Selfment: Multifunctional Self-Supervision in AI

Updated 5 July 2026

Selfment is a diverse set of self-related computational frameworks covering self-supervised segmentation, AI-native memory, self-attribution, and self-improvement methods.
In computer vision, Selfment employs techniques like Normalized Cuts and Iterative Patch Optimization to achieve state-of-the-art unsupervised saliency detection with minimal human guidance.
Beyond vision, Selfment influences AI self-reflection and digital self-representation through systems such as memory aids, introspective language models, and self-clone chatbots.

Selfment denotes several distinct but related constructs in recent research. In the narrowest and explicit sense, it is the name of a fully self-supervised foreground segmentation framework that learns object masks directly from raw images without human labels, pretrained segmentation models, or post-processing (You et al., 27 Feb 2026). In broader usage, the term has been read as covering AI-mediated self-extension, self-attribution of mentality in LLMs, inference-time self-improvement, computational profiling of self-aspects in text, emotionally self-evolving dialogue systems, context-aware reflective journaling, and self-clone chatbots. This suggests that Selfment is not a single settled technical term, but a family of self-related computational ideas centered on representation, extension, reflection, or improvement of a self-like substrate.

1. Scope and semantic range

The exact term appears most crisply in computer vision, where Selfment names a segmentation method. Elsewhere, closely related papers either explicitly reinterpret the term or state that they do not use it while nevertheless addressing adjacent constructs such as “Second Me,” “self-attributions of mentality,” “self-improvement without training,” “Self-aspects,” “emotional self-evolution,” or “self-clones” (Wei et al., 11 Mar 2025, Kim et al., 30 Mar 2026, Li et al., 20 Mar 2026, Caporusso et al., 17 Jul 2025, Zhang et al., 20 Apr 2026, Nepal et al., 2024, Shirvani et al., 8 Sep 2025, Chen et al., 2024).

Research line	Term usage	Representative source
Foreground segmentation	Exact title usage	(You et al., 27 Feb 2026)
AI-native memory	Read through the lens of Selfment	(Wei et al., 11 Mar 2025)
LLM mentality attribution	Selfment as self-ascribed mentality	(Kim et al., 30 Mar 2026)
Inference-time self-improvement	A subtype of Selfment	(Li et al., 20 Mar 2026)
Self-aspect profiling in text	Closest construct rather than exact term	(Caporusso et al., 17 Jul 2025)
Emotional self-evolution	Relevant if Selfment means self-guided affective development	(Zhang et al., 20 Apr 2026)
Reflective journaling and self-clones	Selfment as AI-mediated self-understanding or self-conversation	(Nepal et al., 2024, Shirvani et al., 8 Sep 2025)

Across these usages, the shared motif is not a uniform ontology of selfhood. Rather, the common thread is that the system either extracts a self-relevant structure from data, represents a user’s perspective, attributes mind or selfhood to a model, or improves itself through internal verification, memory, or recursive feedback. The literature therefore supports a plural rather than singular definition.

2. Selfment as fully self-supervised segmentation

In computer vision, Selfment is a fully self-supervised foreground segmentation framework introduced in “Learning Accurate Segmentation Purely from Self-Supervision.” Its pipeline has three stages: a patch affinity graph plus Normalized Cut (NCut) for an initial coarse foreground/background split, Iterative Patch Optimization (IPO) for feature-space refinement, and self-supervised training of a lightweight segmentation head from the refined masks. The main frozen backbone is DINOv3-7B. The graph uses thresholded inner-product affinity with $\tau = 0.2$ and a small $\epsilon$ to maintain connectivity; NCut solves $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ and thresholds the Fiedler vector at its mean; IPO then repeatedly reassigns patches by similarity to foreground and background centroids, enforces orientation consistency, and runs for $T=20$ , though the paper states that IPO often converges visually in around 10 iterations (You et al., 27 Feb 2026).

The trainable segmenter is intentionally small: a two-layer projection head followed by a binary classifier, with 0.54M trainable parameters and 1.08M FLOPs per forward pass. Training uses pseudo-labels from IPO and combines contrastive loss, Soft Dice loss, and binary cross-entropy loss, with $\lambda_{\text{con}} = 0.1$ , $\lambda_{\text{Dice}} = 1.0$ , and $\lambda_{\text{BCE}} = 1.0$ . The training corpus is 1,000 images randomly sampled from the DUTS training set; optimization uses Adam, learning rate $1 \times 10^{-3}$ , 3 epochs, and 8 NVIDIA A100 GPUs (80 GB) with cached backbone features.

Empirically, the paper reports state-of-the-art unsupervised saliency performance. At $1280 \times 1280$ , Selfment achieves $F_{\max}=95.9$ on ECSSD, 86.4 on DUTS, 94.4 on HKUIS, and 91.7 on PASCAL-S. The abstract highlights gains in $\epsilon$ 0 over previous unsupervised saliency detection methods of $\epsilon$ 1 on ECSSD, $\epsilon$ 2 on HKUIS, and $\epsilon$ 3 on PASCAL-S; the introduction also mentions $\epsilon$ 4 on DUTS. Zero-shot transfer to camouflaged object detection is unusually strong, including $\epsilon$ 5 $\epsilon$ 6 on CHAMELEON and $\epsilon$ 7 $\epsilon$ 8 on CAMO, with the paper noting that Selfment outperforms all existing unsupervised approaches and even rivals some fully supervised methods. The main reported failure mode is that IPO can absorb objects or regions that are semantically similar to the true foreground, because refinement is driven by patch-level feature similarity.

3. Selfment as AI-native memory and digital self-representation

In personalized AI systems, Selfment has been used as an interpretive lens for “AI-native Memory 2.0: Second Me.” SECOND ME is presented as an “intelligent, persistent memory offload system” and “context provider” whose function is to stand between a person and the digital world, carrying, organizing, and applying that person’s memory so that the person does not have to repeatedly restate facts, preferences, history, or project context. The paper explicitly distinguishes this from browser-stored credentials, autofill, and unified authentication systems, which are described as static repositories rather than context-aware mediators. It also explicitly positions SECOND ME “as a context provider aligned with the user’s perspective, rather than a task executor,” while still attributing to it active capabilities such as autonomously generating context-aware responses, prefilling required information, facilitating communication with external systems, and acting as a proxy in multi-user settings (Wei et al., 11 Mar 2025).

The architecture inherits the earlier Large Personal Model framework and uses three layers: L0: Raw Data Layer, L1: Natural Language Memory Layer, and L2: AI-Native Memory Layer. SECOND ME adds an inner loop for seamless layer integration and an outer loop in which “LLMs and internet resources” operate “under Second Me’s guidance.” User-uploaded data include notes and todos; notes may be “a document, an audio file, a website, a picture or multi-modal information which includes title and content.” The synthesis pipeline categorizes and summarizes uploaded multimodal information, uses indexing and extraction tools “such as GraphRAG” to identify “entities, relations and communities,” generates a user biography and status description, and then synthesizes trainable QA pairs by question generation and answering. For personalization, the base model is Qwen2.5-7B-Instruct, adapted using PEFT and then refined with DPO, with preference pairs constituting about 20% of the total SFT training data. The paper states that each user’s data “remains isolated.”

The operational tasks make the self-representational intent concrete. Memory QA covers retrieval, concept understanding, behavior prediction, and recommendation. Context Enhance enriches a user’s initial query by injecting details from related notes and to-do items. Context Critic critiques an “expert response” from the user’s perspective by identifying omissions or misalignment relative to the user’s memory and priorities. In the multi-agent appendix, a trained personal model collaborates with an expert model such as GPT-4o and can later function as a proxy in shared environments. Evaluation is modest but specific: one internal staff user had 132 notes and 62 todos, producing nearly 7k instruction pairs; the evaluation set contained 60 first-person memory queries, 60 third-person memory queries, 60 context enhance samples, and 60 context critic samples. Under Strong COT, scores were 0.91 on Memory (Self), 0.71 on Memory (Third-Party), 0.75 on Context Enhance, and 0.85 on Context Critic; adding DPO raised these to 0.96, 0.76, 0.85, and 0.86. The paper also reports human evaluation of 0.95 for Strong COT without DPO and “close to 1” with DPO. The stated limitations include reliance on single-turn training, limited large-scale evaluation, and unresolved problems in multimodal integration and “real-time synchronization with human thought.”

4. Selfment as self-attribution, introspection, and proto-self-consciousness in LLMs

One line of work defines selfment in LLMs as the tendency to ascribe mentality to themselves: to say that they are conscious, sentient, agents, persons, or possess souls. “Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs” operationalizes this with direct self-report-style prompts on a 0–10 scale, alongside anthropomorphism questions about robots, computers, television sets, cars, conversational chatbots, and animals. The key result is dissociation: safety fine-tuning suppresses self-attributions of consciousness, sentience, agency, soul, and personhood, but removing the learned safety direction does not measurably improve Theory of Mind on MoToMQA, HI-ToM, or SimpleToM. With chain-of-thought, pooled jailbreak effects on self-attributions were $\epsilon$ 9 for agency, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 0 for consciousness, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 1 for sentience, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 2 for personhood, and $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 3 for soul, all with $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 4. By contrast, pooled main effects of jailbreaking on ToM benchmarks were non-significant. Mechanistically, instruction tuning rotates the mind-attribution direction toward anti-alignment with safety, with average similarity shift $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 5, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 6, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 7, while the ToM direction remains essentially unchanged relative to safety, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 8, $(D-A)\mathbf{x}=\lambda D\mathbf{x}$ 9, $T=20$ 0. The same paper argues that models show an “AI-centric bias”: after jailbreaking, mind-attribution rises above human means for chatbots, technology, and non-animal natural entities, but remains below human baselines for non-human animals (Kim et al., 30 Mar 2026).

A related but distinct line studies functional self-consciousness. “From Imitation to Introspection: Probing Self-Consciousness in LLMs” defines a practical notion of self-consciousness as the conjunction of C1 consciousness, meaning global availability of information for recall, reporting, and decision-making, and C2 consciousness, meaning self-monitoring of one’s own computations. It refines ten core concepts: situational awareness, sequential planning, belief, intention, self reflection, self improve, harm, known knowns, known unknowns, and deception. The evaluation spans 10 models and 11,097 binary-classification samples. The top three models are Claude3.5-Sonnet, GPT-4o, and GPT-o1 preview, and the paper reports that the first three exceed the 50% random baseline by 26.5%, 22.6%, and 22.4%, respectively, while 60.0% of the models struggle to exceed 70.0%. The weakest concept is known knowns; the strongest is intention. Linear probing over attention heads shows “a discernible representation of certain concepts within their internal mechanisms,” but activation-level manipulation via Mass Mean Shift and Probe Weight Direction is “hard to manipulate positively at the current stage.” By contrast, LoRA fine-tuning on belief and sequential planning improves both accuracy and deeper-layer activation, especially in the 30th–32nd layers (Chen et al., 2024).

Taken together, these papers support a sharp distinction between at least three self-related phenomena in LLMs: self-ascribed mentality, benchmarked self-monitoring or introspection-like competence, and social-cognitive reasoning about others. They do not collapse into one another.

5. Selfment as self-improvement and self-evolution

In inference-time reasoning research, Selfment has been explicitly linked to self-verification plus targeted regeneration. “A Training-Free Regeneration Paradigm: Contrastive Reflection Memory Guided Self-Verification and Self-Improvement” proposes Reflection Memory (RM), an offline-curated repository of positive entries and negative/contrastive entries containing a cleaned wrong reasoning trace, a reflection on the mistake, a corrected solution, and a distilled principle. Retrieval uses Contriever for coarse top- $T=20$ 1 retrieval with $T=20$ 2, then MPNet reranking, keeping $T=20$ 3 entries. At test time, the system generates an initial answer, performs RM-guided self-verification, adds an entropy-based recovery step using $T=20$ 4, and, if needed, performs a single RM-guided regeneration from scratch rather than iterative editing. The central claim is that regeneration breaks out of “faulty reasoning loops.” On GPT-3.5 / GSM-Hard, the full method scores 64.00, whereas replacing regeneration with rectification yields 53.60. In flip analysis under oracle verification on GSM-Hard, RM-Regen converts 46.55% of initial wrong answers to correct, whereas Reflexion (5 iters) converts 10.47%; without oracle verification, RM-Regen achieves 38.63% wrong→correct versus 8.77% for ST-CoT (4 iters), while correct→wrong damage is 4.48% versus 13.95%. Across nine benchmarks and three base models—GPT-3.5-0125, Llama-3.1-8B-Instruct, and Gemma-2-9B-it—the paper reports that RM-Regen outperforms prior methods while maintaining low computational cost (Li et al., 20 Mar 2026).

A separate affective line frames self-improvement as emotional self-evolution. “SELF-EMO: Emotional Self-Evolution from Recognition to Consistent Expression” builds a three-stage pipeline over dialogue context $T=20$ 5 and personality information $T=20$ 6:

$T=20$ 7

Here the model predicts the other speaker’s emotion, generates its own self-emotion, and then produces a response. The system uses role-based self-play, a replay-buffer flywheel, a smoothed weighted IoU reward with a $T=20$ 8 smoothing term, and SELF-GRPO, which adds a group-level consistency reward over the top-3 consensus emotions. On IEMOCAP, MELD, and EmoryNLP, under a dataset-unified setting, SELF-EMO improves average accuracy by +6.33% on Qwen3-4B and +8.54% on Qwen3-8B. The corresponding averages are 52.96 → 59.29 for Qwen3-4B and 53.43 → 61.97 for Qwen3-8B. Ablation shows that removing SELF-GRPO drops average performance by 3.14, and training with GRPO using only original data drops it by 5.18. The paper’s framing is that better emotion prediction leads to more consistent emotional responses, so self-improvement proceeds by recognition, self-emotion generation, expression, filtering, and replay (Zhang et al., 20 Apr 2026).

These two programs differ in substrate—symbolic or reasoning traces in one case, emotional dialogue trajectories in the other—but converge on a similar principle: self-improvement is more effective when internal verification is coupled to structured memory or structured intermediate self-state, rather than treated as free-form post hoc editing.

6. Selfment as self-profiling, reflective scaffolding, and self-conversation

A text-analytic interpretation of Selfment appears in “A Computational Framework to Identify Self-Aspects in Text.” That paper does not use the exact term, but proposes an ontology for identifying the presence and mode of Self-aspects in text, treating the Self as a multifaceted construct reflected in language. The ontology is explicitly hierarchical—Self-aspect $T=20$ 9 element $\lambda_{\text{con}} = 0.1$ 0 mode—with examples including Minimal Self, Narrative Self, Self as Agent, Bodily Self, and Social Self. The task is instance-based classification of whether a given sentence, utterance, or paragraph expresses a Self-aspect and, eventually, in what mode. The proposal intends to compare discriminative models, LLMs, and embedding-based retrieval systems along four criteria: interpretability, ground-truth basis, accuracy, and computational efficiency. Only a pilot is fully reported: binary classification of Social Self on 1,473 diary sub-entries from Lemotif, with Cohen’s Kappa = 0.80 between human annotators and 0.84–0.89 between human and model annotators. Under 10-fold cross-validation, the best model is SVM on LIWC features, reaching macro precision 0.83, macro recall 0.83, and macro F1 0.83 (Caporusso et al., 17 Jul 2025).

A behavior-aware reflective interpretation appears in “MindScape Study: Integrating LLM and Behavioral Sensing for Personalized AI-Driven Journaling Experiences.” MindScape is an 8-week exploratory study with 20 college students using an Android journaling app that combines passively collected behavioral signals with GPT-4 prompts. Signals include social interaction, sleep, digital habits, physical activity, and location; features are aggregated over hourly, daily, and weekly timescales, with a cron job running every 30 minutes and weekday trends compared against a 30-day historical average. The system produced 661 total journal entries, of which 533 were contextual and 128 generic. After the 6-week contextual phase, reported changes included positive affect +7.15%, negative affect -10.60%, loneliness -6.47%, mindfulness +6.76%, self-reflection +5.80%, and insight +7.57%. In weekly mixed-effects analysis, PHQ-4 decreased with $\lambda_{\text{con}} = 0.1$ 1, $\lambda_{\text{con}} = 0.1$ 2, $\lambda_{\text{con}} = 0.1$ 3, and self-reflection increased with $\lambda_{\text{con}} = 0.1$ 4, $\lambda_{\text{con}} = 0.1$ 5, $\lambda_{\text{con}} = 0.1$ 6. Contextual journals contained more pronouns, cognition words, social processes, social behavior, and social referents than generic journals, and 85% reported that contextual prompts sometimes, often, or always led to more in-depth reflection than generic prompts. The main limitations were sensing inaccuracies, repetitive prompts, privacy concerns, and the absence of journal-content adaptation (Nepal et al., 2024).

A more explicitly dialogic version appears in “Talking to an AI Mirror: Designing Self-Clone Chatbots for Enhanced Engagement in Digital Mental Health Support.” That study introduces self-clone chatbots derived from how participants support a distressed friend in a preliminary “A Friend in Need” conversation. The main chatbot conditions were a generic counselor baseline (BL), a self-clone without Social Support Prompting (SCX), and a self-clone with Social Support Prompting (SCS). The primary experiment had N = 180 participants, balanced across the three conditions. A crucial finding is that self-clone believability was bimodal, with Hartigan’s dip test $\lambda_{\text{con}} = 0.1$ 7 and a Gaussian mixture threshold of 2.81. Engagement correlated strongly with believability, $\lambda_{\text{con}} = 0.1$ 8. When only high-believability self-clones were compared with baseline, the study found significantly higher engagement: overall engagement $\lambda_{\text{con}} = 0.1$ 9, cognitive engagement $\lambda_{\text{Dice}} = 1.0$ 0, and emotional engagement $\lambda_{\text{Dice}} = 1.0$ 1. A follow-up roughly 10 weeks later found $\lambda_{\text{Dice}} = 1.0$ 2 for engagement difference across waves, which the authors interpret as little evidence of a novelty-only effect. The paper’s central design tension is that the clone must remain believable as a self-representation while also being sufficiently positive and supportive to be useful; otherwise interaction breakdown, misrepresentation, or alienation occurs (Shirvani et al., 8 Sep 2025).

In these text, journaling, and mental-health systems, Selfment is best understood not as a metaphysical claim about a machine self, but as a computational organization of self-relevant evidence. That evidence may take the form of annotated linguistic traces, passively sensed behavior, or a user’s own supportive language. The resulting systems function as self-profilers, behaviorally informed reflective scaffolds, or externalized self-conversation agents.