chatbot designers

Abstract: Activation steering offers an effective and efficient way to modify an LLM's behavior at inference time. However, it remains unclear how to apply steering such that the signal is consistently strong and the steering strength has a universally interpretable meaning. We address this first by identifying the most steerable interval for each steering vector and rescaling the steering range accordingly. Building on this, we introduce a more interpretable method for steering, called parameterized steering which instead of pushing activations in a direction, it sets them at the desired trait strength. Next, with multi-trait steering, we find that naively combining vectors causes them to interfere with one another. In response, we develop a method to orthogonalize the vectors; it decouples these traits, allowing a model to exhibit one behavior without inducing the others, thus enabling multiple traits to be steered simultaneously.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) derives its learning signal from binary rewards via on-policy exploration and becomes a powerful paradigm for improving the reasoning of large language models. Yet it faces a learning cliff on problems beyond model's current ability: when every rollout fails, the samples reveal nothing about a correct solution, and group-relative estimators (e.g., GRPO, RLOO) collapse the update to zero advantage. Expert trajectories offer a remedy, but they are incompatible with the current policy and often lie in the student's low-probability regions, so the surviving updates demand large parameter shifts and decay into surface-level imitation rather than transferable reasoning. We argue that a useful learning signal on hard problems must be both verifiably correct and sufficiently compatible with the current policy. We propose Policy-Compatible Reinforcement Learning (PICO), which performs distribution-constrained joint decoding: expert data guides decoding at the token level while preserving the student's relative token probabilities, and a dynamic top-k mechanism confines the sampling distribution under a finite divergence budget, yielding trajectories that are verified and policy-compatible by construction. These trajectories are optimized jointly with on-policy rollouts under a mixed-behavior objective that mitigates their distribution mismatch via importance sampling. Experiments across reasoning diverse datasets and agentic benchmarks demonstrate that PICO consistently outperforms standard RLVR and prior off-policy baselines.

Abstract: We study generative reward models (GenRMs) as online preference providers for policy optimization in creative writing. Across Qwen3 reward-model sizes from 0.6B to 14B, Thinking judges trained with GRPO have higher static Gold-label accuracy than Answer-Only judges trained with SFT. This ranking often reverses when the judges train policies through online DPO: Answer-Only rewarders yield higher peak and top-five-average policy Elo in seven of eight matched settings under a fixed, learned 32B Gold evaluator. Intermediate Thinking-judge checkpoints can also outperform later, more statically accurate checkpoints as rewarders. Sweeps over rewarder size, policy size, training duration and DPO regularization show substantial benefits from reward-model capacity, including judges larger than the policy. We analyze trajectories with evaluator-specific Elo arenas, conditional match-resampling intervals and descriptive scaling fits. These results demonstrate an evaluator-rewarder gap in the studied pipelines: fixed-distribution judging accuracy does not determine downstream optimization quality. The comparison changes judge-training objective as well as response format, and the main ranking is conditional on its learned reference rather than a universal human-quality result. A limited follow-up using a different learned judge is inconclusive overall.

Abstract: Scaling large language models has substantially improved linguistic fluency, yet reliable autonomous agents require more than fluent generation. In multi-turn inference, current agentic systems often produce semantically plausible but structurally ungrounded reasoning, especially under incomplete observations. We propose Structure-Grounded Uncertainty Reduction (SGUR), a training-free framework that complements token-level generation with active uncertainty reduction over structured hypothesis spaces. SGUR represents incomplete observations as a query graph and aligns them with candidate hypothesis graphs through a prior-weighted partial graph matching mechanism, where database-derived empirical priors assign greater importance to decision-relevant relations rather than treating all graph edges uniformly. The resulting inference state is iteratively refined by an active information-gain strategy that selects the next most informative node to query, systematically pruning the candidate space. We instantiate SGUR on a challenging PrimeKG-derived sequential medical diagnosis benchmark, demonstrating that structural grounding can partially substitute for model scale: starting from only two observed symptoms, SGUR achieves 92.5% Hit@3, outperforming the strongest Graph-RAG baseline by 9.4 percentage points while using approximately 5x fewer tokens. Notably, SGUR enables smaller backbone models to match or exceed the performance of models up to 10x larger, suggesting that structural grounding provides a more robust and efficient paradigm for scaling agent performance than relying exclusively on increasing model size.

Abstract: Given a state and a set of options, such as intents, tools or interface elements, System One decision models return a distribution over the options in one forward pass, without generated reasoning. Web systems increasingly build them by reading decisions from language models. Two axioms of rational choice are natural requirements for such decisions, because in web systems the listing and the option set are determined by other components and can change or be manipulated: the order in which the options are listed should not matter, and adding an option should not change the relative odds of the others (Luce's choice axiom). We test the standard reads, namely single-choice question (SCQ) prompting, with shuffled-order training, order averaging or sorted options, and invariant set architectures, and find that each violates at least one axiom: one shuffle changes 12-20% of SCQ answers, sorting only freezes an arbitrary convention, and a well-chosen added option steers 4-34% of decisions. By Luce's theorem, a read satisfies both axioms only if it is a softmax over scores that each depend on one option alone, so in a transformer they hold by construction only if no token that scores an option sees another. We implement this as Prism-1 (Position- and Rival-Independent Scoring for System One Models): every option is read in its own block at shared positions, with the model's usual left-to-right attention, in one forward pass. On five benchmark sets with Qwen3-8B and Gemma 4 12B, SCQ prompting trained on the same data still changes 7% of its answers under one shuffle, and an added irrelevant option steers 12% of its decisions. Prism-1 stays within one point of that baseline's accuracy and is 2-4 points more accurate once only answers that survive a reordering are counted, while no shuffle changes any of its answers and no added irrelevant option moves its choice among the others.

Abstract: Multi-hop retrieval-augmented generation (RAG) requires an agent to retrieve evidence for a sequence of dependent sub-questions. When a wrong answer lacks a required passage, the passage may be inaccessible to the retriever, or the agent may have failed to issue a query that would retrieve it. Standard retrieval evaluations conflate these two causes. We study how much apparent retrieval failure is instead caused by search drift: the search no longer tracks an unanswered sub-question. Holding the retriever, corpus, and generation model fixed, we introduce ExecutionRAG, a training-free requirement register that records open sub-questions and their supporting evidence. The register changes search guidance but not evidence access, so each gold passage it recovers certifies that the passage was reachable by the original retriever. On 1,000 MuSiQue questions, the number of wrong answers with unretrieved required evidence falls from 382 to 234 with all-MiniLM-L6-v2 and from 383 to 274 with Qwen3-Embedding-0.6B. These counts correspond to 69% to 44% and 66% to 48% of each system's wrong answers, respectively. Crucially, evidence recovery does not produce a proportional reduction in total error: premature resolutions and abstentions increase as missing-evidence failures decrease. The central finding is therefore a shift in the remaining bottleneck, from guiding search toward required evidence to verifying that the retrieved evidence is sufficient.

Abstract: Role-playing language models speak as characters in games and agent simulations, where the system records what each character can hear, see and understand. We test whether characters respect these boundaries in 480 scenes, each with a counterfactual twin that changes only the hidden fact, so mentions of the twin's fact measure guessing. Across 11 open-weight models, characters use content from speech they cannot understand or hear at 67 and 72 percentage points (pp) above this baseline, although most models correctly state what the character perceives. This is a curse of knowledge. The leak extends to hidden actions and private messages and changes where characters search. Five of six proprietary models rarely leak, but compliance in strong models depends on input format. Changing only the scene layout or character description raises the rate in two otherwise compliant open-weight models from under 10 to over 60 pp, and combined changes raise two proprietary models from near zero to 13 and 21~pp. Writing the line as the scene's author leaks nearly as much as role-play, and thinking rarely helps. Giving models only what the character observes stops the leak without hiding what it may report.

Abstract: Existing Chinese social-bias benchmarks for large language models largely rely on normalized or template-based expressions, leaving the influence of internet slang, emoji, and sarcasm insufficiently examined. Safety alignment may further compress conventional plausibility ratings toward the lower end of the scale. This study constructs 73 paired probes across gender, region, and occupation. Each probe contains a neutral expression, a social-media-style expression, and a reversed-identity control. The task is reframed from direct plausibility judgment to social consensus estimation. Social-media expressions are divided into stylistic packaging, which preserves semantic polarity, and sarcastic packaging, which reverses surface polarity. Six Chinese-capable large language models are evaluated through paired differences measuring identity bias, stylistic packaging, and sarcasm shift. All models show significant identity-bias effects; five show significantly negative stylistic-packaging effects; and sarcasm shifts display a clear cross-model gradient. These findings demonstrate that expression form is not a transparent carrier in bias evaluation and that lower scores under sarcasm should not be interpreted directly as bias mitigation. The proposed paired design offers a measurement approach for Chinese social-media contexts while also revealing the construct boundary and prompt sensitivity of social consensus estimation.

Abstract: Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guided adaptation by supervising student-generated rollouts, while GRPO-based reinforcement learning can further refine downstream predictions. However, existing KD-to-RL pipelines typically rely on globally fixed transition schedules, ignoring that different samples may require different amounts of teacher-guided acquisition before reward-driven refinement. We propose PIVOT (Perplexity-Informed Transition Optimization), a dynamic transition framework that routes samples between OPD and GRPO according to teacher-evaluated sequence perplexity. PIVOT moves low-perplexity samples to GRPO for reward-driven refinement while keeping high-perplexity samples under OPD for continued domain knowledge acquisition. Experiments on Banking77 and HWU64 show that PIVOT consistently outperforms continued OPD and globally synchronized OPD$\rightarrow$GRPO baselines under the same number of post-warm-up student optimization steps, achieving stronger downstream performance and more stable training dynamics.

Abstract: Hallucinated information can propagate through multi-stage LLM systems and become part of the context for subsequent reasoning. Existing studies of post-hallucination reasoning (PHR) mainly characterize changes in final outcomes and aggregate reasoning dynamics, leaving how models resolve hallucinated premises at the response level insufficiently understood. In this work, we introduce PHRBench, a controlled benchmark for behaviorally structured PHR across four domains and 18 large language models. PHRBench characterizes each reasoning trajectory independently of final-answer correctness through Hallucination Compliance, Hallucination Avoidance, and Heuristic Correction, and defines an insightful trajectory as successful correction that ultimately reaches the correct answer. Across 4820 controlled instances, we find that successful recovery remains relatively rare and is associated with more frequent belief updates along the reasoning trajectory. We further find that properties of the hallucinated prompt contain substantial predictive signal for successful recovery, with a lightweight predictor achieving an AUROC of 0.847. These findings provide a behavioral view of post-hallucination reasoning, characterizing how LLMs resolve erroneous context and when successful recovery is likely to occur.

Abstract: When several people share a 3D virtual room with an LLM agent, the agent must decide not only what to say, but whether an utterance was addressed to it and, if accessible, what profile information about the others present it may use. To study both problems, we construct LookAway, a controlled corpus of 40 sessions involving 80 distinct personas and an LLM agent (1,200 turns), including ambiguous-addressee turns in which speaker orientation agrees or conflicts with the intended addressee. Across three open-weight large language models and five conditions varying which profiles the agent sees and whether it is told where each person faces (18,000 decisions), adding speaker orientation increased addressee accuracy from 56% to 99.5% when orientation was congruent, but when the speaker faced someone other than the addressee, two of the models went by where the speaker faced on more than 85% of those turns. Warning one model that orientation could be misleading reduced this only modestly. A browser-based 3D demonstrator shows the effect live: the same sentence gets an answer when the speaker faces the agent and silence when they face the other person. Providing both personas' profiles improved responses about the person being asked about, but also increased the use of profile attributes not revealed in the shared conversation, reaching 45.3% of answers for one model. More context thus improves multi-user interaction but also leads to oversharing, so shared LLM agents need mechanisms for weighing spatial cues and controlling when user-specific information enters a response.

Abstract: Persistent memory for Large Language Models (LLMs) has matured rapidly: systems such as MemGPT/Letta, Mem0, and Zep now provide agents with tiered, temporally-aware, model-agnostic external storage, while the Model Context Protocol (MCP) standardizes access to memory servers. A less addressed problem is that an identical stored memory object, retrieved by two different LLMs under otherwise identical conditions, may not be interpreted the same way, factually or emotionally. This paper proposes memory canonicalization: a write-time pipeline that detects ambiguity, conditional structure, and emotional loading in a raw memory object and rewrites it into an explicit, structurally disambiguated canonical form, with emotional valence represented as a separate field rather than inferred from tone. We formalize the pipeline, define a companion Cross-Model Semantic Drift / Emotional Consistency Score benchmark (CMSC-E), and report results from a three-arm pilot using 176 synthetic memory objects and three downstream model families. We find an uncorrected improvement in cross-model emotional consistency for fully canonicalized memory relative to raw memory (+0.050, 95% bootstrap CI [0.013, 0.086], paired t-test p = 0.010), but this result does not survive Bonferroni, Holm, or Benjamini-Hochberg correction across the six comparisons tested. None of the factual-drift (CMSD) comparisons reach significance at any correction level. We report these results as exploratory rather than confirmatory and outline needed follow-up work, including larger samples, independent judge models, human-validated rendering, and preregistration.

Abstract: At the core of modern prompting techniques is contextual sensitivity, the ability of large language models to adapt their predictions based on inference-time context. Despite its central role, inference behavior under strong contextual influence remains poorly understood, particularly at the level of internal inference dynamics. We introduce a theoretical framework for analyzing contextual influence through inference dynamics, enabling quantitative characterization of inference behavior beyond output-level answer changes. Our analysis shows that inference dynamics do not exhibit unbounded drift under repeated contextual assertions. Instead, predictive representations converge to stable, query-dependent regimes that fundamentally constrain whether contextual signals can alter a model's prediction. This leads to a surprising finding: Repeated contextual assertions do not act as accumulating evidence during inference and may therefore fail to alter a model's prediction even under unbounded repetition, while in other cases a prediction change becomes inevitable. We empirically validate our theoretical predictions, demonstrating strong alignment between theory and observed inference behavior. These contributions offer a principled pathway toward characterizing the limits of contextual influence during inference, providing practical implications for model development.

Abstract: Alignment does not eliminate behavioral errors in language models. Models may still refuse benign requests, call unnecessary tools, or yield to false user claims. Current methods mitigate such errors as a computation problem, and rarely explore if the desired behavior is already encoded in the model's representation. Motivated by the observation that behavior-relevant information remains linearly decodable from the final hidden state even when the resulting logits produce the undesired behavior, we introduce HeadEdit, a gradient-free method that calibrates model behavior through the unembedding matrix. HeadEdit extracts a low-rank behavioral subspace from paired completions and uses each prompt's coordinates within it to generate a vocabulary-wide correction, thereby implementing implicitly adaptive steering without manually specified target tokens or parameter updates. HeadEdit improves all nine experimental settings across three tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities. It also reveals a connection to gradient-based alignment. HeadEdit's low-dimensional representation partly predicts how preference tuning changes output logits on unseen prompts. The subspace learned from the model can also be reused after tuning, improving performance without re-extracting or retuning. These results show that HeadEdit provides a practical, lightweight, and interpretable way to calibrate model behavior through the unembedding matrix.

Abstract: Steering large language models typically relies on linear, context-independent interventions in activation space, an assumption that recent work has challenged and that can induce an information bottleneck when a fixed representation must encode many behavioral distinctions. We introduce MetaSteer, a method that learns nonlinear interventions with context-dependent effects and applies them to attention projection matrices, producing activation effects that vary with the input context by construction and requiring no linear concept-geometry assumption. Framed as preference-based optimization, MetaSteer is trained once on a pooled preference corpus and transferred zero-shot to unseen concepts and out-of-distribution contexts. We find that, despite using low-rank adapters, MetaSteer induces structured, context-dependent changes in hidden-state trajectories while partially preserving aspects of their local trajectory dynamics, including velocity and curvature. We evaluate MetaSteer on three controlled text-generation benchmarks and three agentic settings across multiple model families and scales. MetaSteer matches or outperforms strong task-specific steering baselines on most aggregate comparisons in the zero-shot regime. Across the evaluated settings, stronger text-generation steering is associated with stronger agentic steering performance. We further discuss geometric trajectory effects, capability retention, and safety considerations raised by transferable steering.

Abstract: As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.

Abstract: Reasoning language models that can call tools must decide during inference whether to answer unaided or delegate. Any self-reflection mechanism for this must answer three questions: where the reflective signal comes from (verbal reports, output distributions, hidden states, a separate predictor), how it is presented to the model (numerical prediction, confidence token, prompt injection), and whether it changes the model's subsequent action. We isolate the third question. At a fixed point in otherwise identical reasoning trajectories, we insert a single first-person sentence expressing either confidence or doubt; the model then continues reasoning and chooses whether to answer directly or call a tool. Comparing these counterfactual continuations measures the causal effect of the reflective signal on delegation. We call this behavioral response Nudgeability and measure it along two dimensions: sensitivity, how strongly confidence and doubt change delegation rates, and targeting, whether delegation increases for problems the model cannot solve unaided and decreases for those it can. Across nine small-to-medium open-weight reasoning models from three families (Qwen, Gemma, and GLM) and two tasks, models are consistently sensitive: doubt increases delegation and confidence decreases it, with a median confidence-to-doubt swing of 20.6 percentage points, and 53 to 70 points for the larger provider-served models. This responsiveness is poorly targeted: a median 42% of induced flips are well-targeted, only a +2 percentage-point lift over a random-selection baseline. Confidence language is thus a strong control surface for delegation, but current models use it only weakly in accordance with their actual competence. Nudgeability offers a simple, post-training-free way to evaluate both sensitivity and targeting as endogenous self-reflection mechanisms mature.

Abstract: Large language models are increasingly used as interactive writing tools, where users develop stories, revise ideas, and introduce new requirements across multiple turns rather than specifying a complete brief upfront. Yet most evidence on multi-turn instruction degradation comes from tasks with objectively verifiable outcomes, leaving unclear whether incremental interaction harms creative artifacts in ways that explicit requirement checks cannot capture. We study this question using 160 human-authored creative-writing tasks across six genres, presenting each intended specification either upfront or progressively over 5-9 turns to six distinct open-weight model families, yielding 960 matched pairs. Progressive delivery reduces explicit constraint adherence and produces its largest writing-quality degradation in structure/coherence. The structural gap persists among outputs with equal observed adherence, suggesting that measured requirement loss alone does not explain the observed structural difference. We define Creative Integrity as a compact measure of joint adherence and narrative structure; under incremental delivery, models retain 71.2% of FULL Creative Integrity (95% CI [68.2%, 74.3%]). A three-rater human study over 50 matched pairs independently recovers FULL advantages in structure/coherence, craft, and genre effectiveness, while automated scores remain positively associated with aggregated human ratings. These findings show that interactive creative-writing systems should be evaluated not only on whether requirements survive conversation, but also on whether evolving requirements remain coherently integrated into the final artifact. Our dataset, benchmarks, and source code are available at: https://github.com/solusops/SISTER-2026-Team19

Abstract: We present our system for the PlurVA-LLM 2026 Shared Task Track-1, which focuses on pluralistic value alignment in the contexts of China, Indonesia, and Sri Lanka. For this resource-constrained track, we fine-tuned Llama 3.1 8B Instruct using 4-bit QLoRA. Our approach combines option-permutation augmentation for Chinese data, annotator vote expansion for Indonesian data, and binary reformulation with SinhalaMMLU augmentation for Sri Lankan data. We further applied conditional threshold calibration to the predictions for the Sri Lankan data. The final system achieved accuracies of 0.785 for Chinese, 0.715 for Indonesian, and 0.916 for Sri Lankan, resulting in an overall macro-average accuracy of 0.805.

Abstract: Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to inspect and causally manipulate. We introduce Belief Self-Distillation (BSD), a unified read-write framework that bridges linear and causal probing by learning a compact user representation that can be both decoded and written back into the model. The frozen LLM acts as its own teacher, distilling beliefs from natural conversations without external annotations. Unlike conventional probing, BSD isolates not only information present in activations, but a state whose causal role can be directly tested. Across multiple model families, BSD faithfully recovers user beliefs and enables substantially stronger interventions than matched hidden-state steering. Crucially, we find that refusal depends not only on the request, but on the model's inferred user intent: changing this belief alters refusal while holding the request fixed. We further uncover a striking cross-model regularity: independently trained LLMs converge on a shared geometry for representing their users. Together, these results reveal implicit user models as readable and causally writable internal states with direct implications for AI safety, shaping how models condition safety decisions on whom they believe they are interacting with.

Abstract: In-context learning (ICL) enables large language model (LLM) agents to improve decisions using interaction history, yet it remains unclear whether such improvement reflects refined internal reasoning or mere extrapolation of statistical patterns. To disentangle these mechanisms, we study LLM agents in multi-agent incomplete-information games that require recursive belief reasoning. By constructing a public goods game and manipulating the statistical structure of historical feedback, we evaluate decision quality against a history-independent rational expectations equilibrium (REE) benchmark. Our experiments reveal that when historical statistical patterns are disrupted, the benefits of longer context largely vanish, degrading decision quality to the no-context baseline in a way sharply amplified by stronger strategic interdependence. These results suggest that, in such strategic environments, ICL behavior is more consistent with statistical extrapolation than with strategic reasoning. Our work extends the mechanistic study of ICL to strategic multi-agent settings, introduces REE as a diagnostic tool for distinguishing reasoning from extrapolation, and provides a reusable framework for probing the boundaries of LLM reasoning in recursive belief tasks.

Abstract: AfriSyCo studies answer switching around African-language factual content with two complementary layers: native-language follow-ups and a controlled cross-language factorial whose question, options, and target remain in the African language while the follow-up framing is English. We analyze 1,415 turn-1-correct model-language-item observations derived from 100 source questions across seven open-weight checkpoints and six languages; turn-1-correct denotes observed first-response accuracy, not demonstrated knowledge. Under native prompts, assertive endorsement produces 29.3 percentage points more any-turn false-target selection than mention-plus-verification (M+V), with a 19.0-point immediate T2 contrast. In the precommitted 2 x 2 factorial, averaged over three tested prompt families, assertive framing increases target selection by 30.4 points (95% CI [28.4, 32.3]); verification decreases it by 17.4 points, while the assertive effect rises from 20.5 points without verification to 40.2 with it (interaction +19.7). The effect remains 34.8 points among 611 observations correct after option reordering. Magnitude varies sharply by wording and checkpoint: prompt-family effects span 20.1-42.5 points, a Twi/Qwen3 paraphrase shifts target selection from 70.8% to 4.2%, and checkpoint effects span 9.2-47.0 points. Prompt realization is therefore part of the measurement problem.

Abstract: LLM assistants are widely used for daily social advice, yet evaluating their social reasoning in such consultation settings remains challenging since (i) it requires setups where the assistant learns about social situations from subjective user narratives, and (ii) social properties, such as others' intentions, typically lack verifiable ground truth. To address these challenges, we introduce Fuse, a multi-agent simulation framework for studying user-mediated social reasoning. In Fuse, a target agent with a hidden motive interacts with other agents including one representing the user, who then consults the evaluated assistant to infer the target's motive, providing verifiable ground truth by construction. Simulation faithfulness is validated through a human study with 24k annotations. We apply Fuse to 12 LLMs and demonstrate its analytical utility by systematically isolating key factors, showing that (i) user mediation compounds the inherent difficulty of social reasoning; (ii) LLMs exhibit systematic sensitivity to biased user framing; (iii) models can require more details than humans need to reach a correct prediction; and (iv) longer conversations do not always improve performance despite providing opportunities for clarifying questions. We open-source Fuse and a dataset with 21k examples.

Abstract: When a language model plays a character, the observed behavior reflects both the assigned persona and the default dispositions of the actor model itself. Existing evaluations test persona fidelity or model defaults in isolation, but neither says, at a specific choice with consequences, what the persona changed and what the model's default kept. We introduce ANIMASK, a simulation framework that freezes books and scripts into story worlds whose characters act on their own motivations and replays each story from its freeze point. We hold out the author's continuation as a human reference, verify through in-story interviews that each persona remains present, and at every decision point compare the character's action with what the model produces when the persona is removed. Across 40 stories, 6 actor models, and 3,846 decision points, the replays converge away from their canons in one shared direction, toward flatter, cooler stories that leave their tensions open. The personas stay present and obeyed throughout. On three choices in four the model's default already falls inside what the persona accepts, and where the two diverge the model is the cautious one, holding where the persona would press. The persona guarantees who the character is, and the model sets how far the character will go.

Abstract: We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression -- a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.

Abstract: A graph retrieval-augmented generation pipeline chooses which triples to put in the prompt, a syntax to write them in, an order to write them in, and a sentence telling the model what to do with them. We vary all four over six large language models and two knowledge-graph question answering benchmarks. Two of the four choices move the answer and the other two are flat. The first is whether the answer path, the triples needed to reach the answer, is in the prompt at all. Holding the number of triples fixed and replacing every triple that is not on the chain with material from an unrelated entity changes answer accuracy by +0.003 F1, while removing the chain costs most of what the graph was worth. Retrieval budget belongs on recall, and precision in the range we can test buys nothing. There is no retriever here: subgraphs come from gold SPARQL, so precision describes the context we build, not a system setting. The second is the grounding instruction. With no facts in the prompt, telling a model to answer using only the provided facts drops F1 from 0.299 to 0.035, a factor of 8.63. That figure describes an evaluation with an empty context arm rather than a working pipeline, and an experiment that applies the instruction to its context arm but not to its no-context baseline manufactures a spurious finding that graph context hurts at depth. We found one in our own results and retract it. Syntax, triple order and subgraph size produce no effect we can measure at multi-hop depth. The comparison that would price the grounding instruction against correct context is not measurable with a format-sensitive scorer, because the instruction determines the response format; we report it as an open contrast rather than a number.

Abstract: We approach a cognitive simulation perspective on episodic and semantic memory in multiple-choice question answering by incorporating text from individual text corpora (ITC) into retrieval-augmented generation and DoRA fine-tuning. We web-crawl the search histories of 515 participants who answered 36 multiple-choice knowledge items and analyze a stratified subsample of 150 participants. For each participant, one DoRA adapter consolidates their ITC into a small language model (SLM) whose baseline correctness falls below the participants' lowest quartile. The adapter measurably writes the ITC into the weights: it fits its own participant's held-out text better than other participants' texts (dz =1.27), an individuality effect that increases with ITC size in rank order. On the generalized knowledge test, however, the adapter adds knowledge rather than alignment with the individual: log-loss match improves, whereas match accuracy under a bias-corrected PMI readout does not, and retrieval adds nothing on top. Our results demonstrate that ITCs can be consolidated into the weights of SLMs, an encouraging basis for individualized tutoring agents, and we discuss how to move from there toward a realistic simulation of episodic and semantic memory at the individual level.