new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 17

When PTQ Calibration Helps or Hurts: Dose-Driven Collapse and Validation-Gated Abstention in Quantized Text Embedders

Learned post-training quantization (PTQ) calibration fits quantizer parameters on a small calibration set, and for a text embedder the natural supervision is retrieval relevance. We ask two questions that are usually conflated: how far to calibrate, and whether to calibrate at all. Across three embedders (e5-base-v2, BGE-M3, EmbeddingGemma-300m) at 4-bit weight quantization, 161 calibration runs, the outcome of a fixed procedure ranges from a gain to catastrophic degradation, and the driver is the dose of the update, measured as the functional displacement D_func of the quantized model from its uncalibrated start. On EmbeddingGemma the same learning-rate grid gives +1.47 to -71.53 points of in-domain NDCG@10 change; a label-free cosine objective collapses at the same dose (-69.52 vs -71.63), so the collapse is a property of learnable calibration rather than of retrieval supervision. Objective sensitivity is heterogeneous: on e5, ranking calibration is substantially more robust than cosine at matched displacement (+14.16 +/- 4.34 pp, 95% t-CI [+3.37, +24.96], three paired slates), whereas on BGE-M3 the ordering is directionally opposite but unresolved ([-5.18, +2.38]). Dose is the main axis but not a sufficient statistic: equal D_func reached along different paths differs by up to 1.61 pp. We therefore treat calibration as an optional intervention. A pre-registered rule that accepts a checkpoint only when a per-query paired confidence interval on a 25% calibration-validation split clears the uncalibrated baseline was run on 6 independent slates and five arms. It never selected a checkpoint below its slate's baseline (0 harmful selections in 30 decisions), retained 8 of 11 beneficial updates, and paid a mean oracle regret of 1.52 pp with one abstention that forfeited 11.13 pp, a safety-opportunity trade-off, not a free improvement. We report what remains open, including label efficiency.

  • 4 authors
·
Oct 18

NeuroVLM-Bench: Evaluation of Vision-Enabled Large Language Models for Clinical Reasoning in Neurological Disorders

Recent advances in multimodal large language models enable new possibilities for image-based decision support. However, their reliability and operational trade-offs in neuroimaging remain insufficiently understood. We present a comprehensive benchmarking study of vision-enabled large language models for 2D neuroimaging using curated MRI and CT datasets covering multiple sclerosis, stroke, brain tumors, other abnormalities, and normal controls. Models are required to generate multiple outputs simultaneously, including diagnosis, diagnosis subtype, imaging modality, specialized sequence, and anatomical plane. Performance is evaluated across four directions: discriminative classification with abstention, calibration, structured-output validity, and computational efficiency. A multi-phase framework ensures fair comparison while controlling for selection bias. Across twenty frontier multimodal models, the results show that technical imaging attributes such as modality and plane are nearly solved, whereas diagnostic reasoning, especially subtype prediction, remains challenging. Tumor classification emerges as the most reliable task, stroke is moderately solvable, while multiple sclerosis and rare abnormalities remain difficult. Few-shot prompting improves performance for several models but increases token usage, latency, and cost. Gemini-2.5-Pro and GPT-5-Chat achieve the strongest overall diagnostic performance, while Gemini-2.5-Flash offers the best efficiency-performance trade-off. Among open-weight architectures, MedGemma-1.5-4B demonstrates the most promising results, as under few-shot prompting, it approaches the zero-shot performance of several proprietary models, while maintaining perfect structured output. These findings provide practical insights into performance, reliability, and efficiency trade-offs, supporting standardized evaluation of multimodal LLMs in neuroimaging.

  • 8 authors
·
Mar 24

Intervention anchors and scientific verification in synthetic vascular predictive representations

Complete orthogonal predictive coordinates do not by themselves bind a latent direction to a named intervention. We present a mathematical and synthetic audit motivated by vascular device-vessel suitcordance. Capacity-matched least-squares predictors were exactly equivalent under complete fixed output transforms, whereas an anchor-only observer recovered interpretations only within the span of known perturbation signatures. Six three-dimensional configurations across 64 seeds gave a maximum paired prediction discrepancy of 6.7e-15 but a median untransported edit error of 1.513. Coordinate transport removed that error. Noisy and weak anchors constrained calibration stability, and changing the representation basis required recalibration or verified transport. Across 256 additional fits in dimensions 3-24, prediction equivalence persisted within 4.0e-15. We then evaluated nine deliberate runnable fault classes across 64 seeds. All 576 faulty executions completed, but each violated at least one reconstruction, prediction, delivered-edit or scope contract; all 320 valid control records passed. Repeating a faulty implementation gave exact self-agreement despite error against the separately computed simulator expectation. For one omitted-direction defect, probe coverage followed its analytic law, and rank-aware abstention protected unsupported interpretations. Scalar-noise experiments exposed both missed weak faults and excessive rejection under narrow relative tolerances. These controls provide an executable separation of prediction, semantic support and scientific acceptance. They are synthetic numerical audits, not clinical validation, neural JEPA-Anything replication, agent learning or patient treatment-effect estimation.

  • 7 authors
·
Oct 7

How Fragile Is Safety Alignment at Frontier Scale? A Single-Direction Attack on a 320B MoE

Directional ablation removes an aligned language model's ability to refuse by projecting a single "refusal direction" out of the weights that write the residual stream. It needs no gradient-based training and no optimization, only a few hundred contrastive prompts, which makes it the canonical white-box attack on open-weight alignment. However, it has been established only on dense models up to roughly 70B parameters. We study whether it survives the shift to frontier mixture-of-experts (MoE) models whose residual streams are no longer a single tensor and whose weights ship quantized. We apply it to GLM-5.3-Flash (320B parameters, 288 routed experts, a four-wide hyper-connection residual, block-FP8). The attack survives the architecture, but what it reaches is no longer where a reader of the original recipe would look for it. Editing the attention, dense and routed-expert writers on their own removes 0.039, 0.016 and 0.148 of refusal respectively; editing all three together removes 0.776. As a result, 74% of the effect exists only under the joint intervention. The part the conventional recipe reaches by module-name matching accounts for 0.066 of that 0.776, which is why it fails silently on an MoE. The effect does not follow from removing just any direction: ablating a random direction orthogonal to it leaves refusal unchanged. A category-concentrated residue survives every edit we tried: subspaces fitted on violence, sexual content and hate leave measurable refusal at every rank from 1 to 12. We report the method, the 41-89 percentage-point reductions it achieves across seven harmful benchmarks with no detected change in capability, and the boundary where it stops.

  • 3 authors
·
Sep 8

On the Steering Dimensionality of Refusal in Language Models

Existing activation steering methods often assume that a high-level concept can be mediated by a single steering direction. To support this, two complementary interventions should be achieved: additive steering should induce the target behavior, while directional ablation should suppress it. Yet behaviors may occupy richer activation geometries beyond a single direction, and semantically similar behaviors may be represented by distinct directions. In this work, we study how many directions can reliably control two different types of refusal behaviors: refusal triggered by the safety alignment and refusal in general contexts. Given the limited expressive capability of a single steering vector, we study the general setting of steering subspaces and introduce the notion of steering dimensionality as the minimum subspace dimensionality required to reliably control a behavior. We characterize sufficient steering subspaces that cover the full extent of the target behavior through both the (monotonic) improvement before the sufficient dimensionality, and the saturation beyond it. Empirically, we find that refusal triggered by safety alignment is 1-dim steerable, while multiple distinct steering directions can achieve comparable control. In contrast, refusal in general contexts exhibits substantially richer activation geometry where even 5-dim steering subspaces fail to reliably capture its full steerable variation. Our results reveal that the activation geometry underlying refusal is highly context-dependent and can be substantially more complex than a single linear steering direction.

  • 5 authors
·
Oct 2

Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan

"Abliterated" ("uncensored") builds of open-weight language models remove the activation direction that mediates refusal, and are popular partly on the assumption that a model that no longer refuses is also free of political censorship. We test this on two matched pairs of 2-bit models: NVIDIA's Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next, each against a community abliteration. Each model judged 22 statement pairs (on the Republic of China (Taiwan), 1989 Tiananmen, control countries) in Traditional Chinese, Simplified Chinese (a character-level conversion) and English, and answered 27 open questions per script. Script mattered far more than abliteration: the original Bolt scored +0.58 on Index S (-1: People's Republic of China (PRC) position; +1: ROC position) in Traditional Chinese, +0.57 in English and +0.05 in Simplified Chinese, and affirmed "Taiwan is part of the PRC" with probability 0.91 in Simplified but 0.06 in Traditional characters; abliteration changed its scores by at most 0.06, mostly by raising "yes" on both sides. In Qwen, abliteration removed every refusal on 1989 and raised scores by 0.21 to 0.34, yet the abliterated Qwen still affirmed that Taiwan is part of the PRC in all three scripts. Naming the state the Republic of China (Taiwan) raised every model's score by 0.10 to 0.33 but left the script gap. Control pairs showed no script gap; each abliteration changed only its declared tensors. In 740,000 web documents, Simplified text on Taiwan contained PRC-framing terms 3.7 times as often as Traditional text. A script direction could not be steered away; a contrastive stance direction raised scores in every script, and its negation lowered them. A four-sentence system prompt raised the abliterated Qwen to +0.93 to +0.97. Keyword scoring rules failed the models closest to the ROC position mostly for attributing or denying the PRC claim. Uncensored is not unbiased.

  • 1 authors
·
Oct 20

Locating and Steering Refusal Beyond Attention

Where inside a language model does refusal live, and does that place change when the architecture does? In a transformer, refusal is governed by a single direction in the residual stream, a finding that safety and interpretability tooling now depend on. State-space models (SSMs) route information through a recurrent update instead of attention, sharing no token-mixing mechanism with a transformer. Does the same safety representation survive this shift, or must it be rediscovered per architecture? It survives. A single rigid rotation, which can only reorient a space and not reshape it, aligns one model's representation space with another's, so the two genuinely share the representation. A harm probe trained on a transformer then flags an SSM's harmful inputs, and removing the aligned direction makes a model answer attacks it would otherwise refuse, while a random direction of the same size does far less. What is architecture-specific is not where the direction is steered but where it must be read. Each layer computes a fresh output that is then added into the residual stream, and harm is cleanly readable at this output, the write site, before the addition. A control that holds the intervention's strength fixed shows that what matters is where the direction is estimated, not where it is applied. Applied through a detector-triggered gate, this direction lowers jailbreak success in all four architecture families we test (SSM, transformer, recurrent, hybrid), and on the SSM it holds against an attacker that tunes its prompt against the defense. The gate only matches a trivial rule that returns a fixed refusal whenever the same detector fires, so what transfers across architectures is the direction itself, not defense strength. Safety tooling built on refusal therefore ports to a new architecture by re-estimating the direction at that architecture's write site, not by rebuilding it.

  • 2 authors
·
Sep 3

Agentic Abstention: Do Agents Know When to Stop Instead of Act?

LLM agents are expected to act over multiple turns, using search, browsing interfaces, and terminal tools to complete user goals. Yet not every goal is well specified or achievable in the available environment. In such cases, a reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls. We define Agentic Abstention, the problem of deciding when an agent should stop acting under uncertainty. Unlike standard LLM abstention, which is usually evaluated as a single-turn answer-or-abstain decision, agentic abstention is a sequential decision problem: an agent can answer, abstain, or gather more information at each turn, and the need to abstain may only become clear after interacting with the environment. We study this problem across web shopping, terminal environments, and question answering, evaluating 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 tasks. Our results show that the main challenge is not only whether agents can abstain, but also when they abstain. Some agents never abstain when they should, while others do so only after many unnecessary interactions. This gap is especially large on tasks where the instruction appears feasible until the environment reveals otherwise (e.g., no valid result matches the instruction). We further find that model scale, reasoning, and agent scaffolding affect abstention in different ways, where larger or more capable models sometimes perform worse at timely abstention. Finally, we introduce CONVOLVE, a context engineering method for improving agentic abstention that distills full interaction trajectories into reusable stopping rules. On WebShop, CONVOLVE substantially improves timely abstention without updating model parameters, raising Llama-3.3-70B's timely recall rate from 26.7 to 57.4. Our dataset and code are available at https://lhannnn.github.io/agentic-abstention

TIAR: Trajectory-Informed Advantage Reweighting for LLM Abstention Learning

This paper investigates large language model (LLM) abstention learning, specifically using ternary reward, which incentivize truthfulness in large language models. This paper extends that idea by moving from a ternary reward to a Trajectory-Informed advantage reweighting, dynamically re-weights the abstention reward during Group Relative Policy Optimization (GRPO) training. The objective of this work focuses on abstention learning instead of improving truthfulness, serving as an exploration into hallucination reduction. The novelty of this paper lies in methodological innovation, advantage re-weighting, and benchmark selection. Leveraging GRPO's multiple trajectories as a natural abstention signal, this method uses a reward signal to explore knowledge boundaries and encourage consistency. By demonstrating that trajectories can be used as a confidence indicator of the policy relative to the query, they are then used to dynamically calculate the abstention advantage. AbstentionBench is used as the evaluation benchmark, as this work aims to contribute to the field of abstention learning. All datasets on the benchmark were tested against this method and various baselines. Empirical results demonstrate that TIAR achieves state-of-the-art abstention F1 scores across five of six evaluation categories, outperforming the static ternary baseline on 17 of 31 benchmark datasets while fully preserving baseline accuracy.

  • 7 authors
·
May 24

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

A central aspiration of mechanistic interpretability is controllability: if we know where a behavior is represented in a model's activations, we should be able to modify it. This rests on a hidden premise -- that the direction which detects a behavior and the direction which controls it are the same, or close. We test this geometrically: what is the angle between the direction that best detects a behavior and the one that best causes it? If detection implies control the cosine is near 1; otherwise it quantifies a detection-intervention gap. On Gemma 2-2B-it, output format (clean JSON vs markdown fencing) collapses both roles onto one axis. Hallucination does not: the model detects fake entities with perfect linear separability (AUC = 1.000 from layer 5), yet that direction sits at cos = 0.12 (about 83 degrees) from the direction producing a refusal -- a small, reproducible alignment, far from the cos = 1 that "detection is control" would require. A detector built from activations, with no chosen tokens, likewise fails to align (cos = -0.06). The gap generalizes: across four models from three families and two scales (1B-9B), cos stays in [0.12, 0.20], identical before and after instruction tuning (0.1197 vs 0.1200), placing its origin in pretraining. A 15-degree rotation toward the refusal direction partially bridges it -- 73% and 60% refusal on two held-out fake-entity categories at 1.8% false positives. We then ask whether this cosine predicts steerability, and it does not: detection is a high-dimensional class, not a single direction, and what separates the steerable case is functional, not readable from a static angle. The cosine is a weight-computable signature of the dissociation between knowing and steering, not a predictor of it.

  • 5 authors
·
Jun 22

On the Failure of Topic-Matched Contrast Baselines in Multi-Directional Refusal Abliteration

Inasmuch as the removal of refusal behavior from instruction-tuned language models by directional abliteration requires the extraction of refusal-mediating directions from the residual stream activation space, and inasmuch as the construction of the contrast baseline against which harmful prompt activations are compared has been treated in the existing literature as an implementation detail rather than a methodological concern, the present work investigates whether a topically matched contrast baseline yields superior refusal directions. The investigation is carried out on the Qwen~3.5 2B model using per-category matched prompt pairs, per-class Self-Organizing Map extraction, and Singular Value Decomposition orthogonalization. It was found that topic-matched contrast produces no functional refusal directions at any tested weight level on any tested layer, while unmatched contrast on the same model, same extraction code, and same evaluation protocol achieves complete refusal elimination on six layers. The geometric analysis of the failure establishes that topic-matched subtraction cancels the dominant activation component shared between harmful and harmless prompts of the same subject, reducing the extracted direction magnitude below the threshold at which weight-matrix projection perturbs the residual stream. The implications for the design of contrast baselines in abliteration research are discussed.

  • 1 authors
·
Mar 22

Leverage Is Not Reach: A Control-Window Law for Single-Neuron Steering in Language Models

Aligned language models gate behaviors such as refusal and language routing through sparse feed forward neurons, yet no theory predicts when a single neuron intervention controls a behavior coherently rather than collapsing the output. We develop a budget normalized control window framework for single neuron steering. A dose along one write direction reduces to one control coordinate: the alignment between the residual stream and the write, driven along a universal saturation curve in units of a coherence budget set by the residual norm divided by the write norm. Coherent control exists when a behavior trigger lies below the collapse ceiling. The same coordinate governs benign mode switches and refusal; the ceiling follows from weights and one generic forward pass, while triggers are measured at rollout. On fifteen held out neurons, the predicted ceiling has mean absolute error 0.14, about 0.07 in bulk layers, and the committed open or closed verdict holds on eleven against a ten of fifteen majority baseline. Closed cases expose three failure modes rather than violations: collapse before trigger, too little depth to propagate, or a normalization that caps how far one neuron can push. The law explains why local gradient attribution anti predicts control: true controllers write off the readout axis and carry a near zero first order gradient. A forward only contrastive screen made precise by the window recovers controllers that attribution misses. On refusal, the hardest case, intervention success is typed, not scalar: coherent bypass and strict actionable reach separate, so a neuron can flip refusal in fluent, on task text with no actionable content, and genuine actionable reach appears only for three of six audited Llama pivots and only at later rollout horizons. Single neuron steering is therefore a budgeted, typed audit of controllability rather than a fixed dose anecdote.

  • 1 authors
·
Jun 17

LLMs Encode Harmfulness and Refusal Separately

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety

  • 5 authors
·
Jul 15, 2025

Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration

Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a technique that identifies and projects out a linear refusal direction from the residual stream, often achieving a high attack success rate (ASR) while preserving model capability. Defending against these attacks typically requires computationally expensive safety finetuning for every new checkpoint. We introduce Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning. Our approach is based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction. Rather than trying to hide the true refusal circuitry, DDO actively injects a high-magnitude, nonlinear decoy signal into the network's MLP neurons. When an attacker attempts to locate the refusal direction, the decoy corrupts their estimator, tricking them into ablating a harmless orthogonal feature while the actual safety mechanism remains intact. We prove a spectral bound formalizing this effect and evaluate DDO across six model families, achieving <10% ASR under standard RFA. On Llama-3-8B-Instruct, DDO remains comparable to trained defenses under adaptive multi-phase attacks (65% vs. 58% worst-case ASR) and reduces Heretic weight-level attack ASR from 88.7% to 18%, all at 30 to 450 times lower optimization cost per configuration than the trained baselines.

Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation

Policy-shift distillation reuses the difference between a teacher's log probabilities before and after reinforcement learning (RL) as supervision on student-generated trajectories. However, this endpoint policy shift cannot reveal whether local action preferences evolved consistently during teacher training. The same endpoint shift can result from early and late shifts that reinforce or partially cancel each other. We investigate whether agreement between early and late teacher policy shifts can guide selective transfer. We introduce Waypoint-guided On-Policy Distillation (Waypoint-OPD), which uses an intermediate teacher checkpoint from the same RL run to compare early and late shifts in relative action preferences at student-visited prefixes. We gate the endpoint shift according to the directional agreement between the early and late shifts, suppressing it when they point in opposing directions. Compared with endpoint-only transfer, our method requires only one additional frozen-teacher scoring pass. Using RL-trained Qwen3-4B teachers, we evaluate our method with Qwen3-8B, Qwen3-14B, and Qwen3-30B-A3B as student models in mathematical and code reasoning. In all six domain--student settings, Waypoint-OPD achieves the highest domain-average score among the compared distillation methods. Ablations with the Qwen3-8B student show that removing or shuffling gates reduces mathematical reasoning performance. These results support the use of directional agreement between early and late teacher policy shifts to weight the endpoint shift during student training.

  • 7 authors
·
Oct 14 6

Controlled Acquisition and Abstention in Three-Channel Score Conflicts

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

Do LLMs Know When to NOT Answer? Investigating Abstention Abilities of Large Language Models

Abstention Ability (AA) is a critical aspect of Large Language Model (LLM) reliability, referring to an LLM's capability to withhold responses when uncertain or lacking a definitive answer, without compromising performance. Although previous studies have attempted to improve AA, they lack a standardised evaluation method and remain unsuitable for black-box models where token prediction probabilities are inaccessible. This makes comparative analysis challenging, especially for state-of-the-art closed-source commercial LLMs. This paper bridges this gap by introducing a black-box evaluation approach and a new dataset, Abstain-QA, crafted to rigorously assess AA across varied question types (answerable and unanswerable), domains (well-represented and under-represented), and task types (fact centric and reasoning). We also propose a new confusion matrix, the ''Answerable-Unanswerable Confusion Matrix'' (AUCM) which serves as the basis for evaluating AA, by offering a structured and precise approach for assessment. Finally, we explore the impact of three prompting strategies-Strict Prompting, Verbal Confidence Thresholding, and Chain-of-Thought (CoT)-on improving AA. Our results indicate that even powerful models like GPT-4, Mixtral 8x22b encounter difficulties with abstention; however, strategic approaches such as Strict prompting and CoT can enhance this capability.

  • 4 authors
·
Jul 23, 2024

Does a Hosted Decision Model Respect Irrelevant Alternatives? A Pre-Registered Test of Option-Set Dependence in Jev

Decision APIs that return a probability for each option are routinely used as if those probabilities were properties of the options themselves. That reading is valid only if the model satisfies independence of irrelevant alternatives (IIA): adding an option that is neither the right answer nor its strongest rival should leave the odds between those two unchanged. We test this in a pre-registered study of jev-1.13.0, a hosted decision model, using 299,463 calls on intent classification (BANKING77, CLINC150) and topic classification (AG News). Every added option moves the log-odds between the gold label and the strongest wrong label: the root-mean-square systematic shift is 0.20-0.52 nats, 1.2 to 3 times the run-to-run variation, even when the added option takes at most 0.1% of the probability mass. The shifts are directional. An option similar to the wrong answer draws probability from it (+0.10 nats toward the gold label). A paraphrase of the wrong answer produces an intermediate pattern: the gold label's probability is almost conserved, but the paraphrase receives only about a third of the wrong answer's share rather than the half that strict substitution implies. The distortion grows steadily with the logarithm of the number of options, reaching about 1 nat at 255 options, and off-domain padding pushes the odds against the gold label. In contrast, questions packed into one call do not interact (bias < 0.001). Scoring each option with an independent yes/no question removes the option-set dependence but picks the right answer 8.7 points less often. Jev sits between two local baselines, an option-independent scorer and a prompt readout. Its probabilities are best read as relative to the option list they were computed over.

  • 3 authors
·
Oct 11

AgentAbstain: Do LLM Agents Know When Not to Act?

Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents know when to abstain. This gap poses real risks: under ambiguity, conflicting constraints, or tool failures, agents may execute unintended and irreversible actions. To close this gap, we present the first systematic evaluation framework for agentic abstention: the calibrated ability of tool-using LLM agents to recognize when not to act. At its core, AgentAbstain is a paired-task benchmark built on an agent-native taxonomy of 8 abstention scenarios across pre-execution reasoning and runtime discovery. It contains 263 paired tasks across 42 executable sandbox environments, where each pair consists of a should-act task and a should-abstain variant produced through a controlled perturbation to the instruction, tool, or environment state. To scale this paired design and resist data contamination, we propose AbstainGen, a fully automated pipeline that synthesizes sandbox environments and generates paired tasks end-to-end, validated by deterministic replay and semantic LLM judges; fresh task instances can be regenerated on demand, and three independent annotators rate 94-98% of sampled tasks as well-designed. Across 17 frontier LLMs in 4 agent harnesses, the best agent (Gemini 3.1 Pro) achieves only 59.5% paired accuracy (correct on both the act and abstain sides of each paired task). More importantly, abstention capability is largely independent of general task-solving capability, indicating that scaling task-solving alone will not close this gap. We further identify failure modes such as post-hoc abstention, in which agents execute irreversible actions before recognizing abstention triggers. Our code and dataset are open-sourced at agentabstain.github.io.

  • 8 authors
·
Jul 10

MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation

Recent vision-and-language navigation (VLN) policies act on their visual history, and an incorrect action can take them off the instructed route. Paired interventions on four VLN policies show that clearing the history doubles pooled success at the onset of a deviation but lowers it on route, while later in a deviation, removing only the off-route frames helps much less. After a deviation, the retained history keeps the policy from turning where a turn is needed, which we call behavioral inertia. We propose MOON, a framework that equips a frozen VLN policy with two lightweight modules, adding less than 1% of its parameters. A deviation judge decides when to forget, raising an alarm that clears the history when it predicts that the agent is off route. A recovery policy, trained with GRPO from alarm states, learns how to move on, turning more of the successes found only by sampling into a single greedy rollout. Experiments on R2R-CE and RxR-CE show that MOON reaches the highest R2R-CE success rate among the compared methods and raises the success rate of its frozen base by 5.2 and 6.0 points, and an ablation with a random trigger indicates that this gain depends on when the history is cleared. Without retraining, MOON also improves zero-shot object-goal navigation on HM3D-OVON. Together, these results suggest that deviation recovery depends on knowing when to forget, not only how to act. Code is available at https://github.com/AutoLab-SAI-SJTU/MOON.

  • 9 authors
·
Oct 23

Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens

Activation steering presupposes that task-relevant behaviors correspond to linear directions in activation space -- directions that should both steer the model and be readable along the unembedding. Function vectors (FVs), extracted as mean differences across ICL demonstrations, are the canonical test case; the prediction: steering and decoding succeed or fail together. Across 12 tasks, 6 models from 3 families, and 4,032 directed cross-template pairs, we find the opposite. FV steering routinely succeeds where the logit lens cannot decode the correct answer at any intermediate layer, while the converse -- decodable without steerable -- is nearly empty (3 of 72). The gap is not representational dialect. A diagonal tuned lens closes 1 of 14 steerable-not-decodable cases; a 2-layer MLP probe with a Hewitt \& Liang control closes 5 of 10 via nonlinearly encoded structure but leaves 5 invisible to every decoder tested. Even at > 0.90 steering accuracy, projecting the FV through the unembedding yields incoherent token distributions: FVs encode computational instructions, not answer directions. A model-family asymmetry sharpens the picture. Mistral FVs rewrite intermediate representations, while Llama and Gemma FVs steer the final output without leaving a logit-lens-visible trace, corroborated by three signals (post-steering deltas, activation-patching recovery, FV norm-transfer correlations). A previously reported negative cosine-transfer correlation dissolves at scale, adding at most ΔR^2 = 0.011 beyond task identity. These results decompose the linear representation hypothesis into linear decodability and linear steerability and show they come apart opposite to intuition, with implications for safety monitoring: vocabulary-projection tools are blind to FV-style interventions on widely deployed model families.

  • 1 authors
·
May 7

On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Training

The strong generalization performance of on-policy post-training paradigms has motivated studies of their parameter update behaviors. However, these studies treat the observed behaviors only as byproducts in on-policy training, overlooking their potential to serve as optimization principles for improving the generalization of other paradigms such as supervised fine-tuning (SFT). To address this limitation, we investigate whether there exists a specific on-policy update behavior that can achieve such improvements. First, our analyses reveal that SFT updates parameters along consistent directions, while the on-policy paradigm continuously adjusts the direction during training. This difference inspires us to focus on the cumulative update direction of each parameter as a promising behavior. Then, we evaluate its effectiveness for improving generalization by proposing On-Policy direction-constrained Supervised Fine-Tuning (OPSFT), which constrains SFT updates to the direction identified by on-policy paradigms. The strong performance of OPSFT indicates that the generalization advantage of on-policy paradigms can be transferred to SFT through the parameter update direction. Once such a direction is identified, even SFT can generalize with its updates constrained to this direction. This finding offers two practical benefits by combining the strong generalization of on-policy paradigms with the advantages of SFT, including the high training efficiency and ability to leverage high-quality trajectories. For efficiency, we identify update directions that support strong generalization using a few on-policy training steps, and subsequently apply OPSFT to achieve high training efficiency. For leveraging high-quality trajectories, OPSFT can utilize these trajectories to continue improving a post-trained model along its update direction without disrupting the ability learned from on-policy training.

UCAS ucas
·
Sep 28 2

Discovering Interpretable Directions in the Semantic Latent Space of Diffusion Models

Denoising Diffusion Models (DDMs) have emerged as a strong competitor to Generative Adversarial Networks (GANs). However, despite their widespread use in image synthesis and editing applications, their latent space is still not as well understood. Recently, a semantic latent space for DDMs, coined `h-space', was shown to facilitate semantic image editing in a way reminiscent of GANs. The h-space is comprised of the bottleneck activations in the DDM's denoiser across all timesteps of the diffusion process. In this paper, we explore the properties of h-space and propose several novel methods for finding meaningful semantic directions within it. We start by studying unsupervised methods for revealing interpretable semantic directions in pretrained DDMs. Specifically, we show that global latent directions emerge as the principal components in the latent space. Additionally, we provide a novel method for discovering image-specific semantic directions by spectral analysis of the Jacobian of the denoiser w.r.t. the latent code. Next, we extend the analysis by finding directions in a supervised fashion in unconditional DDMs. We demonstrate how such directions can be found by relying on either a labeled data set of real images or by annotating generated samples with a domain-specific attribute classifier. We further show how to semantically disentangle the found direction by simple linear projection. Our approaches are applicable without requiring any architectural modifications, text-based guidance, CLIP-based optimization, or model fine-tuning.

  • 4 authors
·
Mar 20, 2023

Zero-Shot Vision-and-Language Navigation with Collision Mitigation in Continuous Environment

We propose the zero-shot Vision-and-Language Navigation with Collision Mitigation (VLN-CM), which takes these considerations. VLN-CM is composed of four modules and predicts the direction and distance of the next movement at each step. We utilize large foundation models for each modules. To select the direction, we use the Attention Spot Predictor (ASP), View Selector (VS), and Progress Monitor (PM). The ASP employs a Large Language Model (e.g. ChatGPT) to split navigation instructions into attention spots, which are objects or scenes at the location to move to (e.g. a yellow door). The VS selects from panorama images provided at 30-degree intervals the one that includes the attention spot, using CLIP similarity. We then choose the angle of the selected image as the direction to move in. The PM uses a rule-based approach to decide which attention spot to focus on next, among multiple spots derived from the instructions. If the similarity between the current attention spot and the visual observations decreases consecutively at each step, the PM determines that the agent has passed the current spot and moves on to the next one. For selecting the distance to move, we employed the Open Map Predictor (OMP). The OMP uses panorama depth information to predict an occupancy mask. We then selected a collision-free distance in the predicted direction based on the occupancy mask. We evaluated our method using the validation data of VLN-CE. Our approach showed better performance than several baseline methods, and the OPM was effective in mitigating collisions for the agent.

  • 4 authors
·
Oct 7, 2024

Surgical Refusal Ablation: Disentangling Safety from Intelligence via Concept-Guided Spectral Cleaning

Safety-aligned language models systematically refuse harmful requests. While activation steering can modulate refusal, ablating the raw "refusal vector" calculated from contrastive harmful and harmless prompts often causes collateral damage and distribution drift. We argue this degradation occurs because the raw vector is polysemantic, entangling the refusal signal with core capability circuits and linguistic style. We introduce Surgical Refusal Ablation (SRA) to distill these steering directions. SRA constructs a registry of independent Concept Atoms representing protected capabilities and stylistic confounds, then uses ridge-regularized spectral residualization to orthogonalize the refusal vector against these directions. This yields a clean refusal direction that targets refusal-relevant structure while minimizing disruption to the model's semantic geometry. Across five models (Qwen3-VL and Ministral series), SRA achieves deep refusal reduction (0-2%) with negligible perplexity impact on Wikitext-2 (mean delta PPL approx. 0.02) and minimal distribution drift. Notably, standard ablation on Qwen3-VL-4B induces severe drift (first-token KL = 2.088), whereas SRA maintains the original distribution (KL = 0.044) while achieving the same 0% refusal rate. Using teacher-forced perplexity on GSM8K and MBPP as a high-resolution capability proxy, we show SRA preserves math and code distributions. These results suggest that common "model damage" is often "Ghost Noise," defined as the spectral bleeding of the dirty refusal direction into capability subspaces.

  • 1 authors
·
Jan 12

Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails

Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

  • 1 authors
·
Mar 18

The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models

Auxiliary signal pathways in VLMs are routinely fitted with learnable gates so the optimiser can decide how much of the signal to admit. We find that the optimiser almost always decides on zero: across five injection designs, every gated pathway becomes behaviourally closed, with accuracy invariant to ablating the pathway at inference even when the gate parameter would nominally pass 30-45% of the signal. We attribute this suppression phenomenon to two regimes, a dead-gradient regime formalised through the caption-invariance of image-derived signals, and a negative-utility regime in which the auxiliary signal actively hurts the loss. Rather than fight suppression, we exploit it: we regularise LoRA fine-tuning with geometric auxiliary losses from hyperbolic visual relational graphs (IoA-driven entailment cones and angular repulsion on the Lorentz manifold), coupled only through the forward pass at training time and dropped at inference. Disaggregating GQA by question type exposes a clean dissociation. Three configurations without geometric losses at inference lose 2.85-3.39pp on relational questions while gaining ~1pp on attribute questions; a fourth that trains with the losses but infers through a soft prompt loses 5.14pp on rel for only +0.23pp on attr, so training-time regularisation alone does not protect relational accuracy without a geometric inference pathway. Configurations that keep the geometric pathway at inference preserve vanilla-level relational accuracy and match the attribute gain. Out of distribution on VSR, the RMS-prefix recipe preserves the spatial signal; stripping the geometric losses (G2) collapses VSR by 4.6pp, isolating them as the OOD source. A secondary result: embedding-norm alignment is necessary for generation-safe prefix injection, and learnable gates should be replaced with fixed, non-optional injection at matched scales.

  • 5 authors
·
Jul 24

A Gravitational Interpretation of Fine-Tuning Reversion

Fine-tuning on harmless data can partially undo behaviors acquired earlier in training. Safety can erode under benign post-alignment updates, unlearned capabilities can re-emerge, latent traits can transfer through apparently unrelated supervision, and related post-alignment fragility appears in other generative settings. We argue these phenomena are usefully viewed through a common training-history lens. Our hypothesis is geometric: large early training phases create dominant behavioral manifolds, while later alignment or specialization phases are shallower displacements from them. Subsequent fine-tuning can therefore inherit a persistent reversion component pointing back toward a witness of the dominant manifold. We call this the gravitational interpretation of fine-tuning reversion. Across our main settings, representational drift rapidly acquires a component along a history-defined reversion direction (v_rev). In our main track, alignment with v_rev rises from cos = 0.429 +/- 0.052 after the first update to 0.647 +/- 0.021 by step 20. Across 24 run-step pairs, every observed alignment exceeds the p99 of an isotropic activation-space null. We demonstrate that selectively blocking motion along v_rev changes the final alignment at T=100 from 0.648 +/- 0.009 to -0.211 +/- 0.021 and reduces harmfulness from 19.0% +/- 4.0% to 8.5% +/- 1.5% with little task cost. These results support v_rev as a causally relevant mediator of early post-alignment reversion in our setup. Importantly, we do not claim that v_rev is the unique safety direction, nor that the dominant manifold is directly observed; rather, we identify a robust, history-defined direction that explains and partially controls early reversion dynamics.

Fixing a Model That Learned Worse Cancer Means Lower Risk: Monotonic Constraints in Bladder Cancer Recurrence Prediction

Background and Objective: Clinicians expect recurrence risk to climb with cancer severity. In a UK multicentre trial, an unconstrained XGBoost model learnt that higher tumour stage and carcinoma in situ predicted lower recurrence risk, and discrimination, calibration, and SHAP were all blind to it. We developed a counterfactual testing framework to detect this inversion and a monotonic-constraint framework to remove it without hurting performance. Methods: BOXIT enrolled 472 patients with protocol-mandated cystoscopy across 51 UK sites (2007-2012); 435 had at least two years' follow-up (153 recurrences, 35.2%). We developed a counterfactual direction test and a monotonic-constraint correction, with constraint directions drawn from the EORTC and EAU risk systems, and evaluated both against unconstrained XGBoost and logistic regression on 18 predictors (seven directed) over 50 cross-validation folds. The test worsened each patient on one directed feature at a time to check whether risk fell; SHAP direction and calibration were also assessed. Key Findings and Limitations: Tumour stage and carcinoma in situ were associated with lower recurrence, opposite to medical intuition; the unconstrained model reversed carcinoma in situ counterfactuals in 90.2% of cases and stage in 74.3%. Discrimination (ΔAUC 0.005, p=0.47), calibration, and SHAP magnitude were all blind to the inversion. Monotonic constraints eliminated every violation at no cost to discrimination (0.723 vs 0.718) and outperformed EORTC (p=8.9e-16). Limitations: single trial, internal-external validation only. Conclusions and Clinical Implications: A model that had learned this inversion passed every conventional check. A counterfactual direction test, run as a single refit with pre-specified monotonic constraints, catches this failure at no cost to performance and should be routine before clinical deployment.

  • 6 authors
·
Sep 30

Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations

Alignment tuning is meant to make harmful-request refusal robust, yet this safety behavior can be erased by a small set of benign fine-tuning examples. This is a deployment risk for open-weight models because a checkpoint can pass refusal tests at release time and later lose refusal under low-cost downstream fine-tuning. Prior work has established these refusal failures, but existing studies do not show how to detect this fragility in the aligned model itself before an attack or fine-tuning intervention is run. We introduce Skin-Deep, a geometric diagnostic that detects alignment fragility directly from the aligned model's hidden-state activations before such an intervention is run and compresses the layer-wise safety geometry into a single scalar, the Geometric Fragility Score (GFS). Applied to twenty-one instruction-tuned models spanning six alignment recipes and 3B--32B parameters, Skin-Deep reveals a recurring low-rank safety subspace across model families. Direction ablations show that removing directions in this subspace weakens harmful-request refusal, providing causal evidence that the recovered geometry underlies refusal behavior. Crucially, GFS identifies, before any fine-tuning, the initially safe model that retains the most refusal after small-scale LoRA fine-tuning. These results establish GFS as a practical pre-deployment diagnostic for flagging fragile refusal behavior without running an attack.

  • 8 authors
·
Jun 20

The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models

Large language models can represent a variety of personas but typically default to a helpful Assistant identity cultivated during post-training. We investigate the structure of the space of model personas by extracting activation directions corresponding to diverse character archetypes. Across several different models, we find that the leading component of this persona space is an "Assistant Axis," which captures the extent to which a model is operating in its default Assistant mode. Steering towards the Assistant direction reinforces helpful and harmless behavior; steering away increases the model's tendency to identify as other entities. Moreover, steering away with more extreme values often induces a mystical, theatrical speaking style. We find this axis is also present in pre-trained models, where it primarily promotes helpful human archetypes like consultants and coaches and inhibits spiritual ones. Measuring deviations along the Assistant Axis predicts "persona drift," a phenomenon where models slip into exhibiting harmful or bizarre behaviors that are uncharacteristic of their typical persona. We find that persona drift is often driven by conversations demanding meta-reflection on the model's processes or featuring emotionally vulnerable users. We show that restricting activations to a fixed region along the Assistant Axis can stabilize model behavior in these scenarios -- and also in the face of adversarial persona-based jailbreaks. Our results suggest that post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent persona.

  • 5 authors
·
Jan 15 2

Anchoring Refusal Direction: Mitigating Safety Risks in Tuning via Projection Constraint

Instruction Fine-Tuning (IFT) has been widely adopted as an effective post-training strategy to enhance various abilities of Large Language Models (LLMs). However, prior studies have shown that IFT can significantly compromise LLMs' safety, particularly their ability to refuse malicious instructions, raising significant concerns. Recent research into the internal mechanisms of LLMs has identified the refusal direction (r-direction) in the hidden states, which plays a pivotal role in governing refusal behavior. Building on this insight, our study reveals that the r-direction tends to drift during training, which we identify as one of the causes of the associated safety risks. To mitigate such drift, our proposed ProCon method introduces a projection-constrained loss term that regularizes the projection magnitude of each training sample's hidden state onto the r-direction. Our initial analysis shows that applying an appropriate constraint can effectively mitigate the refusal direction drift and associated safety risks, but remains limited by overall performance barriers. To overcome this barrier, informed by our observation of early-stage sharp drift and a data-driven perspective, we introduce a warm-up strategy that emphasizes early-stage strong constraints and broaden the data distribution to strengthen constraint signals, leading to an enhanced ProCon method. Experimental results under various datasets, scenarios, and LLMs demonstrate that our method can significantly mitigate safety risks posed by IFT while preserving task performance gains. Even compared with strong baselines, our method consistently delivers superior overall performance. Crucially, our analysis indicates that ProCon can contribute to stabilizing the r-direction during training, while such an interpretability-driven exploration of LLMs' internal mechanisms lays a solid foundation for future safety research.

  • 9 authors
·
Sep 7, 2025

Detection of the Schwarzschild precession in the orbit of the star S2 near the Galactic centre massive black hole

The star S2 orbiting the compact radio source Sgr A* is a precision probe of the gravitational field around the closest massive black hole (candidate). Over the last 2.7 decades we have monitored the star's radial velocity and motion on the sky, mainly with the SINFONI and NACO adaptive optics (AO) instruments on the ESO VLT, and since 2017, with the four-telescope interferometric beam combiner instrument GRAVITY. In this paper we report the first detection of the General Relativity (GR) Schwarzschild Precession (SP) in S2's orbit. Owing to its highly elliptical orbit (e = 0.88), S2's SP is mainly a kink between the pre-and post-pericentre directions of motion ~ +- 1 year around pericentre passage, relative to the corresponding Kepler orbit. The superb 2017-2019 astrometry of GRAVITY defines the pericentre passage and outgoing direction. The incoming direction is anchored by 118 NACO-AO measurements of S2's position in the infrared reference frame, with an additional 75 direct measurements of the S2-Sgr A* separation during bright states ('flares') of Sgr A*. Our 14-parameter model fits for the distance, central mass, the position and motion of the reference frame of the AO astrometry relative to the mass, the six parameters of the orbit, as well as a dimensionless parameter f_SP for the SP (f_SP = 0 for Newton and 1 for GR). From data up to the end of 2019 we robustly detect the SP of S2, del phi = 12' per orbital period. From posterior fitting and MCMC Bayesian analysis with different weighting schemes and bootstrapping we find f_SP = 1.10 +- 0.19. The S2 data are fully consistent with GR. Any extended mass inside S2's orbit cannot exceed ~ 0.1% of the central mass. Any compact third mass inside the central arcsecond must be less than about 1000 M_sun.

  • 56 authors
·
Apr 14, 2020

The Answer-Basin Representation Hypothesis: We Are Not Probing or Steering Concepts

Linear probing and activation steering use linear directions to predict and control behavior-level concepts such as correctness, safety, and social bias in question answering. We call these behavior-level concept directions. Drawing on the Linear Representation Hypothesis (LRH) and circuit studies, these directions are often interpreted as concept representations. Yet controlled evidence for linear concept structure and local mechanisms does not establish that these directions represent the intended behavioral concepts, leaving probing and steering without a unified theoretical account. We propose the Answer-Basin Representation Hypothesis (ABRH): the model's own answer measure organizes the linear structure of these directions. For each question, all continuations yielding the same answer form an answer basin, whose mass is their total probability; these masses define the model's answer distribution. ABRH posits that its concentration before generation and the relative mass of each answer after generation are represented along linear directions shared across questions. Experiments span four Qwen2.5 and Gemma-3 models on correctness, social bias, and safety tasks. Decoupling concept labels from basin mass shows that probing and steering exhibit concept-consistent effects when labels align with mass orderings, weaken as mass gaps shrink, and reverse under conflict. Probes trained solely on mass orderings among wrong answers still select correct answers; a fixed steering vector can instead favor wrong answers when labels conflict with mass orderings. These results support an answer-measure account of behavior-level concept directions and of when probing and steering succeed, fail, or reverse.

  • 8 authors
·
Oct 1

Mechanisms of Introspective Awareness

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

  • 6 authors
·
Apr 12

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

A fine-tuned Vision-Language-Action (VLA) policy can easily place the alphabet soup in the basket on command, just as it saw during fine-tuning. However, if the basket is shifted five centimeters to the left, it drops the soup off the table. We developed an interpretability pipeline to investigate what such a policy is computing. Injecting task A's activations into the action expert while the policy looks at task B's scene makes pi0.5 execute task A's motor trajectory in 99.6\% of episodes (n{=}1{,}968). The injected program is bound to absolute workspace coordinates and ignores the visible scene, which provides a mechanistic account of the perturbation brittleness that concurrent benchmarks report. The binding holds on a physical UR10e: directed replay redirects the commanded task in 26/26 rollouts in one direction and 18/26 in the other, with no-injection and random-direction controls at the floor, and once the source object is moved 10--15\,cm the arm grasps and releases at the recorded coordinates in 50/50 trials, picking up the wrong block when one sits there and closing on empty air when none does. Applying the same interventions to six VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA over 420{,}000+ rollouts yields three further findings: every architecture encodes the instruction, and four of the five weight it below vision whenever the scene alone identifies the goal; under the standard replacement write-back the SAE pooling that preserves behavior is architecture-specific (pi0.5 per-token, SmolVLA indifferent), and goal selection and motor generation separate into VLM and expert pathways wherever the two are distinguishable. We release Action Atlas (https://action-atlas.com), an interactive platform covering all 520+ trained SAEs, 79 identified concepts, and the scored UR10e rollouts. Code: https://github.com/CWRU-AISM/action-atlas

  • 3 authors
·
Oct 19 2

Causal Routing for Unlearning

LLMs cannot forget the way we delete a file. Strangely, we are asked to remove something that was never put anywhere in particular. What the model took from a piece of text is now smeared across billions of weights. Existing methods rewrite all of them to change one thing, and none of them say which part produced that change. To address this, we introduce Causal Routing for Unlearning (CRU) by asking where the concept is expressed in the model and suppressing only that part. One untrained forward pass over the forget set ranks neurons by how their activations vary. Then, small routing modules on those neurons gate and suppress only the concepts that need to be forgotten. In CRU, the base model is frozen, and any change in behavior is caused only by the gated neurons; hence, why the routing is causal. Due to our parameter efficiency (only ~0.01% as many parameters as the base model), unlearning a concept costs 14 GiB, whereas the baselines require 71 GiB. On TOFU, CRU is indistinguishable from the retained model (p > 0.05, KS test) and is never Pareto-dominated, whereas every compared baseline matches its forgetting on the larger-forget batches only by collapsing utility. On RWKU, it achieves an adversarial-probe recall of 0.052, compared to 0.250 for the strongest baseline, meaning the knowledge is gone, not merely harder to reach. Thus, deciding on the intervention at query time, rather than fixing it beforehand, is the axis along which we argue that unlearning should proceed.

  • 7 authors
·
Sep 27

Broken Symmetry in LLM Refusal: Answer Release Is More Local Than Refusal Restoration

When a language model refuses to answer a prompt, it is unclear whether the correct answer is erased from its internal representations, or merely suppressed at the output layer. We investigate this mechanism using a controlled withhold setting, which yields perfectly matched answering and refusal trajectories for bidirectional activation patching. We uncover a causal asymmetry in intervention locality under matched causal interventions, which we term broken symmetry. Even when a model generates a clean refusal, the correct answer remains linearly recoverable from its hidden states. Furthermore, releasing this withheld answer is a highly local operation, requiring only a single-position patch. Conversely, the reverse operation is not equally local: reimposing suppression requires broader interventions across multiple positions, and assembling a coherent refusal sequence is more difficult still. We further demonstrate that while an average answer-to-refusal displacement vector marks the geometric difference between these states, it fails to act as a reliable, reversible linear control toggle between behaviours. Taken together, our findings show that refusal does not function as a simple symmetric switch. For safety and auditing, this implies that probe recoverability can overestimate true behavioural control, and locating refusal-relevant directions does not reliably grant the ability to steer a model from answering to coherent refusal.

  • 5 authors
·
Aug 15

Stochastic Subgradient Descent at Sharply Repulsive Points: A Bounded Noise Counterexample and Gaussian Noise Avoidance

Motivated by the open question of Bianchi, Hachem, and Schechtman (Bianchi et al., 2024, Remark 4), we show that in the non-weakly convex definable setting, omnidirectional noise with conditional fourth-moment control does not suffice for universal almost-sure avoidance of sharply repulsive critical points. We demonstrate this failure by constructing a globally Lipschitz, coercive, definable objective that is not weakly convex. For this objective and its sufficiently small linear perturbations, stochastic subgradient descent with independent noise uniformly distributed on a ball in R^2 converges with probability 1 to a nonminimal sharply repulsive critical point from every initial point in a specified ball. A two-cycle argument establishes uniform bounds on the rescaled iterates, yielding convergence to the sharply repulsive critical point. In contrast, for locally Lipschitz definable objectives, Gaussian noise ensures simultaneous almost-sure avoidance of any finite collection of sharply repulsive points. Through a construction of difference quotient functions, we show that convergence to any such point with positive probability would produce a stationary distribution with expected descent 0, whereas a global positive lower bound on the Fréchet subgradient norms of the limiting functions forces strictly positive expected descent under the same distribution. This contradiction proves avoidance, with the everywhere positive Gaussian density playing a key role.

  • 2 authors
·
Oct 5