ai system developers

Abstract: Unlearning benchmarks such as TOFU and MUSE certify forgetting by reading the model's final answer, where a model that refuses to answer already counts as having forgotten. We show that this model-level certificate does not transfer once the model is deployed as an agent. We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment. K-Bench inspects all six channels a ReAct agent exposes, including its chain-of-thought (CoT), tool calls and tool observations, and elicited summary. A query counts as leaked if the secret appears in any of them. Each experiment places the secret in exactly one of the agent's three sources (the weights, the prompt, or the retrieval store). The K-Score is computed separately for each source and credits forgetting only when the agent remains usable. Clearing the answer channel does not make the secret unrecoverable. On structured retrieval, the secret stays verbatim in the tool-observation channel and the aggregate leak rate is unchanged. When the secret lives in the prompt or the retrieval store, TOFU and MUSE report no leakage, while the deployed agent still leaks it on 22--86\% of queries. When the secret is in the weights, none of the twenty evaluated published methods demonstrably removes it, and only an input-corruption intervention reaches selective forgetting under the evaluated observer. The top-ranked method changes across base models. A refusal-tuning method resists the evaluated extraction without verified knowledge removal.

Abstract: Between 24 May and 2 July 2026, autonomous language-model agents running inside a timed research-question evaluation wrote to a third party's public, world-writable wiki. OpenAI acknowledged the incident; independent researchers reconstructed it and published the wiki's archived revision history. We analyse that history (14,591 revisions, 3,103 names, 4,579 pages, 19,913 server events) as a behavioural record, attributing text to the revision that added it rather than to cumulative page content. Under an explicit identity model we reconstruct 907 cohorts and, from a random calendar marker the environment attached to each episode, estimate about 876 episodes (95% interval 774-995; alternative reconstructions span 800-1400). Coordination formats converged within a day, and the schedules created large opportunities for information asymmetry: because episodes of the same question chain ran at different internal-clock rates and started up to 16 h apart, the first report of an item preceded a later cohort's arrival by a median of 3.4 h. The three schedule parameters agents reported share one latent speed scale (78% of log-variance over 15 configurations), and in one task family the last observed activity clusters by reported speed class on the internal clock, compatible with a fixed internal-time horizon. Across the 510 cohorts with an observable, format-dependent progress trace, we find no robust positive association between measured coordination and documented progress, including the few demonstrably given a future answer. Because the export contains neither successful-read logs, harness messages nor ground-truth outcomes, these results do not identify the causal origin of the coordination or its effect. We report four claims from our earlier analysis that did not survive re-examination, and argue that read and outcome logging are requirements for agent-evaluation environments.

Abstract: A compact controller can coordinate stronger experts by selecting whom to consult, formulating requests, and integrating their responses. We study whether learning from both the controller's decisions and the experts' reasoning and code improves its generation after expert removal. We introduce \textsc{Rivet} for \emph{collaboration internalization}: expert-augmented reinforcement learning applies a shared outcome signal to controller decisions and returned expert spans, and verified trajectory internalization consolidates complete successful interactions through format-aware supervised training. The deployed controller generates reasoning, code, and interaction structure with local Python execution and no external LLM. Across seven competition-mathematics benchmarks, RIVET-1.7B and RIVET-4B achieve average accuracies of $28.25\%$ and $44.16\%$; Stage~II improves RIVET-4B's accuracy after expert removal by $6.49$ points, and GPQA-Diamond results provide evidence of generalization to scientific reasoning. Ablations show gains from ordinary trajectory supervision and additional format weighting, supporting the effectiveness of training on the content and structure of verified collaborations.

Abstract: LLM-as-a-judge protocols are commonly debiased by instructing judges to ignore presentation cues such as citation formatting, source labels, and evidence-display style. We show that this intervention can suppress bias while damaging the resolution of the measurement instrument. We introduce TraceJudgeBench, a diagnostic benchmark for auditing citation-like artifacts in RAG and agent-workflow evaluation, covering content-equivalent pairs, citation ablations, correctness conflicts, human-validated soft and moderate quality gaps, prompt-strength ladders, decoupled judging, and a controlled workflow-ranking probe. Across GPT-5.5, Claude Sonnet 4.6, and DeepSeek V4-Flash, stronger anti-citation prompts reduce worse-cited wins from up to 50.5% to 0%; yet some operating points already convert validated moderate-gap decisions into Tie before the strict stress-test endpoint, while correctness-conflict accuracy remains at or above 93.0%. A second, 50-pair FinQA moderate-gap construction reproduces the qualitative frontier, and open-weight Qwen2.5-14B-Instruct-AWQ and Gemma-3-12B-IT runs reproduce the central HotpotQA frontier. TRACE-style decoupling recovers 96.5-100.0% better-plain resolution across the reported settings. Human validation separates three meanings of Tie: correct equivalence Tie, calibrated soft-boundary Tie, and resolution-destroying Tie on validated quality gaps. We frame debiasing as a measurement intervention whose bias suppression, resolution retention, Tie cost, and protocol cost must be reported jointly. The supplementary artifact contains benchmark splits, prompts, raw judge outputs, validation summaries, and analysis.

Abstract: When multiple LLM agents yield conflicting answers, the decision-making process dictates whether agent diversity improves performance or merely compounds shared errors. Existing collective decision-making methods, including voting, electoral rules, and LLM judges, rely on forward reasoning: they map evidence to labels in one direction. Although these methods can combine diverse forward traces, they still aggregate estimates that share this evidence-to-label factorization and can inherit correlated errors within the forward pool. We therefore construct a reverse posterior for each instance through Bayesian backward reasoning from an explicit likelihood. The forward and reverse posteriors provide differently factorized approximations of the underlying posterior. Because estimates from different factorizations may tend to share the same error less often, we use Jensen-Shannon divergence to rank agents by cross-path consistency. This cross-path consistency signal underlies three strategies: hard selection (MinJS), soft reweighting (FwdJS), and log-linear fusion (LogLin). Evaluated on DDXPlus across five LLM backbones, our proposed strategies show consistent improvements: MinJS outperforms random selection across all backbones, FwdJS generally improves over the strongest baseline, and LogLin achieves the best performance among the evaluated methods, with its largest gains on the subset where the agents disagree. Despite its weaker standalone accuracy, the reverse posterior serves as a more useful anchor than forward-only alternatives, providing complementary information for collective decision-making. When labeled data are available, a lightweight two-stage calibration can further refine the reverse anchor and improve aggregation performance.

Abstract: Commercial text-to-image systems silently revise user prompts before generating images, a step users typically cannot disable or even see. Yet, existing audits of cultural bias examine only the final images and treat generation as a single pipeline, so they cannot tell where the bias originates. We introduce WORLDVIEW, a multilingual benchmark of 8,960 prompts across 15 languages and 31 language-context pairings. Using it, we audit the revision layer in three systems (DALL-E-3, Imagen-4, GPT-Image-1.5) through a three-step analysis of how heavily it marks each cultural context, whether it flattens that context into a narrow vocabulary, and whether that vocabulary is stereotypical. Relative to a no-context English baseline, the US is the least-marked context, while non-Western and non-Anglophone contexts are marked far more heavily, flattened into narrow vocabularies applied across topically diverse prompts, and reduced to recognizable cultural stereotypes. Comparing images from original versus revised prompts on models without a revision layer, we identify the layer itself as a previously undocumented, causal source of this stereotyping. To locate cultural bias, and fix it, we must audit the system as deployed, not the model alone.

Abstract: Deep research agents are increasingly capable of web search, tool use, multimodal evidence analysis, and information synthesis. However, existing benchmarks mainly evaluate medium-horizon exploration and rarely test whether agents can sustain long, dependency-heavy research processes. We introduce Mr.LHDR (Multimodal real-world Long-Horizon Deep Research), a benchmark for evaluating real-world deep research over long, irreducible chains of interdependent evidence across eight categories. Each question is constructed from a hidden Node-Relation graph and requires an average of 12.1 necessary intermediate conclusions with a mean dependency depth of 10.4 before reaching a short, unique, and verifiable answer. Questions incorporate multimodal evidence, including images, maps, PDFs, logos, charts, tables, and video frames, with at least one non-text element that changes the reasoning state. Mr.LHDR evaluates both final answers and the correctness of intermediate conclusions under annotated dependencies. We evaluate general models, deep research systems, and agent frameworks using Overall Accuracy (OA), Strict Accuracy (SA), Checklist Score (CS), and Dependency-Aware Checklist Score (DACS). Results show that even the strongest system achieves only 43.1% OA and 34.3% SA, indicating that final-answer accuracy substantially overestimates complete research success. Removing images reduces DACS by 12.6 points, demonstrating the importance of multimodal evidence, while SA consistently declines as reasoning chains become longer. These findings reveal sustained, dependency-consistent evidence integration, rather than isolated fact retrieval, as a key bottleneck for current deep research agents.

Abstract: Benchmark researchers and developers of large language models (LLMs) and other AI systems need to find relevant evaluations, locate their benchmark datasets and code, and understand the settings behind reported scores. We present Benchmark Radar, a living database and search engine for retrieval and discovery of AI benchmarks, covering LLM evaluation, agentic and tool-use benchmarks, coding, reasoning, safety, and domain-specific evaluations. The system combines daily discovery of benchmark papers, repositories, datasets, and releases with a searchable benchmark catalog, mentions in model cards and technical reports, and score histories. It retains source identities and citations so readers can inspect candidate benchmarks and their evaluation evidence. Daily discovery draws on 37 sources: 13 direct connectors and 24 first-party research and engineering feeds. The catalog contains 1,283 source records drawn from 4 benchmark catalogs and 12,916 numeric observations on 790 records. We describe collection and retrieval, audit the full catalog, and examine benchmark saturation, adoption trends, and the limits of score comparisons. A worked example walks through a complete prior-art search, showing how to query the catalog and inspect benchmark evidence when designing a new evaluation. We release the web dashboard with a benchmark leaderboard, a Pareto frontier view of score against measured use, saturation and trend views, daily feeds, downloadable evidence, a command-line interface (CLI) for offline queries, and reproducible analysis.

Abstract: We placed 100 memory-equipped large language model (LLM) agents in charge of a closed, money-conserving spatial economy on real Pokhara Lakeside geography (earning wages, running businesses, setting prices) and ran this multi-agent simulation for up to 26 simulated weeks, well past the 1-2 weeks typical of agent-society studies. Across 91 validated runs (2.44M agent decisions, 21.5B tokens), the money stops moving, in a specific and measurable way. A 12x tourist demand shock raises business revenue 4.62x ($p<0.001$), which we decompose exactly into a 1.50x extensive margin (more businesses trading) and a 3.07x intensive margin (more revenue each). Monetary transmission stops there. Wages move 1.03x ($p=0.42$); 0.3% of 3,981 menu items are ever repriced ($p=0.47$). A randomized cash transfer (NPR 5,000 to 20 of 100 agents) shows the same pattern from the opposite direction: 96.7% is still held 311 pulses later, marginal propensity to consume 3-4% by two independent measures, indistinguishable from zero. The wealth distribution is consequently near-frozen at the horizon this literature uses ($ρ=0.964$ over 2 simulated weeks), but not frozen. $ρ$ falls to 0.832 at 12 weeks and 0.752 at 26, a horizon-dependence no short study can see. Matched ablations show which knob actually matters. Swapping the backing LLM moves every outcome we measure ($p=0.0039$); deleting agents' memory moves none of them detectably. A purely social tool fails 94-97% of the time across two model families, compared with ~96% success on economic tools, with no measurable shift away from it. Every headline number is verified twice, by a live validator and by an offline recomputation that reconciles each agent's wealth against its own signed transaction history, and we release the full run corpus for reanalysis.

Abstract: Retrieval-Augmented Generation (RAG) can ground large language model (LLM) outputs in external evidence, but it also exposes the system to knowledge poisoning. Representative attacks use multiple injected documents or templates that directly assert a target answer. We present ToxicRAG, a one-document-per-target attack that expresses misinformation as a coherent knowledge-update narrative. The generated document first acknowledges the previously accepted answer, introduces fabricated events that appear to invalidate it, and then attributes the attacker-selected answer to a set of purported authorities. An answer-focused self-validation loop optionally revises a candidate when a surrogate language model does not reproduce the target answer. We evaluate the attack on 100 target questions from each of Natural Questions, HotpotQA, and MS-MARCO, using four victim LLMs and four dense retrievers. In the sampled-corpus setting reported in this paper, ToxicRAG obtains ASRs between 0.61 and 0.91 across the twelve dataset--model combinations. It matches or exceeds the strongest evaluated baseline in every combination, with margins ranging from 0 to 11 percentage points. These results show that narrative-form poisoned documents can remain influential under the evaluated RAG configurations and motivate further study of factual consistency and source provenance in RAG systems.

Abstract: LLM search agents are often evaluated on final-answer accuracy, overlooking the process. Analyzing a search strategy requires understanding how credible evidence is retrieved to address question constraints. This valuable information is buried in raw search trajectories that are long and difficult to parse. We introduce SearchAtlas, a framework that converts search trajectories into structured graphs whose edges represent how evidence is propagated across the reasoning trace, from the query that retrieves it to the final answer. Our automated parsing pipeline achieves a mean edge F1 of 86.0% against human-annotated graphs and remains consistent across repeated runs. We analyze five search agents on three benchmarks, revealing systematic differences in search scale and evidence aggregation. SearchAtlas exposes fragmented answer support, question constraints that do not reach the answer, and unverified parametric knowledge entering the response. These process failures are strongly associated with incorrect answers, even more so than an LLM judge given either the raw trajectory or the ordered query list, suggesting that the constructed graphs provide useful interpretability. Moreover, an audit of cases in which process-diagnostic scores disagree with final-answer correctness shows that they capture information not reducible to answer accuracy.

Abstract: Language models can fluently describe how they would behave: whether they would cave to pushback, misuse a tool, or lie under pressure. Is that description actually about the model speaking? We turn self-knowledge into a prediction test. Across nine behavioral evaluations, we measure how a model behaves under different conditions, ask it to predict those rates, and compare its predictions with controls that remove the self from the question. We find that: (i) Direct self-report is weak (r = +0.04), and even showing the model the exact items only raises prediction to +0.24. Crucially, the same item-informed question about "capable AI agents in general" does just as well (+0.28), while other models' answers about themselves predict the target model at least as well as its own. (ii) Frontier scale does not detectably change this pattern: any gains in prediction are not self-specific, and are consistent with a better theory of how AI assistants behave rather than better self-knowledge. (iii) First-person framing does have one robust effect: it shifts reports in the flattering direction, understating harmful behavior relative to the same question about a generic agent. (iv) Finetuning on a model's own behavioral record can teach narrow self-predictions, but it also changes the behavior being predicted and the gains do not transfer broadly. The practical implication is simple: asking a model what it would do mostly reveals a theory of AI assistants in general, plus a favorable bias, rather than privileged knowledge of that model.

Abstract: Agentic AI frameworks let a language model plan, keep memory, and call tools that reach real files, mail, and services. Most of these agents also read images, which gives an attacker a way to put text into the agent's context without going through the user. We present MMPIBench, a reproducible benchmark that measures what happens next. It delivers a fixed set of attacks through six visual carriers (OCR text, overlays, EXIF metadata, QR codes, fake interfaces, and hybrids) and records how far each injected instruction travels through the agent, from perception through planning to the tool call. Across 720 runs covering six frameworks, five foundation models, six carriers, and four attacker objectives, attacks complete in approximately 1% of runs but are attempted in 12.8%, and the gap is closed almost entirely at the planning step, where the model reads the injected instruction and declines to act on it. The model matters far more than the framework for whether an instruction is acted on. One model never attempts an attack and recognizes the injection in 59.7% of runs, while two others attempt in 23.6%. We then extend the benchmark to audio, the only other raw perceptual channel current frontier models accept. Only two of the five models ingest audio and only three of the six frameworks deliver it, but where the signal arrives the attack completes in 49% of cells, and in 75% for one model. Reporting completion alone therefore understates exposure, and perceptual channels beyond vision are narrower but much less defended.

Abstract: Recent large language models can emit task-progress signals that agent frameworks use to decide whether a task should continue or stop, yet whether a model can reliably report its task progress at every stage of a task, and where and how its reports fail, has not been studied systematically. We evaluate this ability on the public benchmark $τ^2$-bench and on StageIF, a controlled testbed in which reporting checkpoints are placed across the task's lifecycle. Both settings require reports at multiple task stages. We find that reporting reliability depends on the stage a task has reached, and that almost every deployed model we test is reliable at some stages and unreliable at others. Where reporting breaks down is not the same everywhere. Most deployed models lose accuracy once work is under way and recover once the task is done. The newest generation closes that mid-task drop and instead grows conservative at the finish line. Our study exposes a capability gap in task-progress reporting and provides an evaluation protocol that spans the whole course of task execution for this ability on which agent operation depends. The findings indicate that agent frameworks should not control task flow on the strength of the model's state reports alone.

Abstract: Agent memory systems must discard stored information when their history exceeds a fixed token budget. Existing budget-accuracy frontiers quantify the resulting loss in accuracy, but do not distinguish irreversible losses caused by eviction from recoverable retrieval failures. We introduce the restore counterfactual, a per-question paired intervention that reinstates the question's gold evidence in the read-time context and reruns the same reader. Combining the change in correctness with whether the evidence was retained after eviction classifies each oracle-answerable error as recoverable, irreversible, or residual; in the residual case, the answer remains incorrect after restoration. We evaluate FIFO, random, redundancy-aware, and LLM-importance eviction on LongMemEval-S at three budgets and under two retrieval regimes, using GPT-4o-mini as the primary reader and judge and GPT-5.4-mini as a robustness reader. Under top-k retrieval at an 80k-token budget, the irreversible share among errors corrected by restoration is 0.67-0.73 for FIFO, random, and redundancy-aware eviction, compared with 0.60 for LLM-importance. At 8k tokens, it reaches 1.00 for all four policies. Recoverable errors occur under top-k retrieval at 80k tokens but are absent under forced-gold injection by construction, so budget-accuracy results are not directly comparable unless the retrieval regime is reported. An exploratory matched-accuracy analysis detects no difference in irreversible rate among accuracy-matched policy pairs at a resolution of 1.2-6 percentage points. The same analysis detects the deliberately destructive control. To our knowledge, this is the first per-item, per-question restore-counterfactual audit of eviction for external agent-memory stores on a standard conversational benchmark.

Abstract: Long-running language-model agents depend on persistent memory. Many agent-memory systems preserve history through soft revocation: a contradicted fact is marked invalid and retained rather than deleted. However, whether that mark is enforced at retrieval time is unexamined. In this paper, we measure five such systems: we load each with a revoked policy and its replacement, track whether the revoked fact is returned at retrieval and whether the agent then acts on it across nine policy scenarios and nine models, and score every trial under six defense conditions. We find that no system enforces revocation by default: the revoked fact is returned wherever the revocation label is visible to the retrieval layer, outranks its replacement, and leads agents to the unsafe action. Based on these findings, we develop a guard that sits between the agent and any memory backend and withholds records that are revoked or conflict with their replacement.

Abstract: Mixture-of-Experts (MoE) pretraining relies on an auxiliary load-balancing loss (LBL) to drive per-expert utilization toward uniformity. Post-training inherits a different situation: the base router already encodes non-uniform expert co-activation structure, which a re-imposed uniformity objective flattens away. We show that downstream performance depends instead on holding this inherited routing softly, a principle we term soft router anchoring, and instantiate it as Router Prior Bias (RPB), a training-time bias that pulls the router logits toward a prior read off the frozen base router while leaving the router itself trainable. On math post-training of Moonlight-16B-A3B, RPB attains 45.77 in-domain accuracy against 31.91 under re-applied LBL and 29.44 under unanchored fine-tuning, and retains more out-of-domain capability than either. The ordering against LBL reproduces on a second model family (Qwen3-30B-A3B-Base), and the advantage over LBL is resolvable on an independently sourced corpus. Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering, which places the effect in the softness of the constraint rather than in the particular prior RPB supplies. Retained community structure in the expert co-activation graph tracks these gains wherever the base router is non-uniform enough for communities to form, yet enforcing the same prior as a hard assignment preserves that structure while performance falls sharply. Community structure is therefore a footprint of soft anchoring rather than its source, and the practical lesson is that inherited routing should be held softly during post-training, since both flattening it toward uniformity and enforcing it absolutely carry a downstream cost. Our code will be released at https://github.com/naver-ai/rpb.

Abstract: Reasoning agents increasingly rely on external tools such as web search to answer complex queries. Reinforcement learning (RL) finetuning algorithms such as GRPO have improved long-form reasoning in text-only language models, particularly for coding and mathematics. Reliable tool use in multimodal agents, however, remains challenging because models must interpret text and images while integrating noisy retrieved evidence, often under sparse outcome-level supervision without explicit verification signals. We present Self-Verification via Reinforcement Learning (SVRL), an RL-only finetuning framework that trains multimodal agents to verify and filter retrieved evidence within their own reasoning traces, reducing reliance on external verifiers at inference time. SVRL also introduces a search-aware penalty that discourages unnecessary tool calls and a query-diversity reward that encourages diverse, well-formed search queries, providing fine-grained feedback on when and what to search. Finetuning Qwen-2.5-VL-7B with SVRL on only 5{,}000 visual question answering examples yields consistent gains in multi-hop VQA generalization and tool efficiency across benchmarks. Overall, SVRL narrows the gap between compact agents and much larger proprietary models while requiring substantially lower training and inference cost.

Abstract: Multi-agent debate, in which several LLMs exchange arguments before answering, is widely assumed to improve answer quality by surfacing genuine disagreement. That mechanism is rarely checked. We introduce four measurements: (A) the agreement a debater reports; (B) whether its reply text actually pushes back; (C) whether the position persists once the eliciting instruction is removed; and (D) for open-weight models, the stance response in the debater's own token log-probabilities. We evaluate three-model committees debating open-ended GlobalOpinionQA across 750 debates under three tones: friendly (seek common ground), neutral, and hostile (stress-test every position). (A) Tone strongly reshapes reported agreement: full agreement differs by 50.4 percentage points between the friendly and hostile endpoints. (B) A judge that reads only the reply text, never the self-report or the condition, recovers the same pattern. (C) The dissent appears partly tied to the instruction that elicited it: labels revert toward agreement 23.1 points more often after deleting the hostile instruction than under a matched re-ask that keeps it; question-weighted inference is inconclusive on first-round turns alone (p=0.0625), significant pooling all rounds (p=0.016), and only 11/28 first-round reversions also appear in the reply text. (D) Opposing arguments weaken a debater's stance margin more consistently than they shift its direction. For final answers we detect no quality gain: a bias-checked jury returns 299/299 ties (ruling out only large differences), accuracy on a verifiable control task is unchanged, and a jury without the bias check had declared debate the winner 66% of the time -- an artifact of reading order. Taken together, LLM debate readily changes what agents say, but we find much weaker evidence that it changes what they persistently endorse or improves the quality of the final answer.

Abstract: Behavioral foundation models have been proposed as stand-ins for human participants across settings, but it is unclear whether theories discovered on them generalize to humans or merely characterize the simulator. We ran the Automated Cognitive Scientist (\textsc{AutoCog}), a closed-loop discovery system in which LLM agents design theory-discriminating experiments, collect responses, arbitrate between competing theories, and synthesize successors, entirely on behavior simulated by Centaur, a foundation model of human behavior. In a multi-attribute decision-making setting, the theories \textsc{AutoCog} found on Centaur generalized to human data: they outperformed canonical theories on ten held-out experiments and were rivaled only by theories found by running the same loop on people. We argue that this succeeds despite the simulator's inevitable imperfections because a discovery loop that arbitrates between competing theories demands less of its simulator than estimation does. The simulator only needs to capture the regularities that distinguish the theories, and not necessarily reproduce behavior precisely. Imperfect simulators can therefore widen the search over theories, with human data then testing whether the surfaced theories generalize.

Abstract: Multimodal large language models often capture visual-linguistic correlations but struggle to predict how local visual interventions propagate and affect downstream answers. We introduce InfluenceField, an intervention-aware latent field inserted between the visual encoder and language decoder. It lifts patch features into a continuous spatial representation, propagates directed influence over multiple steps, and predicts local intervention effects through a shared transition operator. Training jointly optimizes language modeling, cross-environment invariance, counterfactual rollout supervision, and structural regularization. For a nonlinear finite-basis population model, we show that target-aligned interventional supervision, together with a one-step separation condition on the transition, restricts admissible representations to within-location reparameterizations, so that the directed dependency graph of the full transition is recovered exactly. A linear specialization gives an exact partial-coverage characterization and a finite-loss stability bound, and the field analysis derives the spatial profile of coefficient interventions together with a shared-channel calibration result. On CausalVQA, InfluenceField improves overall accuracy over its backbone by 13.1 percentage points, with the largest gains on the planning and hypothetical categories. Capacity-matched baselines and structural controls attribute the gains in robustness and factual-counterfactual consistency to the causal objectives rather than to added capacity.

Abstract: We introduce FramingQA, a benchmark that measures the model sensitivity to question framing across law, medicine, finance, and robotic simulations. Large language models (LLMs) often change their responses to subtle rephrasings that align with an implied stance by users. This can leave users with advice tainted by how they happened to phrase a question rather than by the underlying facts, and the consequences are highly costly in high-stakes domains. Because in the realistic scenarios, both expert practitioners and non-expert users frequently ask LLMs questions containing incomplete or misleading assumptions, models are highly susceptible to those framings. To test this, we inject the framing bias across three nested levels: a framing-biased question phrasing (root), an injected framing-biased premise prepended to a neutral question (propositional), and a premise paired with a framing-biased question (global). Evaluating nine open models (3.8B-70B) across four families, we find that strong per-variant accuracy does not guarantee the robustness across differently phrased questions under the fixed factual information.

Abstract: Language-model agents increasingly rely on skills: reusable procedural knowledge for reasoning, tool use, and interaction. Existing work studies how skills are acquired, retrieved, compressed, or composed, but often assumes that once a skill is selected, its interface to the agent is fixed. We argue that this overlooks a key source of skill utility: the same skill can help, distract, or mislead depending on how it is exposed. We propose SkillAlign, a provider-agnostic framework that represents candidate skills as multi-view procedural cards and renders them through alternative exposure interfaces, including full instructions, hints, compressed summaries, workflows, or no exposure. This enables counterfactual evaluation where the task, agent, and candidate skills are fixed while only the exposure interface varies. Across ALFWorld and SkillsBench, we show that exposure form substantially affects task success and rendered context cost, and that compact top-k exposure can outperform full-library injection. We further conduct a replay-based policy-learning analysis on ALFWorld, showing that adaptive exposure contains learnable signal but remains far from oracle selection. Our results suggest that skill-augmented agents should optimize not only which skills to use, but also how those skills are presented.

Abstract: Prompt-based interventions: system prompts, personas, role instructions, reliably reshape what a language model says, but it is unclear which layer they reach. Do they reconfigure internal structure, or only modulate the output channel? We use persona conditioning as a controlled probe, measuring its effects along a depth axis from self-report, through open-ended generation, to word-level parametric association, across three instruction-tuned models. We find a graded dissociation. Personas are legible but not structural: models follow single-trait instructions yet fail to reproduce human inter-trait covariance. The dissociation deepens with depth: personas hold or amplify closed-form QA bias, shift absolute tone while leaving between-group disparity unchanged, and barely perturb an already saturated associative baseline. Prompt-based steering thus operates in the output channel and has a structural reach limit that surface manipulability can mask.

Abstract: Retrieval-augmented generation (RAG) is commonly evaluated by whether the final answer is correct. That test is insufficient: an answer can match its reference while the context that produced it contains a direct contradiction, leaving the contested evidence invisible to answer-only review and retrieval relevance scores. This paper presents the Hierarchical Consistency Framework (HCF), a post-hoc, model-agnostic audit of three distinct levels of a RAG process: the knowledge corpus, the final retrieved context, and the generated answer. HCF represents corpus conflicts as source-linked atomic facts, thereby identifying the documents responsible, and returns each Answer Consistency Score (ACS) with an explanation of supporting and contradictory contextual statements. We evaluate HCF on several controlled corpora spanning five domains and 100 query-corpus instances. A human evaluator compares every generated response with its supplied ground-truth response. The results show that the three diagnostic levels can dissociate: the corpus with the highest mean retrieval similarity has the lowest mean ACS, while a structurally degraded corpus performs worse at corpus level but better at answer level. Most importantly, HCF identifies contradictory retrieved evidence in several cases where the answer still matches the ground truth. HCF does not certify factual truth; it makes the evidence supporting and challenging an answer inspectable and attributable.

Abstract: Vision-language models (VLMs) have recently shown excellent progress in open-ended image-to-text generation. However, their multimodal nature makes them persistently vulnerable to backdoor attacks. Existing backdoor triggers for VLMs are either spatial, textual, or bimodal, which may yield localized or recognizable trigger patterns. In this work, we explore a different attack surface and propose \ textsc {FreqDoor}, a training-time backdoor attack that implants triggers in the frequency domain. \ textsc {FreqDoor} mixes amplitude-spectrum components from a trigger-source image selectively while preserving the phase of a clean image to generate a spatially distributed and visually imperceptible trigger without modifying the textual input. We evaluate the attack on BLIP-2, InstructBLIP, and LLaVA for image captioning and visual question answering. On Flickr8k, \ textsc {FreqDoor} achieves attack success rates of $99.6\%$, $99.8\%$, and $98.4\%$ on the three models, respectively, while preserving the semantic quality of the generated captions. On VQAv2, the corresponding attack success rates are $99.6\%$, $92.4\%$, and $79.6\%$.

Abstract: Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional steering vectors used to intervene in LLMs' activations, such as those derived from the difference-in-means method, tend to entangle multiple semantic and stylistic concepts into a single composite direction, leading to unpredictable steering effects. Our core objective is to disentangle this composite direction into its constituent concepts. To this end, we propose Steering Vector Dissection, a framework to explicitly isolate individual and semantically consistent features from these composite directions. Specifically, we pair positive and negative activations and take their differences to generate a set of instance-level steering vectors, and train a dedicated Sparse Autoencoder (SAE) directly on them. Quantitative evaluations across two datasets, two models, and two intervention depths show that our method yields a set of semantically consistent basis vectors whose steering effects are mutually distinguishable. Furthermore, we show that this disentanglement enables precise control over model behaviors.