Competitive Context Displacement: Shielding Multi-Agent Working Memory from Cascading Errors

When a language-model agent works through a long task, its own past outputs are appended, unedited, to the context it reads next. A subtle mistake in turn three therefore becomes an unquestioned premise for turn four onward, and session-level success collapses even when each turn in isolation looks fine: on the PowerPoint task-completion benchmark, GPT-4 scores 75.1% per turn on creation tasks yet completes as few as 6% of complex multi-turn editing sessions (Guo et al.). The failure is structural — the context window behaves like an unprotected surface to which any generated error binds permanently, what Lu et al. call the immutable-error problem — and models are trained to stay consistent with their own history rather than accurate, so they will defend an earlier hallucination against contrary evidence (Zhong et al.).

Molecular neurobiology solved an isomorphic problem decades ago. Toxic misfolded protein oligomers kill cells by binding to lipid membranes; because the oligomers are structurally disordered with no stable binding pocket (Tóth et al.), direct targeting fails. The working solution inverts the target: administer small molecules (trodusquemine, squalamine) with high affinity for the membrane, which saturate its binding sites and outcompete the oligomers — reducing membrane-bound oligomers by 65–71% across three unrelated protein systems while leaving the oligomers themselves untouched, even at tenfold molar excess (Limbocker et al.). Notably, trodusquemine can even accelerate aggregation into fibrils while suppressing toxicity, because what matters is membrane occupancy, not the aggregate's fate (Limbocker et al.). This proposal transplants that design into multi-agent LLM systems: we do not correct, rewrite, or delete erroneous context — we outcompete it.

Core Claim

In a multi-agent pipeline (planner, executor, verifier roles), the working context that downstream agents condition on is the "membrane": a limited surface of slots competing for attention. Errors bind to it and poison everything downstream. The proposal, Competitive Context Displacement (CCD), injects tool-verified "displacer" statements that have high affinity for the same semantic slots as flagged errors, so the verified content occupies the surface and the error — retained verbatim in an append-only archive, exactly as toxic oligomers persist in solution — loses its downstream influence. Three commitments come directly from the membrane analogy and are what make this more than a renamed context filter: displacers never remodel the error's text; displacement is graded and reversible rather than binary deletion; and it is stress-triggered and selective, so legitimate speculative content (hypotheses, plans, dissent) is preserved the way trodusquemine preserves physiological monomer-membrane binding (Limbocker et al.).

Mechanism

Each working-memory item $i$ (a claim, parameter, or intermediate result) carries an affinity score $a_i = \alpha v_i + \rho r_i$, where $v_i$ is a tool-grounded verification score (code execution, retrieval match, or verifier-agent verdict) and $r_i$ is relevance to the current subtask. The working context $W_t$ holds the top $K$ items by soft competition:

$$P(i \in W_t) \;\propto\; e^{\beta a_i}$$

with temperature $\beta$ controlling how sharply the surface saturates. When a stress signal fires — verifier disagreement or a perplexity spike, reusing Tian et al.'s objective-masking signal as a trigger rather than a deletion rule — the verifier agent generates a displacer: a verified statement covering the same semantic slot as the flagged item, entering with high $v_i$. If the flagged item's owner contests the displacement, a short FORD-style focused debate with explicit refute-or-concede turns (Xiong et al.) adjudicates, using evidence-anchored rules to avoid the debate-hacking pathology documented by Chen et al.. Full history lives in an archive; only $W_t$ is curated. Task decomposition (Direction 3) supplies the verification capacity — a dedicated verifier with tool access, in the spirit of MACM's Thinker/Judge/Executor split (Lei et al.) and Dong et al.'s Analyst/Coder/Tester team (Dong et al.) — while debate (Direction 2) is reserved for the rare contested case, keeping cost low.

Research Program

The program builds CCD as a modular layer over existing agent pipelines, then tests whether the displacement mechanism — not merely the verification it wraps — is doing the work. Three studies. First, implementation and calibration: fit $\beta$, $K$, and the stress-trigger threshold on held-out development splits, and characterize the dose-response curve (displacer density versus error occupancy of $W_t$), the direct analog of concentration-dependent membrane shielding. Second, mechanism isolation: ablations that replace graded displacement with hard deletion (MAD-M²-style masking), with rewriting (ACR-style fact rectification, Shen et al.), or with unconditional always-on filtering, each holding the verifier fixed, so any difference is attributable to the displacement rule alone. Third, equilibrium predictions: remove displacer refresh mid-session and measure whether errors re-bind to $W_t$ over subsequent turns — the membrane analogy predicts they will, which no deletion- or rewrite-based account predicts and which would falsify the equilibrium framing if absent.

Baselines

The naive reference point is the unmodified single agent, whose fragility is already quantified: GPT-4 achieves 75.1% turn-based accuracy on PPTC creation tasks but session-level completion as low as 6% on complex editing (Guo et al.), so any method must be judged at session level, not turn level. The strongest existing competitor on the debate-plus-cascade question is MAD-M², which masks erroneous debate-round memories via self-rating and perplexity filters and reports Qwen2.5-Math-7B rising from 71.2% to 80.2% on MATH, a 6.6% gain on AIME, and roughly 30% token savings, at a 13–41% latency penalty (Tian et al.); we will run it with the same backbone, agent count, debate rounds, and total inference budget as CCD. On decomposed role pipelines, MACM is the comparator, reporting GPT-4 Turbo Level-5 MATH accuracy of 54.68% rising to 76.73% and 99% on the 24-Point Game (Lei et al.). On multi-turn state management, ACR reports HotpotQA exact match improving from 21.7% to 35.1% while cutting cost from 1,678 to 278.5 tokens per turn (Shen et al.); we compare both accuracy and token cost, since MAD-M² shows masking and latency trade off directly. On error-detection oversight, ColMAD beats competitive debate by 19% on ReaLMistake (Chen et al.). The key ablation converts CCD's graded competition into binary deletion with identical verifier output: if deletion matches CCD everywhere, the displacement mechanism adds nothing and the verification layer was the real driver — a result we would report as a null finding for the mechanism. Beating MAD-M² means higher session-level completion on PPTC editing and higher MATH/AIME accuracy at equal or lower token cost, since MAD-M²'s own MATH number under identical settings is 80.2%.

Why not just…?

Why not just delete the errors, as MAD-M² already does? MAD-M²'s masking is binary and destructive: pruned content is gone, so correct material entangled with an error is lost with it, and its latency cost runs 13–41% (Tian et al.). CCD keeps the full transcript, replaces rather than removes, and — critically — masks on self-rating and perplexity, signals with no ground-truth anchoring, whereas displacers are tool-verified statements. The head-to-head with the deletion ablation settles whether the difference matters empirically; if deletion wins at equal cost, the membrane framing earns nothing.

Why not just rewrite the context, as ACR's fact rectification does? Rewriting is the "remodel the oligomer" strategy the membrane literature found infeasible: the rewriting model faces the same self-verification blind spot as the original, and every rewrite risks injecting a new hallucination while destroying provenance. ACR's rectification works within a single agent and reports strong multi-hop QA gains (HotpotQA EM 21.7% to 35.1%, Shen et al.), and LLM-as-RNN rewrites a hidden state every step with a critic's semantic gradient, self-healing 54.8% of clinical-note errors (Lu et al.). These are real threats. Our experiments address them directly: the rewrite ablation (rectify rather than displace, same verifier) and a provenance-audit metric counting cases where the rewrite introduces a new error — the mechanism the analogy predicts will favor displacement, since it never generates replacement text beyond tool-verified output.

Why not just gatekeep every step with a Judge, as MACM does? MACM's Judge is an LLM self-check with no external anchoring, and its Condition List is strictly additive — no eviction, no competition for limited context — which is why the authors flag repetitive Thinker-Judge loops as its dominant cost (Lei et al.). CCD differs on both axes: verification is tool-grounded, and the surface is competitive rather than append-only, so the context cannot saturate. The PPTC and MATH comparisons against MACM, plus a context-occupancy audit at long horizons, test whether competitive eviction is what prevents the saturation failure MAD-M² observed at long debate horizons.

Why not just fine-tune the model to recover from errors? The pathology is architectural — immutable history conditioning — not a knowledge gap, and role-based methods already degrade below 13B parameters (Dong et al.), so a training fix re-encounters the consistency-over-accuracy bias that Zhong et al. show persists through decoding-time interventions like their residual visual decoding. Training is also incompatible with our frozen-backbone, backbone-agnostic evaluation design. We treat this as out of scope rather than refuted.

Proposed Experiments

Four experiment groups, all on benchmarks with reported reference numbers. First, long-horizon execution on PPTC (Guo et al.): 279 sessions of 2–17 turns, scoring session-level completion, position-relation accuracy (GPT-4 drops to 24% there), and a new error-rebinding trace measuring how often a displaced error re-occupies $W_t$ after refresh lapses — the decisive equilibrium check. Second, mathematical reasoning with debate: MATH, AIME, and GSM8K with Qwen2.5-Math-7B for direct comparability to MAD-M²'s 71.2%→80.2% and 6.6% AIME gains (Tian et al.), and with GPT-4 Turbo against MACM's 76.73% Level-5 figure (Lei et al.). Third, multi-turn multi-hop QA on HotpotQA against ACR's 35.1% EM at 278.5 tokens per turn (Shen et al.), reporting accuracy and cost jointly. Fourth, oversight on ReaLMistake against ColMAD's 19% margin over competitive debate (Chen et al.). Throughout, we run the physiological-binding control: tasks requiring speculation (hypothesis ranking on APPS-style design subtasks) where a too-aggressive filter should hurt — deletion-based methods should lose more than CCD, whose displacers compete only on verified slots. Success is defined narrowly: CCD beats MAD-M² and the deletion ablation at matched compute on session-level metrics, shows no rewriter-introduced errors in the provenance audit, and exhibits the predicted dose-response and reversibility signatures. If the deletion ablation ties CCD across all four groups, the honest conclusion is that tool-anchored verification, not displacement, carries the gains — and that would itself sharpen where the field's effort belongs.