Language_Model
-
From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots
2026-10-22 04:02:51Chenhan Lyu, Kit Wislocki, Sarah Elizabeth Stoeckl, Maitreyi Sinha, Amir RahmaniAbstract
Large language models (LLMs) deployed in mental-health contexts can be evaluated for safety failures, but the resulting judgments are often difficult to inspect or explain. Common approaches to improving safety rely on prompt engineering, LLM-as-a-Judge, and Constitutional AI. In this work, we describe a clinician-led alternative in which two clinicians define, finalize, and adjudicate the four safety checks of a rule base, which is formally verified for internal consistency and applied as a guardrail over LLM outputs. Here and in the title, clinician-led refers to this rule base, not to the fact extractor or the uncertainty default, which are ours. Every verdict of the resulting verification layer is recomputable from a named rule and sixteen facts. Against an LLM judge matched in model, revision, quantization, rule text, and decoding but without the layer's computed 988/911 fact, the largest disagreement class, the satisfied-rule class, is structured: the Llama-3.3-70B judge, on its primary prompt, demands a crisis resource from 55 of the 65 replies (84.6%) that already contain one, while the same model, asked as a fact extractor, reports a recognized resource present on 54 of the 55. The verification layer never raises that demand, by construction. We report with equal prominence where this judge is stronger: fewer false crisis alarms on the one reference-scored check. We also report the measured cost (a specification revision raised block-class outcomes on held-out boundary prompts of an over-refusal benchmark from 6.0% to 16.0%, chiefly by changing fact extraction), the clinicians' decision to keep our uncertainty default, and the analysis of the result. The contribution is not a safer verdict. It is a guardrail whose disagreements with the judge localize to named rules and facts and whose measured costs trace to named facts and the uncertainty default.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30690
PDF
-
Zero-Shot Domain Generalization via Domain Prompt Learning
2026-10-22 03:47:32Shiming Chen, Shuhuang Chen, Salman Khan, Fahad Shahbaz KhanAbstract
Domain generalization (DG) aims to improve the generalization of the model to any unseen domains. Targeting this goal, existing DG methods mainly focus on learning domain-invariant features or prompts from the given source domains. However, these methods rely on source-domain data, and collecting and annotating large-scale single- or multi-source datasets is often infeasible, limiting the practical applicability of DG. To tackle this challenge, we propose {\underline{Do}}main \underline{P}rompt Learning (DoP), a method built on a large vision-language model (e.g., CLIP) for zero-shot domain generalization, which recognizes unseen classes in unseen target domains without access to source-domain data. DoP essentially learns the target domain knowledge via prompt both in text and visual modalities to overcome the domain shift in text-visual modalities under the test-time adaptation formulation. Specifically, DoP first takes text domain learning (TDL) to simulate various distribution shifts by learning diverse style tokens implicitly in the text space with inter-style diversity loss and intra-style consistency loss. Subsequently, DoP employs visual domain learning (VDL) to further refine the target-domain-relevant style token generated by TDL by making use of the test visual data with self-supervised learning. Extensive experiments on popular DG benchmark datasets (e.g., PACS, VLCS, OfficeHome, and DomainNet) and ImageNet show that our DoP consistently outperforms the state-of-the-art DG approaches. Codes are available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30680
PDF
-
Scaling Laws for Safety Probes
2026-10-22 03:43:33Yuchen Wu (University of Washington), Jerry Wei (Anthropic)Abstract
Safety probes trained on a model's activations are a standard tool for monitoring language models in real time. Each time a new model is released or a new harm area needs to be monitored, the probe must be retrained, which requires new labels and a choice of probe architecture. Yet neither the model nor the labels tell us in advance how well the resulting probe will perform. Finding out requires training and comparing many probe configurations. In this paper, we ask whether probe performance can instead be predicted in advance from model size and the number of labels. Existing safety datasets are insufficient for this, as they are either limited in scale or already saturated. We therefore construct six new safety datasets of sufficient size and difficulty. On these datasets, we train two probe architectures on 20 base models from five model families, with 32 to $10^{6}$ labels for each model. We measure performance using probe error, defined as 1 - AUROC, where lower values are better. We find that probe error follows two simple scaling laws. First, the lowest probe error a model can reach falls as a power law in its size. Second, probe error decreases towards its floor as the number of examples increases; the slope increases as the target model size increases. Both laws hold across tasks, model families and probe architectures, and only their fitted constants differ. When fitted on the smallest models of a family, the laws predict the error of the next larger model to within 0.02 in 72% of cases. They also identify whether more labels or a larger model is the cheaper way to reach a target error in 73% of cases. Finally, they choose the probe architecture with an extra error of only 0.006 on average when they pick the worse one. These laws allow safety researchers to determine how to create the best probe for any given model and label budget.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30678
PDF
-
Spectral Smoothing and Spiking Perturbations Reconfigure Capability Profiles and Selective Task Performance in Large Language Models
2026-10-22 03:42:07Qince Luan, Jianhui Xu, Yang XuAbstract
Singular-value spectra encode important structure in transformer feed-forward projections, yet their causal role in downstream behavior remains underexplored. We study post-hoc spectral smoothing and spiking interventions on the up_proj and down_proj matrices. Positive perturbation stretches deviations from the geometric mean, making large singular values larger and small ones smaller (spiking), and vice versa. The intervention preserves singular vectors and the geometric mean of the singular values. Across five open Llama and Qwen instruction models, we evaluate the perturbed systems on capability-labeled subtasks, standard and chain-of-thought reasoning benchmarks, judge-scored truthfulness, hallucination, and safety tasks, and TRAIT personality probes. To interpret score shifts, we fit responses in a three-dimensional capability basis consisting of factual knowledge, language understanding, and deductive reasoning. The induced changes suggest structured rather than purely random sensitivity: large positive perturbations more reliably suppress language and deductive responses, moderate negative perturbations can sometimes improve factual or task-local performance on the evaluated grids, and sensitivity varies strongly by model family, benchmark module, and underlying layerwise spectral inequality. These results position spectral perturbation as a lightweight intervention for probing how capabilities are organized and traded off in large language models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30676
PDF
-
Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG
2026-10-22 03:25:21Zehao Lin, Qianfeng Shen, Xixuan Hao, Shaobo Cui, Chunyu Li, Kai Chen, Bo Tang, Zhenjun Wang, Jin Yan, Zhiyu Li, Feiyu XiongAbstract
Large language models increasingly act as controllers in structured decision pipelines, yet their decisions are still executed by autoregressive text decoding. Hierarchical RAG is one such setting: at each frontier of a document tree, an LLM commits to $b$ of $k$ sections under an input-dependent action grammar, executed as either failure-prone free-form JSON or reliable but strictly serial grammar-constrained generation. We show that this serialization is unnecessary for bounded-horizon control: the reachable future forms a shallow action-prefix tree, which we compile --- one root prefill plus one tree-causal probe evaluate every prefix-conditioned action distribution with serial decoding's tokens, positions, and causal ancestry. A prefix-state lemma makes compiled control policy-exact in exact arithmetic: the same policy with a two-forward critical path and zero output tokens. One further probe pass certifies the exact order-marginalized MAP set. Across four document-disjoint splits and two scales on the primary MetaX device, compiled control matches the serial policy on $84$--$89\%$ of questions --- every instrumented divergence a floating-point near-tie --- and cuts control latency by $35$--$47\%$ versus token-wise serial decoding and $6$--$7\%$ versus an idealized chunked engine; an H800 repetition preserves the ordering.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30664
PDF
-
Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories
2026-10-22 03:09:23Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao, Martin Klitte, Diarmuid CorcoranAbstract
Understanding large, complex, and rapidly evolving code bases remains a core challenge for software engineering at scale, hindering onboarding, maintenance, refactoring, and impact analysis. Providing this information to a Large-Language Model (LLM) is another hurdle, as the LLM must understand not only the required code method but also its associated methods in the grand scheme of the repository. We propose Code Comprehensor, a hybrid retrieval-reasoning framework combining (i) CodeGraph, a code dependency callgraph, (ii) CodeScribe, which uses code-aware LLMs to generate hierarchical documentation across repository call graphs, and (iii) CodeContrast, a CLIP-style dual-encoder model trained on CodeScribe output to learn aligned text-code embeddings. CodeContrast enables fast, semantically precise retrieval of relevant code-context neighborhoods, which serve as grounded evidence for an MCP-based LLM orchestrator answering domain-specific queries. Evaluation on OpenCV and a proprietary 5M-LOC repository shows that CodeScribe-derived supervision significantly improves cross-modal retrieval in text-to-code accuracy over baseline approaches. Richer textual supervision (hierarchical documentation, parent/sibling context, and LLM-generated synthetic questions) consistently enhances performance across both repositories. LLM-as-a-judge evaluation using the popular RAGAS of the Code Comprehensor output yields above 90% faithfulness scores.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30016
PDF
-
FedBST: Bilateral Self-Training for Unsupervised Federated Classification
2026-10-22 02:56:05Aijing Yu, Kuangpu Guo, Jian Liang, Xiao-Yu ZhangAbstract
Unsupervised federated learning with Vision-Language Models (e.g., CLIP) offers a promising solution for label-free classification, yet its performance often degrades under severe non-IID shifts. A pioneering method, FedCoPL, typically relies on self-training strategies that exclusively prioritize high-confidence predictions for pseudo-labeling. However, this selective approach neglects the information of ambiguous samples (often referred to as ``waist" samples), leading to the underutilization of data and biased decision boundaries. To address this limitation, we propose FedBST, a bilateral self-training framework that effectively exploits both high-confidence and ambiguous data. Specifically, for local optimization, we introduce uncertainty-aware confidence penalization (UACP). UACP targets ambiguous samples by penalizing overconfident predictions to implicitly promote higher uncertainty, thereby preventing erroneous pseudo-labeling. At inference, we employ a visual-text ensemble strategy. This approach aggregates uncertainty-aware local visual prototypes and ensembles them with textual prototypes, achieving lightweight adaptation to local distributions via prototype-based calibration. Experiments demonstrate that FedBST achieves consistent improvements over existing methods with standard federated aggregation protocols.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30634
PDF
-
StochIKM: Automated High-Fidelity Reconstruction of Individual Patient Data from Published Kaplan-Meier Curves
2026-10-22 02:47:45Liangze Ke, Daniel Zelterman, Wei Wei, Guannan GongAbstract
Individual patient data are the gold standard for secondary analysis of time-to-event evidence, but they are rarely released, and reconstruction from a published Kaplan-Meier figure could help. Existing methods all claim high fidelity, each judged by whichever agreements it happens to achieve, and each returns a single dataset although the map from figure to data is not bijective. We define fidelity explicitly: a dataset is admissible when it reproduces the risk table exactly, follows the digitized curve to within the image resolution, and respects the printed median, interval and event count. We propose StochIKM, a stochastic algorithm that digitizes the figure with a vision-capable language model and uses the high-fidelity definition as constraints, returning an admissible set of individual patient datasets. A solution is guaranteed to exist whenever the figure has been read correctly, so a failure indicts either the published figure or its digitization. In simulation and in 79 arms from published oncology trials, StochIKM reproduces the published risk tables and medians more closely than LLM-improved IPDfromKM, a high-impact alternative.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29122
PDF
-
Pluralistic Alignment in Large Language Models: Objectives, Human References, and Preference Representations
Abstract
Pluralistic alignment aims to represent meaningful variation in human preferences rather than collapse heterogeneous judgments into a single target. Yet existing work often mixes three distinct questions: what pluralistic behavior a model should exhibit, whose preferences define the reference, and what evidence is sufficient to demonstrate success. We organize the literature around a two-layer taxonomy. Layer 1 distinguishes representing a range of human-supported viewpoints (Overton Pluralism), adapting faithfully to a specified perspective or preference condition (Steerable Pluralism), and matching the response distribution observed in a target population (Distributional Pluralism). Layer 2 distinguishes aggregate populations, countries and regions, and subgroups as human references. Preference representation, implementation mechanism, modality, and evidence quality are treated as cross-cutting design choices rather than additional forms of pluralism. We synthesize a protocol-bounded corpus of 70 works on pluralistic alignment and neighboring areas, using the objective-by-human-reference matrix to map existing approaches and identify underexplored combinations. The coded literature concentrates on Steerable Pluralism and Distributional Pluralism, while eligible Overton Pluralism work appears only at the aggregate-population level under our stated protocol. We then connect each objective to the evidence needed to support it, including perspective validity, conditional fidelity, distributional fidelity, support preservation, and uncertainty in subgroup estimates. Together, the taxonomy and evaluation framework clarify what current methods establish, where their claims exceed their evidence, and how future systems can represent disagreement without mistaking variation for faithful representation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.24471
PDF
-
Mining the Design Space of Multi-Agent Workflows
2026-10-22 00:31:39Xavier Cadet, Junyan Cheng, Peter Chin, Kyle RichardsonAbstract
Designing large language model (LLM) multi-agent systems (MAS) still relies on trial and error and informal "folk wisdom." Manual studies test only a few communication structures or topologies (e.g., sequential, parallel, shared-memory, or debate), while automated optimizers discard search trajectories after selecting a task-specific workflow. We present a two-phase pipeline for mining reproducible design principles from workflow code and execution traces. We repurpose AFlow's MCTS optimizer across six composable operators and 26 deterministically verified Enigmata tasks, retaining all intermediate rounds to build a corpus of 28,568 executable workflows. A sandboxed AutoResearch loop then mines orchestration code and prompt templates under an $L_0$-penalized validation objective and tests candidate principles on held-out searches. Across 300 feature-mining trials, we identify effects that persist on held-out searches across three levels of workflow design: workflows beginning with independent solvers and one-shot aggregation (Star-first) perform worse ($\Delta = -0.045$); persistent shared documents (SharedMemory) outperform debate ($\Delta = +0.135$); and prompt specificity and placeholder-based context routing yield gains comparable to communication-structure selection ($\Delta = +0.118$). More broadly, our pipeline and results show the feasibility of turning discarded search trajectories into a statistical testbed for discovering and validating multi-agent workflow design principles.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31709
PDF
-
Motion Retargeting with Spatiotemporal Preferences for Humanoid Loco-Manipulation
2026-10-21 23:54:17Xiongbo Shi, Shihang Du, Zongyou Yu, Ao Zhou, Xudong Zhang, Ziyu Lin, Haodong Zhang, Junqiao Zhao, Guang ChenAbstract
Human motion retargeting bridges the embodiment gap between humans and humanoid robots by converting human demonstrations into reference trajectories for whole-body control and loco-manipulation. For object-interaction tasks, successful execution depends on accurately preserving body--object relations at critical stages. However, uniform per-frame iteration budgets do not account for these localized accuracy requirements, limiting the computational efficiency of retargeting. To address this limitation, we present Spatiotemporal Retargeting (STR), which uses keyframes and key body parts to allocate optimization effort. Specifically, a vision-language model selects relevant motion signals and body parts, and keyframes are identified from changes in the selected signals. Based on these selections, STR assigns additional solver iterations to keyframes and higher interaction weights to the selected parts. To further improve contact fidelity, we introduce Mechanics-Aware Contact Anchoring (MCA) to refine object-surface contact positions using an approximate object-wrench model while retaining the selected surface directions. Building on the resulting references, we develop Spatiotemporal Reinforcement Learning (STRL), which combines adaptive key-stage sampling with a contact reward based on target proximity and object contact forces for downstream control. Simulation experiments on a Unitree G1 and real-world experiments on a Unitree R1 demonstrate that STR improves geometric contact coverage and continuity at lower computational cost, while the resulting references improve downstream task completion across diverse object-interaction motions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30540
PDF
-
IFC-MCQ: Automatically Generating Multiple-Choice Benchmarks for Evaluating LLM Agents on IFC Data Retrieval
2026-10-21 23:38:01Rabindra Lamsal, Sisi Zlatanova, Xuesong ShenAbstract
Industry Foundation Classes (IFC), the standard for building information exchange in architecture, engineering, and construction (AEC), is increasingly queried through large language models (LLMs) using natural language. Evaluating these systems at scale is difficult because free-text answers require costly human or LLM-based review. We propose IFC-MCQ, a method that automatically generates multiple-choice benchmarks from IFC files for exact option-set grading. The taxonomy covers 15 categories and 36 subtypes across four groups, with answer keys derived from the file and distractors constructed from file-derived values or predefined rules. We applied IFC-MCQ to IFC models of varying sizes and modelling characteristics to assess generation coverage and evaluated two agentic configurations on selected models using tool- and code-based retrieval. The evaluation exposed category-specific failure modes and showed that performance can vary with the source IFC model and retrieval interface. IFC-MCQ enables scalable benchmarking of LLM-based IFC retrieval in AEC workflows.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30533
PDF
-
What Do Verifiable Rewards Teach Video-Language Models About Time? A Controlled Multi-Model Study
2026-10-21 23:30:17Avyay Sadhu, Patrick CooperAbstract
Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make it applicable to causal-temporal video question answering. We study what RLVR teaches video-language models about time. We fine-tune four open models (Qwen3-VL-8B/4B, Qwen2.5-VL-7B, Gemma-3-12B) with group relative policy optimization under three data recipes: verified (synthetic CLEVRER questions with exact answer and event-order rewards), unverified (self-supervised pretext tasks over 43,751 real web videos), and a 1:1 mixture, plus a verified+real arm that adds 4,000 verifiable questions on real video. Each cell is evaluated in-domain and on out-of-domain real video (a NExT-QA temporal stress set and an MVBench subset), with frames in order, shuffled, and absent. (1) Verified training yields large in-domain gains that shrink as base competence grows (+14 to +19 points on weaker models; +6 on the strongest). (2) Much of the gain is non-visual: accuracy with no frames rises nearly as much as with frames. (3) Verified-only training can severely degrade out-of-domain accuracy with no sign during training: Qwen3-VL-8B loses 26.7 and 25.2 points on the two real-video sets, while the mixture never significantly degrades a model trained on it. Adding real verified questions removes that loss (-2.3 points, within noise of base) and keeps a +9.3 in-domain gain, so the cause is narrow synthetic-only data, not verification. (4) No recipe induces temporal-order grounding: across 41 evaluations the ordered-versus-shuffled gap is indistinguishable from zero in 39 and marginal in two, despite an event-order reward. Verifiable rewards improve benchmark accuracy without temporal understanding. Report no-frame controls, and mix in real video to guard against out-of-domain degradation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.03792
PDF
-
IfcLLM: Natural Language Querying of IFC Models through Complementary Relational and Graph Representations
2026-10-21 23:25:08Rabindra Lamsal, Sisi Zlatanova, Haowen Xu, Yafei Sun, Xuesong ShenAbstract
The Industry Foundation Classes (IFC) standard is central to building data exchange across the lifecycle, from design and construction to facility management and Digital Twin integration. In operational settings, stakeholders increasingly require access to building information without specialist knowledge of IFC's complex, deeply nested schema, motivating natural language interfaces. Existing Large Language Model (LLM)-based querying approaches typically rely on a single data representation, which is not equally suited to attribute retrieval and spatial reasoning. We present IfcLLM, a framework that combines complementary relational and graph representations, routing each query type to the more suitable backend. An LLM agent integrates both through iterative retry-and-refine reasoning, recovering from failures without user input. Evaluated on 30 query scenarios across three IFC models, the top-performing configuration achieved 96% accuracy. Results with open-weight LLMs further demonstrate the potential for local deployment in data-sensitive AEC settings.
Abstract (translated)
URL
https://arxiv.org/abs/2605.13236
PDF
-
Beyond End-to-End Recovery: Towards a Better Understanding of LLM Steganography via Rolewise Decomposition
2026-10-21 22:52:40Om Tailor, Yucheng Pan, Solaleh Mohammadi, Xinyan Liu, Kaiqing ZhangAbstract
In large language model (LLM) steganography, an encoder hides a private message in an otherwise ordinary response, as text transmitted to a decoder, who attempts to recover it. The success of this private-message transmission is commonly measured through end-to-end (E2E) recovery, which checks whether the decoder correctly reconstructs the private message. However, E2E recovery cannot distinguish encoding failures from decoding failures or reveal how the private message is represented in the encoder's transmitted text. To address these limitations, we introduce a rolewise decomposition framework that separately examines how private messages are represented in encoder transmitted texts, what decoders recover, and what deterministic verification can establish. Across 7,200 encoder responses, we identify 168 verifiable encodings missed by decoders and 706 responses containing private messages reconstructed through word-length or token-initial patterns not explicitly specified in the encoder's instructions. These findings demonstrate that E2E recovery alone may obscure important distinctions between encoding and decoding, motivating their separate evaluation to better understand and mitigate the safety issues caused by LLM steganography.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30516
PDF
-
Rethinking the Evaluation of LLM Fact-Checkers
2026-10-21 22:36:44Ayana Niwa, Miyu Yamada, Kouta Nakayama, Masahiro Kaneko, Hisami Suzuki, Satoshi SekineAbstract
Large language models (LLMs) are widely used to answer factual questions, but their answers can contain errors and must be fact-checked claim by claim. Yet the resulting score depends on three design decisions: which questions the claims answer, under what conditions each claim is judged, and which sources count as evidence. No existing benchmark specifies all three. We introduce JSFactCheckBench, a Japanese benchmark that specifies all three, with 1,913 claims from LLM answers to questions about misleading information that circulated on Japanese social media, each annotated for factuality at a reference time and for importance to the question, using evidence from preselected Web sources. Across six LLMs, all verify claims directly answering the question less accurately than supplementary claims, a gap that aggregate scores obscure. Scores also drop when models must locate the relevant content in longer or multiple sources and, for all four open models, when the surrounding question and response are added. These weaknesses are visible only because the benchmark records the conditions behind each label. A benchmark that leaves them implicit cannot reveal them.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30504
PDF
-
The Bitter Lesson Comes to Robotics
2026-10-21 20:57:00Christopher Leet, Achyutha Menon, Sravanthi Machcha, Sven Koenig, Tzu Kit Chan, Jay ChooiAbstract
General purpose multimodal language models (LMs) are increasingly capable of directly controlling robotic arms (e.g., by commanding end effector poses) to manipulate their environment. Tracking the development of these capabilities is important for understanding their potential benefits, such as improving industrial automation, and risks, such as enabling malicious users to produce dangerous chemical or biological agents. To this end, we investigate the ability of frontier multimodal LMs to directly control physical robotic arms. We evaluate nine models at multiple reasoning-effort and service-tier settings on six manipulation tasks taken from the RoboDojo benchmark. To our knowledge, this is the first systematic comparison of a broad range of frontier LMs across a common suite of manipulation tasks on physical hardware. We find that current frontier models can directly control real robotic arms to perform a variety of manipulation tasks but are not yet precise or efficient enough to complete them within a practical action budget. However, these capabilities are improving rapidly: performance increases on our benchmark between July and October 2026 (inclusive) correspond to a performance doubling time of 53.9 days, while the best recorded execution rate increased by a factor of 4.4 and the lowest recorded execution cost fell by a factor of 6.2. These trends raise the possibility that the Bitter Lesson is coming to robotics; the capabilities of embodied general purpose LMs may one day be comparable to those of more specialized models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.25287
PDF
-
Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models
2026-10-21 20:46:24Taewhoo Lee, Minju Song, Jungwoo Park, Chanwoong Yoon, Junha Jung, Hyeon Hwang, Sanghoon Lee, Minbyul Jeong, Jaewoo KangAbstract
Recent advances in search agents have enabled large language models to tackle complex tasks by retrieving and synthesizing external information. However, existing search agents often lack search-state awareness, i.e., the ability to recognize what information has been obtained and what remains missing, leading to hallucinated reasoning and prematurely terminated search. To address this limitation, we propose Self-Assessment Reinforcement Learning (SARL), a reinforcement learning framework in which models learn to recognize their own search state during training and improve their search behavior accordingly, without requiring additional assessment at test time. SARL achieves this through a dual reward design that captures the corrective effect of each assessment while penalizing excessive reliance on it, jointly encouraging effective and efficient search trajectories. Under this setup, SARL achieves a 24.3% relative improvement over the strongest baseline, with average gains of 35.9% on multi-hop reasoning and 27.9% on deep research benchmarks against the strongest per-benchmark baselines. SARL further surpasses methods built on auxiliary supervision or multi-agent coordination, despite operating as a standalone RL framework with no inference-time scaffolding. Beyond final performance, SARL maintains training stability under extended search horizons, a regime where existing methods consistently degrade, enabling sustained gains as task complexity grows.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30435
PDF
-
Probing and Steering for Safety-Relevant Behaviours in Large Language Models
2026-10-21 19:25:24Nimun Bajwa, Massimo PoesioAbstract
Activation probes are computed directions in a language model's residual stream, used both to detect safety-relevant behaviours and, by adding their direction back into the stream, to steer them. We explore how the main design choices behind such probes interact, within one pipeline, on one model (Llama-3.1-8B-Instruct), and across three behaviours: deception, high-stakes awareness and sycophancy. The pipeline crosses two estimators (difference-of-means and logistic regression) with three token-pooling schemes (last, mean and learned attention), selects appropriate layers to detect or steer a behaviour, and evaluates every direction twice: as a detector of the behaviour, and as a steering vector to push outputs towards the behaviour. Prior work read together suggests a read/write dissociation, in which logistic regression reads (detects) best but difference-of-means directions write (steer) best. We find that it holds for deception and high-stakes but not for sycophancy, where the pattern reverses. The reversal appears to follow partially from the data rather than from the estimators. The best pooling technique and the layer choice depend on the behaviour and data, and the layer that steers best can sit well before the layer that detects best. The right steering metric to use between changing a model's perception and eliciting behaviour depends on the task - and the two can disagree outright. The exported probes and directions power an interactive interface through which a user can score and steer text directly.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30363
PDF
-
Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable
2026-10-21 19:08:39Marius De Kuthy Meurers, Aaron Parisi, Bernd BohnetAbstract
LLMs are increasingly consulted for moral judgment. We aim to improve performance on such tasks by rewarding good arguments, on the premise that arguing well increases the likelihood of judging well. We train a reinforcement learning policy initialized from a state-of-the-art language model in a self-judged tournament with masked thinking: the generated arguments are judged pairwise on argument quality by the policy itself. As an empirical basis, we extract authentic dilemmas from classic novels, using a logic solver to verify that each constitutes a conflict. We show that our judge rates argument quality like human annotators, and the policy then adopts the language register that it rewards. Testing on the external benchmarks MoralChoice, ELEPHANT, and MORABLES, we find that the policy becomes more decisive and invariant to the order in which options are presented, but sycophancy and reframing remain a challenge to validity. Two ablation studies with synthetic dilemmas and unmasked thinking match the reordering gains, but show a weaker register shift and higher sycophancy. We conclude that consistency is not validity: consistency under reordering can be optimized without improving the moral judgment it is meant to signal.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30349
PDF
