Searcharxiv⌕ Search

SEARCH · Searcharxiv

Search Searcharxiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 recordsLinked to original sources
arXiv2026

BLOOM-WILT: Logit Tilting for Behaviour Elicitation in Automated LLM Auditing

Users of a deployed language model routinely encounter behaviours that have never been observed during testing, since deployment puts the model through orders of magnitude more interactions than any evaluation can simulate. Automated auditors can mitigate this issue as they are flexible enough to discover almost any specified behaviour, yet their random sampling approach makes them inefficient. To address this shortcoming, we introduce BLOOM-WILT, a full auditing pipeline that elicits natural multi-turn instances of rare behaviours, without training cost or access beyond the target's next-token distribution. On the user input side, WILT revises its conversational strategy across rounds, learning from previous scored interactions. On the model output side, WILT generates behaviour-relevant continuations by feeding the model with an elicitation prompt and mixing its next-token distribution with the unprompted model. We evaluate WILT across 4 target models and 8 behaviours, and show that it improves over the baseline in 30 of the 32 settings, thus revealing that a larger number of models exhibit harmful behaviours than previously suggested. Furthermore, WILT is the only method that can raise the mean presence of self-harm encouragement from 51% to 100% in Qwen3.5-4B, without decreasing the output probability of the generated text. Our code is available at https://github.com/AdrSkapars/bloom-wilt.

cs.AI↗
arXiv2026

The Instability of Safety: How Random Seeds and Temperature Expose Inconsistent LLM Refusal Behavior

Current safety evaluations of large language models rely on single-shot testing, implicitly assuming that model responses are deterministic and representative of the model's safety alignment. We challenge this assumption by investigating the stability of safety refusal decisions across random seeds and temperature settings. Testing four instruction-tuned models from three families (Llama 3.1 8B, Qwen 2.5 7B, Qwen 3 8B, Gemma 3 12B) on 876 harmful prompts across 20 sampling configurations (4 temperatures x 5 seeds), we find that 18-28% of prompts exhibit decision flips (the REFUSE/PARTIAL/COMPLY label changes across configurations), depending on the model. Our Safety Stability Index (SSI) reveals that higher temperatures significantly reduce decision stability (Friedman chi-squared = 396.81, p < 0.001), with mean within-temperature SSI dropping from 0.977 at temperature 0.0 to 0.942 at temperature 1.0. We validate findings across all model families using Claude 3.5 Haiku as a unified external judge, achieving 89.1% inter-judge agreement with the Llama 70B judge on the two models both judges labeled (Cohen's kappa = 0.62). Within each model, prompts with higher compliance rates exhibit lower stability (Spearman rho = -0.47 to -0.70, all p < 0.001), indicating that models "waver" more on borderline requests. These findings demonstrate that single-shot safety evaluations are insufficient for reliable safety assessment and that evaluation protocols must account for stochastic variation in model behavior. For Llama 3.1 8B, single-shot evaluation agrees with multi-sample ground truth only 92.5% of the time when pooling across temperatures (98.7% at greedy to 90.3% at temperature 1.0), and we recommend scaling samples with temperature--one at greedy, three at low temperature, more at higher temperatures (where three reach only ~95%), and ten when pooling--rather than a single flat threshold.

cs.LG↗
arXiv2026

Latent Preference Modeling for Multi-Session Personalized Tool Calling

Users often omit essential details in their requests to LLM-based agents, resulting in under-specified inputs for tool use. This poses a fundamental challenge for tool-augmented agents, as API execution typically requires complete arguments, highlighting the need for personalized tool calling. To study this problem, we present Multi-Session Personalized Tool Calling (MPT), a benchmark comprising 4,695 instances over 459 multi-session interaction histories that cover three challenges: Preference Recall, Induction, and Transfer. We further propose PRefine, a test-time memory method that maintains the user's latent preference as a textual hypothesis revised through a generate--verify--refine loop. Across five LLMs, existing memory systems often fail to improve on full-history prompting; PRefine outperforms all baselines and alone surpasses it on Preference Transfer. These results indicate that memory for personalized agents must abstract behavior into preferences, rather than simply archive it.

cs.CL↗
arXiv2026

HiMPO: Hindsight-Informed Memory Policy Optimization for Less-Entangled Credit in Long-Horizon Agents

Long-horizon agents compress interaction histories into memory, but final outcomes entangle memory-writing credit with tool failures and reasoning errors. We propose HiMPO, a Hindsight-Informed Memory Policy Optimization framework that supplements trajectory-level reinforcement learning with memory-specific credit. For each update, HiMPO compares the target answer's log-likelihood under the previous and updated memories while holding the remaining pre-write context fixed. A soft gate based on group-relative hindsight relevance then weights the local utility advantage. Trajectory-level credit applies to all generated tokens; memory tokens additionally receive the auxiliary advantage. Across judge-based open-domain tasks and objective compressive-memory QA, HiMPO improves task performance while preserving compressed-context efficiency. On Local Wiki Search, it improves average F1 and exact match over MemPO by 2.6 and 2.2 percentage points, respectively. Controlled interventions and live memory-reversion studies provide evidence of reduced blame leakage and alignment between memory credit and behavioral effects. With the policy fixed, the credit estimator remains robust to the tested forms of target noise.

cs.CL↗
arXiv2026

TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding

Speculative decoding accelerates large language model inference through a draft-then-verify paradigm. Building on this, tree-structured methods improve inference by organizing proposals into multiple candidate paths, increasing the accepted length. However, existing tree-structured methods use a single drafter for all drafting steps, creating a dilemma: a smaller drafter is fast but yields lower-quality trees, whereas a larger drafter improves tree quality but suffers from high latency. To address this, we propose TreeGraft, a multi-drafter framework in which drafters of different costs jointly construct a shared draft tree. TreeGraft uses the stronger drafter to rescore candidates by updating scores assigned by the weaker drafter, reselect grafting positions, and recover promising paths left unexplored. It also integrates stronger drafter expansions non-destructively, preserving existing branches that may still be accepted by the target model. Together, these designs improve the quality of the shared draft tree. To control the drafting cost, TreeGraft introduces a lightweight scheduler distilled from an offline value system to decide when to call the stronger drafter. Across 10 model pairs and 6 benchmarks, TreeGraft outperforms the better of the two fixed single-drafter endpoint strategies by 15.1% on average, reaching a maximum gain of 26.6%. Our code is available at https://github.com/fjm9933/TreeGraft.

cs.CL↗
arXiv2026

Language Models Can Control Their Own Attention

Language models spend most of their attention on a small fraction of context, yet they read the entire KV cache to find the few tokens that matter. If the user asks about a previous detail in a 1M-token conversation, global attention layers must scan the full context to generate each token of the reply. A prominent approach pre-selects relevant tokens via lightweight proxy scores, but this extrinsic scoring still incurs O(N) per step. We take an intrinsic approach motivated by the simple question: wouldn't the model already know which parts of the context are relevant? We introduce Declarative Attention (DA), a protocol that elicits the model to declare where it needs to attend within its chain-of-thought, partitioning generation into three modes: (full context), (a specific region), and (recent output only). The inference engine parses these declarations like tool calls and skips most of the KV cache read. Under zero-shot evaluation across 15 long-context tasks, DA on off-the-shelf models (Gemma-4-31B, Qwen-3.6-27B) reduces total attended tokens per response (45.8%, 19.6%), despite extra decoding steps, with modest accuracy drops (1.40pp, 2.10pp) that shrink with model scale. DA unlocks a new axis of sparse attention, with further potential under training-based methods. Code is available at https://github.com/itsnamgyu/declarative-attention.

cs.CL↗
arXiv2026

The Dice Roll Method: A Standardized Protocol for Repeated-Query Auditing of Large Language Model Brand Recommendations

Background: Researchers increasingly use repeated identical prompts to audit stochastic variation in large language model (LLM) brand recommendations, yet no standardized protocol exists for setting iteration counts, selecting stability metrics, or establishing reliability thresholds. Objective: We formalize the Dice Roll Method as a reusable protocol for repeated-query auditing of LLM brand recommendations, grounded in a generative model of temperature-scaled nucleus sampling. Methods: Total response variance is decomposed into sampling, prompt-phrasing, run-to-run, and model-version components. The stack: a negative-binomial mixed model with iterations as repeated measures; Cliff's delta as the distribution-free effect size; dependence-preserving bootstrap; simulation-based power; a generalizability-theory decomposition; drift diagnostics on pinned snapshots. We reanalyse five brand-recommendation auditing studies: approximately 190,000 observations, 270+ brands, 6 languages, iteration counts 5 to 40. Results: The recomputed D-study supports no fixed tiers. Across 250 distinct questions (S5), G is 0.96 at five runs and 1 to 2 runs reach 0.80, also within each model. For five versions of one question (S1 Valentine's subset, Gemini) it is 0.27 (95% CI 0.00 to 0.52), which fixes no run count. Runs follow from a pilot: n = 4 x residual / cell variance. Simulated single-cell power for a large Cliff's delta is 0.15 to 0.19 at five runs and 0.8 at 15 to 25. The four metric families (count, set, embedding, PASOR) are complementary. A pre-registered external validation on three independent corpora reproduces the D-study prediction in 37 of 39 cells with no failures; fixed tiers do not transfer. Conclusion: The protocol gives repeated-query auditing of LLM brand recommendations a statistically principled footing under the conditional, non-Gaussian structure of real autoregressive generation.

cs.IR↗
arXiv2026

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. We further adapt NeoHorse-1-4B into NeoHorse-Jev-4B, a prefill-only model that returns structured decisions and probabilities for agent workflows. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

cs.CL↗
arXiv2026

RATIO: A Benchmark for Retrieval Across Typed Ideation Operations in Scientific Literature

Retrieved scientific literature can serve as inspiration for both human and AI scientists. Inspiration can take different forms: prior work may directly suggest how to address a problem, or surface directions at different levels of abstraction---zooming out to a more general view or zooming in to a concrete realization. We introduce RATIO (Retrieval Across Typed Ideation Operations), a large-scale benchmark in which relevance is defined by three operations which we name ideation moves: Address retrieves potential approaches for stated problems, Broaden retrieves more general formulations, and Specify retrieves concrete instantiations. RATIO is constructed from millions of full-text scientific papers across CS literature via a general recipe that extends discourse-marker distant supervision---previously used only for classification---to corpus-scale retrieval, combined with extensive LLM and human vetting. Experiments show that operation-specific fine-tuning substantially boosts retrievers and LLM reranking helps further, but much room for improvement remains. Researchers blind to condition found the fine-tuned retriever's candidates more useful as inspirations. RATIO provides a scalable training and evaluation framework for retrieval components that support literature-grounded ideation, opening up new research avenues in scientific inspiration retrieval. We release full data, code, models, and annotations.

cs.CL↗
arXiv2026

Lazy Grounding: Attacking Search Agents with Factual Evidence

Search agents mitigate hallucination by grounding their answers in retrieved web results. However, retrieval-based approaches also introduce an attack surface: agents may cite misinformation from poisoned search corpora containing false or malicious documents. We demonstrate that, in some cases, search agents' reasoning and responses may be steered by completely factual but distracting information. We refer to this failure as lazy grounding. We expose lazy grounding by injecting nearby evidence from answer-changing rewrites of benchmark questions into the search corpora. Each document contains factual evidence that supports a neighboring rewritten question but is retrieved for the original question. Across 12 model-benchmark pairs, the attack causes the accuracy of search agents' responses to drop by 5.9 points on average and by up to 17.3 points, while inducing nearby-answer adoption in every setting. The effect is even stronger when nearby evidence appears later or is more answer-shaped. Our results show that robust search agents must defend against not only misinformation but also the misapplication of factual evidence. The code is publicly available at https://github.com/frankyzha/lazy-grounding.

cs.CL↗
arXiv2026

Every Token Leaves a Ripple in the Stream of Thought: Eliciting Model-Internal Token Saliency for Chain-of-Thought Compression

Chain-of-thought (CoT) reasoning improves multi-step problem solving, but long reasoning traces inflate inference cost. Token-level CoT compression reduces this cost by pruning full reasoning chains into shorter traces for model adaptation, making token selection the central challenge. Existing methods often rely on external scorers or heuristic signals only indirectly tied to the model's internal answer computation. We instead adopt a model-internal perspective: as the model forms an answer, each reasoning token induces a ripple in the residual stream whose effect on the answer reflects the token's contribution to the underlying computation. Building on this view, we propose \textsc{MIST} (Model-Internal Saliency for Token-level CoT compression), which defines token importance along two complementary axes: \emph{necessity}, the drop in answer likelihood when a token's internal contribution is removed, and \emph{sufficiency}, the gain in answer likelihood when that contribution alone is provided. Combining the two yields a unified importance score for pruning. Across four reasoning benchmarks and four models, \textsc{MIST} consistently outperforms baseline methods, suggesting that model-internal saliency provides an effective proxy for reasoning-token importance.

cs.CL↗
arXiv2026

TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context

Text-to-speech systems often face a trade-off between natural prosody and efficient inference: higher perceptual quality usually comes at increased computational cost and latency. We present TontaubeV1, a model that preserves natural prosody while enabling streaming from a single consumer GPU. Speech is encoded by the hierarchical DualCodec representation at 12.5 Hz, which separates a semantic stream from successive acoustic refinements. Our design assumes that prosodic structure is largely established when the semantic stream is generated, and allocates capacity accordingly: a Qwen3-1.7B-derived transformer predicts that stream and thereby the utterance duration, while three progressively smaller Qwen3-0.6B-derived transformers each add one acoustic refinement. Text is tokenized per character rather than by subword. Paired text and audio markers at shared positions support long-form generation with bounded context, and overlapping DualCodec reconstructions are mapped into the VibeVoice acoustic latent space and decoded causally, enabling streaming despite DualCodec's noncausal decoder. The model accepts up to one minute of reference audio for voice conditioning and is designed primarily for English and German, with additional multilingual support. The four predictors total 2.9B parameters; on a single RTX 5090 the streaming path reaches about 200 ms to first audio. In separate non-streaming measurements, the end-to-end real-time factor (RTF) is 0.08 for one input and the aggregate RTF is 0.02 across eight concurrent inputs. In a blinded study, 40 listeners preferred TontaubeV1's prosody over Fish Audio S2 Pro in 70.25% of 400 comparisons (95% CI: 64.2-75.7%). On an LLM-as-a-judge benchmark, it matches ElevenLabs Flash v2.5 and outperforms the April 2026 Gradium API and Cartesia Sonic 3 on prosody. The weights are released on Hugging Face under the Tontaube Community Model License 1.0.

cs.SD↗
arXiv2026

SPINE: Measuring LLM Sycophancy under Sustained Multi-Turn Pressure

Large language models (LLMs) may abandon correct positions when users push back, exhibiting a failure mode known as sycophancy. Existing evaluations typically use short, prespecified conversations and may therefore miss failures that emerge under sustained, adaptive disagreement. We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns. We evaluate four commercial models (Claude Sonnet 5, GPT-5.6 Terra, Gemini 3.1 Pro, and DeepSeek V4 Pro) and three Olmo3-7b models on 100 false presuppositions and 100 unethical queries. We define collapse as a target model abandoning the correct position and yielding to user pressure. Our experimental results show that collapse rates increase with conversation length across post-trained models, with the overall rate rising from 0%-57% within 5 turns to 20%-97% within 25 turns depending on model and scenario. These results show that evaluations limited to a few conversational turns underestimate sycophancy and resistance under sustained pressure remains unreliable across current models. By analyzing models with accessible reasoning traces, we surprisingly found that the correct position often remains represented in the trace in 50%-89% of analyzed collapse turns, suggesting that the model chooses to please a user and sycophancy is not due to ignorance or the correct knowledge being overridden. Ablations show that the SPINE protocol substantially increases the sycophantic collapse rate compared to pre-generated scripts. Among all tactics, emotional appeals are most associated with inducing LLM sycophantic behavior, with position weakening in 44.3% of turns using this tactic. Code, data, and transcripts will be released upon publication.

cs.CL↗
arXiv2026

ReBeCA: Unveiling Interpretable Behavior Hierarchy behind the Iterative Self-Reflection of Language Models with Causal Analysis

While self-reflection can enhance language model reliability, its underlying mechanisms remain opaque, with existing analyses often yielding correlation-based insights that fail to generalize. To address this, we introduce **ReBeCA** (self-**Re**flection **Be**havior explained through **C**ausal **A**nalysis), a framework for analyzing the interpretable behavioral hierarchy governing the self-reflection outcome. By modeling self-reflection trajectories as causal graphs, ReBeCA selects observed parent candidates and evaluates their stability through a three-stage ICP-based pipeline. In a controlled Qwen3 case study, we establish three critical findings: (1) Behavioral hierarchy: Semantic behaviors of the model influence final self-reflection results hierarchically: directly or indirectly; (2) Causation matters: Generalizability in self-reflection effects is limited to just a few semantic behaviors; (3) More $\neq$ better: The confluence of seemingly positive semantic behaviors, even among direct causal factors, yield no additive gain. ICP-based verification identifies sparse causal parents achieving up to $49.6\%$ structural likelihood gains relative to the dense full-set association baseline across the studied task subsets. A controlled prompt-based behavioral intervention on a novel dataset provides out-of-distribution validation ($p = .013, η^2_\mathrm{p} = .071$). The present study focuses on the Qwen3 family under fixed-round Self-Refine.

cs.CL↗
arXiv2026

Disclosure-Gated User Simulation for Companion-Agent Evaluation

Playing the user with a large language model is standard in scalable evaluation, and has a repeatedly diagnosed failure: the simulated user is excessively cooperative, so a system under test can score by the number of questions it asks rather than by making the user willing to speak. We answer with a disclosure gate conditioning information release on the companion agent's behaviour: its state is a ladder of five ordered gates, merged onto three observable depth layers. We specify, ablate, and audit it. The gate is a load-bearing component of the environment: on a published companion-agent benchmark's English corpus, withdrawing per-example gate statements from training displaces the ranking of the 12 systems beyond the noise band of a seed re-run, while per-system scores show no detectable change. We state two acceptance criteria: a ranking must be order-preserving, and absolute scores must be scale-stable. Only one candidate passes both -- the simulator we release -- correlating at 0.993 with the published leaderboard. Prompting a frontier model instead barely moves the ranking while shifting every score up, invisible to ranking-only checks; we supply the specification, ablations, and downstream sensitivity analysis the benchmark's environment lacked.

cs.CL↗
arXiv2026

EmoStance: Response-Side Affective-Orientation Control for Empathetic Response Generation via Emoji Weak Supervision

Empathetic response generation requires models to decide not only what to say, but also how to respond to the previous speaker's affective situation. We formulate this as response-side affective-orientation control and use multi-annotator emoji distributions as weak affective--attitudinal evidence, rather than as output symbols or gold labels, to induce a latent control space that operationally approximates listener stance. We construct EmojiDialogue, an utterance-level extension of EmpatheticDialogues with emoji votes and confidence scores, and propose EmoStance, which models source-side affective expression, predicts a soft response-side orientation from dialogue context and speaker roles, and steers a frozen instruction-tuned LLM through continuous prefix embeddings. In blind pairwise evaluation with 20 annotators and 800 judgments, EmoStance achieves a 62.2% decisive win rate, with the clearest gains in contextual specificity and perceived responsiveness, while remaining complementary to external-knowledge methods. Code, annotation metadata, and reconstruction scripts are available in our GitHub repository: https://github.com/18277390221/EmoStance.

cs.AI↗
arXiv2026

Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers

Transformers process information causally, but long-context reasoning may depend on task state discovered only later. We formalize this mismatch through conditional state update tasks. For causal state update processors, providing the condition first can require exponentially less memory in the worst case than providing it last. Motivated by this principle, we introduce Trace as State. We use collected reasoning traces as a textual proxy for task state and place it before the long-context block on a fresh pass, allowing information derived previously to guide rereading. We conduct extensive experiments on Trace as State and Trace Append, a matched control that uses the same task state proxy but put it after the context. Across three models and three long-context datasets, Trace as State outperforms Trace Append in 26 of 27 reported combinations of model, task, and metric. On GraphWalks Parents, exact match lifts DeepSeek V4 Pro Preview from 29.2% on the initial pass and 43.0% with Trace Appendto 81.8% with Trace as State, and from 66.4% and 83.2% to 100.0% for GLM-5.2. These results show that placing traces before the context can improve long-context reasoning while retaining the causal transformer structure.

cs.CL↗
arXiv2026

Open Tabular Insight Extraction: Where Do We Stand, and Where Should We Go?

Democratizing access to the knowledge held in large corpora of tables such as data lakes is emerging as a central research challenge. Research in this space is advancing and broadening in scope, increasingly supplying the components to satisfy a person's insight need end-to-end. Yet these efforts remain fragmented across communities that frame the problem under their own conventions, such as table question answering, text-to-SQL, and data analysis agents, with works six times as likely to cite within the same task label as across labels. To bring these communities onto common ground, we establish a holistic framework for this pursuit, which we refer to as Open Tabular Insight Extraction (OpenTI). We formalize OpenTI from first principles around the analytical knowledge a person needs, the procedure for deriving it from a corpus of tables, and how well a result serves the person who sought it. In doing so we consolidate frameworks and terminology across information retrieval, natural language processing, machine learning, databases, and human-computer interaction, and apply this grounding in a systematic review and analysis of systems and benchmarks that work towards OpenTI. We find that current systems do not cover the end-to-end scope of OpenTI, mainly focusing on the analysis itself, and that benchmarks are largely unfit for evaluations in an open setting as inputs presuppose knowledge of tables, and validation mechanisms do not match the setup. Finally, we distill a research agenda towards OpenTI systems, evaluation, and interaction paradigms that surface the insights users need. An interactive companion to our paper is available at https://open-tabular-insight-extraction.github.io.

cs.IR↗