Probe-based Companion in AI

• Probe-based Companion is a secondary AI module that inspects hidden states, memory stores, or runtime telemetry to determine system performance.
• It employs methodologies like logistic regression over hidden layer activations and cosine-similarity based retrieval to inject corrective guidance without affecting the primary agent.
• Its design spans various applications—from reasoning degradation detection and memory augmentation to post-failure recovery and on-demand user intervention—each tailored to specific task demands.

A probe-based companion is a companion architecture that remains separate from a primary agent or interaction loop and derives its interventions by explicitly inspecting some externalized or internal representation rather than relying only on the base model’s next-token dynamics. In current arXiv usage, the term is explicit in hidden-state monitoring for LLM reasoning, where a parallel companion reads activations and injects guidance when degradation is detected; a broader inferred usage includes memory-augmented companions that probe external memory stores, post-failure recovery layers that probe runtime telemetry, and technology probes that sit beside AI companion applications to mediate breakdowns (Khan et al., 15 Apr 2026).

1. Conceptual scope and definitional variants

The most literal use of the term appears in "The cognitive companion: a lightweight parallel monitoring architecture for detecting and recovering from reasoning degradation in LLM agents" (Khan et al., 15 Apr 2026). There, a Probe-based Companion is a zero-measured-overhead observer that reads hidden states from the main model’s own forward pass, classifies whether the agent is ON_TRACK or DEGRADED, and selectively injects guidance without altering the primary agent’s architecture. The paper’s formalization makes the separation explicit: the companion takes recent history $H_{t-k:t}$, hidden-state information $\Phi_t$, and companion parameters $\theta$, then outputs an intervention decision and guidance.

A broader, inferred usage emerges in "Embedded AI Companion System on Edge Devices" (Gupta et al., 13 Jan 2026). That paper does not use the term explicitly, but its active-phase retrieval stack repeatedly probes a long-term memory store, a short-term session store, and a structured user profile with an embedding of the current utterance. The current query vector is therefore a probe over externalized memory representations rather than an appeal to persistent latent transformer state alone. This suggests a retrieval-driven interpretation of probe-based companionship.

A third variant appears in "Debugging the Debuggers: Failure-Anchored Structured Recovery for Software Engineering Agents" (Zhao et al., 9 May 2026). PROBE is not a conversational companion, but it is a companion layer that attaches to an existing agent workflow after failure, preserves typed telemetry, generates an anchor-first diagnosis, and emits bounded recovery guidance only when grounded and actionable. Here the probed object is failed-run telemetry rather than hidden states or autobiographical memory.

A fourth usage is HCI-specific. "Minion: A Technology Probe for Resolving Value Conflicts through Expert-Driven and User-Driven Strategies in AI Companion Applications" presents a browser-side intervention layer for Character.AI and Talkie (Fan et al., 2024). In that paper, “probe” denotes a technology probe in the Hutchinson et al. sense: a deployable artifact used simultaneously for engineering, social-science, and design inquiry.

Variant	Probed substrate	Representative source
Hidden-state monitor	Layer activations during generation	(Khan et al., 15 Apr 2026)
Memory-augmented companion	Short-term memory, long-term memory, user profile	(Gupta et al., 13 Jan 2026)
Recovery side channel	Telemetry, evidence, diagnosis objects	(Zhao et al., 9 May 2026)
Technology probe	Live companion dialogs and conflict episodes	(Fan et al., 2024)

This range of usage suggests that the phrase is currently polysemous rather than standardized. The common structure is a secondary component that inspects an otherwise opaque process and then mediates, steers, or repairs the primary interaction.

2. Hidden-state monitoring and reasoning-degradation probes

The most technically specific probe-based companion architecture is the Cognitive Companion (Khan et al., 15 Apr 2026). The framework models the primary agent as a sequential process with cumulative history

$$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$

and next response

$r_{t+1} = M(T, H_t, G_t),$

where $G_t$ is any guidance injected so far. The companion architecture has three components: the Primary Agent, the Companion Observer, and the Intervention Handler. The primary agent is architecturally unchanged and unaware of the companion’s explicit presence.

The probe implementation reads hidden states from Gemma 4 E4B, a decoder-only transformer with $L=42$ layers and hidden dimension $d_{\text{model}}=2560$. Hidden states are exposed by enabling output_hidden_states=True during generation. For each reasoning step $t$, the probe constructs a step-level feature by mean-pooling the final-position hidden states of the last $w=10$ generated tokens at layer $\Phi_t$0: $\Phi_t$1 This yields one $\Phi_t$2-dimensional vector per step. The task is then reduced to binary classification,

$\Phi_t$3

where $\Phi_t$4 and $\Phi_t$5, with DEGRADED collapsing LOOPING, DRIFTING, and STUCK.

The probe itself is deliberately minimal: a binary logistic regression classifier trained independently at each candidate layer,

$\Phi_t$6

with candidate layers $\Phi_t$7. The strongest reported result is at layer 28, where the paper reports Train AUROC 1.000, CV AUROC 0.840, on the v4 session dataset with 35 proxy-labeled examples. The labels are generated by the LLM-based Companion rather than by human annotation, a limitation the paper emphasizes.

Operationally, the probe checks every 2 agent steps and uses a fixed degradation threshold of 0.55 in Whisper Mode, injecting constructive guidance silently into context. The main attraction is that detection requires no additional model calls and incurred 0% measured inference overhead in the reported setup. On the six-task Experiment 2 study, the Probe Companion achieved Mean $\Phi_t$8 (score) = $\Phi_t$9, Mean $\theta$0 (repetition) = $\theta$1, and Overhead = 0% (Khan et al., 15 Apr 2026).

The gains are not uniform. The probe was strongest on Liar Paradox with effect size +2.04, positive on Consciousness at +0.87, neutral on Startup Design and DB Decision, and negative on Ship of Theseus. By task category, the paper reports Loop-prone tasks with Probe mean $\theta$2, Drift-prone tasks with Probe mean $\theta$3, and Structured tasks with Probe mean $\theta$4 (Khan et al., 15 Apr 2026). This establishes the paper’s main deployment constraint: companion benefit is task-type dependent rather than universal.

The paper is equally explicit about limitations. The probe is trained on proxy labels, the strongest result uses only 35 examples, the main v5 probe experiment is in-domain, and the quality metric is self-referential because the same model family judges the outputs. The work therefore supports feasibility of hidden-state probing, but not robust claims about cross-task generalization, cross-model transfer, or production reliability.

3. Retrieval-driven memory probes on constrained edge companions

A second major sense of probe-based companionship is retrieval over external memory. "Embedded AI Companion System on Edge Devices" addresses a child-facing offline companion deployed on an NVIDIA Jetson Orin Nano Super 8GB, where the available generator is only Qwen2.5-7B-Instruct quantized to int4 and practical prompt lengths are much smaller than the nominal context window because latency constraints are severe: around ~5 s requires the prompt to stay below roughly ~2500 tokens, and around ~2 s requires roughly ~1000 tokens (Gupta et al., 13 Jan 2026).

The architecture alternates between active and inactive phases. In the active phase, during live conversation, the system embeds the current utterance $\theta$5 into $\theta$6 using gte-base-en-v1.5, performs cosine similarity search over a long-term memory store and a short-term session store, retrieves top-$\theta$7 and top-$\theta$8 candidates, filters them using $\theta$9, and builds the response prompt from retrieved memories, the user profile, and a recent sliding window of $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$0 messages. Retrieved short-term memories are further expanded with $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$1. The paper does not name this a probe module, but functionally the utterance embedding serves as a probe over external memory banks.

The inactive phase, triggered after $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$2 minutes, performs heavier maintenance. The full session is chunked into $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$3 token segments, later chunks inherit the last $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$4 messages plus an LLM-generated summary, and the model extracts long-term memories and a structured profile, merges chunk-level outputs, reconciles them with the global profile, and runs a memory add/update decision against similar existing memories (Gupta et al., 13 Jan 2026). This is a second, retrospective probing schedule: light probes online, heavy probes offline.

The memory state is explicitly heterogeneous. The paper stores short-term query-response tuples, long-term textual memories, a structured user profile containing at least name, age, gender, a personality summary, inferred Myers-Briggs and Big Five attributes with supporting dialogue examples, and chunk summaries used only during extraction. Long-term memories are stored as

$$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$5

and forgetting follows the MemoryBank-style retention model

$$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$6

with memories deleted in the inactive phase if $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$7, where $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$8. Memory strength $$H_t = \{(p_1, r_1), (p_2, r_2), \ldots, (p_t, r_t)\},$$9 is initialized to 1 and incremented on retrieval. The store is therefore adaptive rather than static.

The benchmark results show where this retrieval-based probing helps and where it fails. In Conversation Quality, the proposed system scores 2.6 vs 1.6 for raw Qwen on naturalness and 3.0 vs 1.6 on personalization. In Generated QA, it reaches 43.56% vs 28.09% on Specific questions relative to raw Qwen, and exceeds raw GPT-3.5 Turbo at 37.74%, but on Inferred questions it scores only 49.0%, below raw Qwen’s 58.5% and GPT-3.5’s 70.83%. In Personalization, it achieves 5.4 hits / 4.8 misses, versus 3.8 / 6.0 for raw Qwen. In Extraction Quality, it reports 77.44% correctness, 29.38% coverage, and 45.14% completeness (Gupta et al., 13 Jan 2026).

The paper’s own explanation of the Inferred-question weakness is especially revealing for the probe interpretation: “we only insert some fixed amount of retrieved memories, and sometimes may not retrieve any relevant memories,” especially when semantic similarity is weak. The system is therefore strongest when the needed fact is semantically retrievable as a stored item, and weaker when the answer requires assembling multiple weakly matching traces.

4. Telemetry-grounded side-channel companions in engineering agents

PROBE generalizes the companion idea away from dialogue and memory into post-failure engineering recovery (Zhao et al., 9 May 2026). The paper presents it not as a replacement agent, not as a new policy, and not as an in-loop controller, but as a portable, framework-agnostic Python package that provides side-channel recovery support after a failed run. It instruments boundary events, preserves fine-grained runtime evidence, and produces artifacts for a second attempt without modifying the agent policy, toolset, executor, or evaluator.

Its architecture is organized as

$r_{t+1} = M(T, H_t, G_t),$0

where $r_{t+1} = M(T, H_t, G_t),$1 is a typed telemetry bundle, $r_{t+1} = M(T, H_t, G_t),$2 structured evidence, $r_{t+1} = M(T, H_t, G_t),$3 a structured diagnosis object, and $r_{t+1} = M(T, H_t, G_t),$4 bounded recovery guidance. The telemetry bundle spans metrics, logs, traces, intent, environment state, and optional outcome signals. The basic recorded unit is the span $r_{t+1} = M(T, H_t, G_t),$5, and the design explicitly distinguishes agent intent $r_{t+1} = M(T, H_t, G_t),$6 from tool–environment state $r_{t+1} = M(T, H_t, G_t),$7. This distinction is central because diagnosis is grounded in mismatches between what the agent was trying to do and what actually happened.

The diagnosis pipeline is anchor-first. Failure localization uses robust z-scores with MAD, empirical quantile thresholds, Isolation Forest on window-level features when no single metric dominates, and a smoothed $r_{t+1} = M(T, H_t, G_t),$8-gram model for transition surprise in traces and intent. Cross-signal fusion then creates evidence units with anchor, source, time scope, severity, and evidence reference. The Guidance Gate applies a Grounding Check, an Actionability Filter, and Guidance Injection. Guidance is injectable only if it has a supported primary cause or failure anchor, can be instantiated as target, operation, verification signal, and boundary condition, and remains within agent-side recovery scope (Zhao et al., 9 May 2026).

On 257 initially unresolved cases, PROBE reports 65.37% Top-1 diagnosis accuracy and a 21.79% recovery rate, outperforming the strongest non-PROBE baseline by 43.58 and 12.45 percentage points, respectively. In AIOpsLab specifically, PROBE and LangSmith Summary both reach 44.90% Top-1, but PROBE recovers 30.61% of cases versus 16.33% for LangSmith Summary. The paper uses this gap to argue that diagnosis quality is necessary but insufficient unless translated into bounded, evidence-grounded recovery guidance (Zhao et al., 9 May 2026).

In encyclopedia terms, PROBE broadens the meaning of probe-based companion. The probed object is not a latent activation or autobiographical memory bank, but a failed execution trace. The companion function is still the same: observe without replacing, synthesize grounded structure, and intervene only through a bounded side channel.

5. Technology probes, value-conflict mediation, and probe-adjacent companions

Minion represents a distinct strand in which probe-based companionship is framed as user-empowerment intervention rather than latent-state inspection (Fan et al., 2024). The paper begins with a formative analysis of 151 complaint posts about AI companions, of which 146 involved value conflicts. Mapped onto Schwartz’s ten basic value types, the most frequent were Universalism: 37, Conformity: 25, and Power: 23. On that basis the authors built a Chrome browser extension for Character.AI and Talkie that appears as a floating HELP button, captures the companion’s introduction and the current conversation history, and returns four candidate user responses: 2 expert-driven strategies and 2 user-driven strategies, shown in random order.

The expert-driven strategies are Proposal, Power, Interests, and Rights. The user-driven strategies are Out of Character, Reason and Preach, Anger Expression, and Gentle Persuasion. Generation uses gpt-4o-2024-05-13 with temperature = 0.2 and top_p = 0.1. The technology-probe study involved 22 participants, 274 tasks, 124 diary notes, and 11 hours of interview recordings. By the paper’s short-term criterion for successful conflict resolution, 258 of 274 tasks were resolved, yielding 94.16%. Minion was invoked 919 times, with 489 expert-driven and 430 user-driven selections; the most used strategies were Reason and Preach: 198, Proposal: 177, and Interests: 136 (Fan et al., 2024).

This is not probing in the representation-analysis sense. Here “probe” denotes an exploratory HCI artifact deployed beside an existing companion platform. Yet it remains structurally companion-like: it is adjacent, non-intrusive, user-invoked, and designed to preserve conversational continuity rather than replace the base system. The paper’s strongest design lesson is that intervention should be available on demand rather than automatically imposed.

An adjacent but distinct design appears in "AIVA: An AI-based Virtual Companion for Emotion-aware Interaction" (Li, 3 Sep 2025). AIVA uses a Multimodal Sentiment Perception Network (MSPN) with cross-attention, a cross-modal fusion transformer, learnable sentiment prototypes, and supervised contrastive learning, then injects the inferred emotion into an emotion-aware prompt for LLaMA2-Chat. The paper explicitly notes that this is not representational probing in the classic sense; it is an end-to-end emotion classifier plus prompt conditioning. MSPN reaches 74.25% accuracy / 72.84% F1 on MVSA-Single and 73.48% accuracy / 70.01% F1 on MVSA-Multi (Li, 3 Sep 2025). In the context of probe-based companions, AIVA is best read as probe-adjacent: it exposes interpretable internal emotional representations, but it does not define a separate companion layer that probes a primary model.

6. Limits, task dependence, and broader terminological boundaries

Taken together, these papers suggest that probe-based companions are most effective when the probed substrate is well matched to the target failure mode. Hidden-state probes help when degradation is expressed in ongoing reasoning trajectories, especially on loop-prone and open-ended/drift-prone tasks (Khan et al., 15 Apr 2026). Retrieval-based memory probes help when personalization depends on semantically retrievable facts, but they can fail on Inferred questions with weak semantic overlap (Gupta et al., 13 Jan 2026). Telemetry-grounded companions help when failures leave structured runtime evidence and when diagnosis can be translated into bounded guidance (Zhao et al., 9 May 2026). Technology probes help when the central problem is not model cognition but user agency under relational breakdown (Fan et al., 2024).

The common limitations also recur. The hidden-state probe literature is currently constrained by proxy labels, tiny in-domain datasets, and lack of significance testing (Khan et al., 15 Apr 2026). Edge-memory companions still rely on fixed-$r_{t+1} = M(T, H_t, G_t),$9 retrieval, embedding similarity, and a weak local model that struggles with nuanced abstraction and temporal reasoning (Gupta et al., 13 Jan 2026). PROBE can diagnose many failures correctly but still recovers only 21.79% of initially unresolved cases, making the diagnosis–recovery gap explicit (Zhao et al., 9 May 2026). Minion reaches high short-term conflict-resolution rates, but extreme bias, output breakdown, and platform moderation still produce unresolved cases (Fan et al., 2024).

There is also a terminological boundary. Outside AI, the words probe and companion often combine in unrelated ways rather than designating a single methodology. JWST/MIRI observations use spectroscopy as a probe of disks around forming companions (Cugno et al., 2024). Companion galaxies are used as probes of AGN radiation in time and direction (Keel et al., 2017). Post-supernova companion monitoring uses the surviving star as a probe of binary interaction physics (Hirai, 2023). Pulsar-wind studies treat a companion as a wind probe that reveals shock geometry, forced reconnection, and orbitally modulated high-energy emission (Richard-Romei et al., 2024). These are genuine probe–companion constructions, but not AI companion architectures.

In contemporary AI usage, the most precise sense of probe-based companion is therefore a secondary module that remains architecturally separate from a primary model or workflow and derives interventions by probing hidden states, external memory, runtime telemetry, or interaction traces under explicit latency, autonomy, or deployment constraints. The term is explicit only in a narrow subset of this literature, but the architectural pattern is broader: a companion that observes first, intervenes second, and treats inspection itself as the core mechanism of assistance.