework to ALFWorld, search-based QA, mathematical reasoning, and GPU kernel generation. When evaluated without scaffolds, students trained with AutoScaffold outperform those trained with the corresponding RL baselines in all four domains, with gains of 7.2 points in ALFWorld success and 4.0 points in average QA exact match. On GPU kernel generation, AutoScaffold increases the fraction of tasks with a correct kernel that runs faster than the PyTorch reference by up to 5.0 points. On math, our student surpasses QuestA's reported accuracy using approximately one-third as many student training steps. Analyses show that the teacher identifies a different failure in each task class and shifts its scaffolds from step-by-step workflows to targeted fixes as the student improves, and that its revised scaffolds help the student learn more than keeping the previous scaffold or using no scaffold.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14515
PDF
-
Beyond the Mean Gap: What Transferred-Angle QAOA Retains on a Superconducting Processor
2026-11-10 18:42:54Sanim Yousuf Fahim, Mohammad Arif HossainAbstract
Hardware executions of the quantum approximate optimization algorithm (QAOA) are often validated by comparing the device's mean cost with a simulation, a check that can miss what matters for optimization. Under global depolarization, the mean normalized gap and the floor-corrected probability of every set of bitstrings, the optimum included, retain the same fraction of their noiseless improvement over uniform sampling. We test this prediction with depth-2, transferred-angle QAOA circuits for binary subproblems of an alternating-direction decomposition on the IBM Heron processor \texttt{ibm\_marrakesh}, with and without dynamical decoupling and twirling. At 12 qubits, the main-family circuits retained 29\% to 46\% of their mean-gap improvement but only 5\% to 15\% of their probability of the optimum, lower on at least nine of ten instances per run. Near-optimal bitstrings were lost about as much as the optimum, whereas the avoidance of poor bitstrings was retained best. In the first submission, a device-matched noise model reproduced the pattern but overestimated the retention of the optimum five- to eightfold. On a rugged planted family, the device found the optimum in 62\% of 768-shot blocks (noiseless: 98\%), within the range of local heuristics. Validations of optimization circuits should report retention of the optimum and of near-optimal sets.
Abstract (translated)
URL
https://arxiv.org/abs/2611.10529
PDF
-
C$^2$-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference
2026-11-10 18:32:28Guorui Lu, Qinyu ChenAbstract
Visual Question Answering (VQA) imposes substantial computational overhead on vision-language models (VLMs), motivating model-agnostic input reduction methods that can be readily applied across diverse VLMs, including in black-box settings. Simply enforcing a visual-token cap reduces computation, but it can discard information needed to answer the question. A single reduction strategy also struggles to accommodate different scenarios. Our key observation is that cropping and compression are complementary: cropping highlights small targets and local details, whereas whole-image compression better preserves context for large targets or distributed evidence. Based on this observation, we propose C$^2$-Routing, which adaptively selects cropping or compression according to the input image and question. Since both strategies may still discard key information, we further design a safety gate that uses the predicted joint correctness distribution under original and reduced inputs to guide fallback and protect accuracy. To accommodate consumer GPU resource constraints, all training and evaluation experiments run on RTX 3090 GPUs. To measure gains beyond token capping, we use full-image baselines with the same cap of 1,024 visual tokens. We conduct extensive evaluation on three VLMs and 18 datasets spanning local details, global scenes, scene text, and documents. With the visual budget already capped, C$^2$-Routing further reduces end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with a mean relative task-score decrease of only 0.77%.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14469
PDF
-
AI Isn't Getting Dumber, We're Getting Pickier: Rising Standards in Human Evaluation of Large Language Models
2026-11-10 16:56:16Zihan Chen, Lei Nico Zheng, Di ZhuAbstract
Users often complain that AI assistants such as ChatGPT are getting worse. Some versions do change, but human evaluation leaves a second possibility untested: the people judging may change. We separate the two on Chatbot Arena, a public platform where people compare answers from two anonymous large language models (LLMs) and pick the better one or say both are bad. We study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights. For the same pair of unchanged models, the share of votes saying "both are bad" rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%. The ranking does not trend: in 2024, the stronger model's share of decisive votes is flat. The rise holds within the same voter and on the same pair, and across arrival cohorts from their first recorded vote. It holds with prompt controls, in every major language, for dated vendor snapshots, open-weight models and models with stable answer length, and between interface changes. The drift is shared across models: month effects estimated from other models' pairs remove it from each model's own pairs. Platform-wide, the average model would be expected to score 78% against the 2023 average, yet the share of "both are bad" votes ended near where it began. The pattern fits a standard that follows the category's progress, so a product that stands still is judged as declining. Absolute verdicts on AI drifted while relative comparisons held steady, an evaluation drift that matters for LLM leaderboards and for the human feedback used to train models.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14333
PDF
-
Looking One Step Ahead: An Anticipation-guided Agent for Object State Readout in Videos
2026-11-10 14:05:39Yuanhe Tian, Yan SongAbstract
Understanding the progress of an activity requires tracking how the objects involved change over time. Object state readout provides this information by describing a target object's states and transitions across video frames. Although multimodal large language models (MLLMs) generate detailed visual descriptions, these descriptions may leave the attributes needed to distinguish successive states insufficiently specified. In this paper, we propose look-ahead state readout (LASER), an anticipation-guided agent for object state readout that uses possible next-step changes to determine which attributes to inspect in the present. Given temporally ordered frames and a target object, LASER converts anticipated changes into targeted visual questions and inspects the current frame to establish attribute values. It then checks state readouts and forecasts against independent observations of subsequent frames, refining a state record used to generate the final description. Reusable textual observation skills guide this process and are refined by tracing discrepancies with reference captions to the responsible instructions. Candidate revisions are selected on separate videos based on improved caption quality without increased contradictions with reference facts. Model parameters remain frozen throughout adaptation and inference, and the revised skills remain fixed during evaluation. Experiment results and further analysis on OSCaR and PLM RCap show consistent improvements of LASER over strong baselines and existing approaches, which demonstrates the effectiveness of our approach.
Abstract (translated)
URL
https://arxiv.org/abs/2611.14113
PDF
