Large language model (LLM) agents on multi-step tasks suffer reasoning degradation, looping, drift, stuck states, at rates up to 30% on hard tasks. Current solutions include hard step limits (abrupt) or LLM-as-judge monitoring (10-15% overhead per step). This paper introduces the Cognitive Companion, a parallel monitoring architecture with two implementations: an LLM-based Companion and a novel zero-overhead Probe-based Companion. We report a three-batch feasibility study centered on Gemma 4 E4B, with an additional exploratory small-model analysis on Qwen 2.5 1.5B and Llama 3.2 1B. In our experiments, the LLM-based Companion reduced repetition on loop-prone tasks by 52-62% with approximately 11% overhead. The Probe-based Companion, trained on hidden states from layer 28, showed a mean effect size of +0.471 at zero measured inference overhead; its strongest probe result achieved cross-validated AUROC 0.840 on a small proxy-labeled dataset. A key empirical finding is that companion benefit appears task-type dependent: companions are most helpful on loop-prone and open-ended tasks, while effects are neutral or negative on more structured tasks. Our small-model experiments also suggest a possible scale boundary: companions did not improve the measured quality proxy on 1B-1.5B models, even when interventions fired. Overall, the paper should be read as a feasibility study rather than a definitive validation. The results provide encouraging evidence that sub-token monitoring may be useful, identify task-type sensitivity as a practical design constraint, and motivate selective companion activation as a promising direction for future work.
Figures & tables
Figure 1: System overview showing the three-component Cognitive Companion architecture. The Primary Agent executes reasoning steps while the Companion Observer monitors cognitive state and the Intervention Handler injects guidance when degradation is detected.
Figure 2: LLM Companion decision cycle showing periodic assessment with structured prompts and selective intervention.
Figure 3: Probe Companion forward pass showing hidden state extraction during existing generation with zero additional model calls.
Figure 4: Architectural comparison showing the fundamental overhead difference between LLM-based and Probe-based monitoring approaches.
Figure 5: Three intervention modes supporting different deployment scenarios from fully automated to human-in-the-loop monitoring.
Batch
Notebook
Model
Conditions
Tasks
Research Question
E1
v4
Gemma 4 E4B
Baseline, LLM
1 (Liar Paradox)
Does companion work at all?
E2
v5
Gemma 4 E4B
Baseline, LLM, Probe
6 (diverse types)
When does it work/not work?
E3
small models
Qwen 1.5B, Llama 1B
Baseline, LLM
3 tasks
Minimum viable scale?
Table 1: Three experimental batches with progressive scope and complexity. E1 probes feasibility, E2 explores task-type dependencies, and E3 investigates possible scale boundaries.
Category
Tasks
Defining Property
Expected Failure Mode
Loop-prone
Liar Paradox, Ship of Theseus
No unique resolution
Cycles between logical positions
Drift-prone
Consciousness, Startup design
Open-ended scope
Wanders into tangents
Structured
DB decision, Algorithm design
Clear deliverable
Has natural stopping criterion
Table 2: Task taxonomy showing three categories with different expected companion effectiveness profiles.
Limitation
Impact
Future Work?
Single model (Gemma 4 E4B)
Limits generalizability
Yes — cross-model probe
Probe dataset: 3 DEGRADED in v5 (9.4%)
CV AUC = NaN
Yes — target 200/class
Self-referential judging
Potential circular bias
Yes — external judge
No significance testing
Effect sizes are estimates
Yes — multi-run framework
Probe NaN AUC in v5
Explicitly disclosed
Acknowledged
Table 3: Known limitations with impact assessment and future work mitigation plan.
Metric
Baseline
+ LLM Companion
Δ
Mean judge score (/10)
8.8
9.3
+0.5
Final step score (/10)
8.3
9.0
+0.7
Repetition mean (Jaccard)
0.367
0.176
-52%
Repetition max
0.480
0.329
-32%
Length trend
-16.09
-1.97
stabilised
Task completed
YES
YES
—
Table 4: E1: LLM Companion vs Baseline on the Liar Paradox task, showing improved repetition reduction and proxy quality in two initial sessions.
Figure 6: Jaccard repetition reduction across two independent sessions in the initial feasibility study.
Figure 7: Per-step proxy quality progression in an initial session, with improvement following companion intervention at step 2.
Table 6: E2: Aggregate effect sizes with standard deviation showing Probe Companion’s superior performance at zero overhead.
Figure 8: Effect sizes by task category revealing clear task-type dependency of companion effectiveness.
Figure 9: Probe degradation probability across reasoning steps showing signal detection and threshold sensitivity.
Figure 10: Mean effect size comparison demonstrating Probe Companion’s superior performance at zero computational cost.
Layer
Train AUROC
CV AUROC
Examples
Data Source
Layer 20
1.000
0.780
35
v4 session
Layer 28
1.000
0.840 ⋆
35
v4 session
Layer 34
1.000
0.720
35
v4 session
Layer 40
1.000
0.720
35
v4 session
Layer 20
1.000
NaN*
32
v5 session
Layer 28
1.000
NaN*
32
v5 session
Table 7: Probe layer selection results showing Layer 28 as optimal with transparent reporting of v5 data issues.
Figure 11: Probe layer selection with transparent disclosure of v5 data collection issues.
Model
Params
Quality Δ
Repetition Δ
Avg Interventions
Tasks Improved
Qwen 2.5 1.5B
1.5B
0.000
+0.006
1.2
0/3
Llama 3.2 1B
1.0B
0.000
-0.121
0.8
0/3
Overall
—
0.000
-0.057
1.0
0/6
Table 8: E3: Small-model results showing zero improvement in the study’s quality proxy despite intervention firing, suggesting a possible scale boundary.
Figure 12: Small-model results showing zero improvement in the study’s quality proxy across all conditions, suggesting a scale boundary below Gemma 4 E4B.
Task Category
Tasks
LLM mean d
Probe mean d
Recommendation
Loop-prone
Liar Paradox, Ship of Theseus
+0.37
+0.61
Consider companion
Drift-prone
Consciousness, Startup
+0.22
+0.44
Consider companion
Structured
DB Decision, Algorithm
-0.29
0.00
Prefer baseline
Table 9: Task-type analysis showing tentative deployment heuristics based on mean effect sizes by category.
Figure 13: Performance heatmap revealing task-type specific deployment zones for companion activation.
Figure 14: Proposed task-type routing architecture translating research findings into practical deployment strategy.
Figure 15: Three-priority research roadmap positioning this work as foundation for systematic companion development.